What changed in T2D3 OS, every day.

New

• Agency payouts now wait for approval. When a client accepts work delivered through an agency package, the agency's earning is held as awaiting approval, with the reason shown, and nothing is paid before a seven-day hold from the order date. A new Agency payout approvals panel on the workers admin page lists each held earning with its delivery evidence: who accepted and how, the rating, revisions used, the deliverable link, the brief and the delivered note. Each row opens with an AI recommendation that a devil's-advocate pass has already argued against; the admin approves, holds or rejects with a reason, and the AI never recommends rejection or moves money on its own. Every decision feeds the recommender.
• Plan a Signal meeting before it runs. Rhythm gains Plan a meeting beside Start Meeting: pick a day, and the meeting opens in prep with no timer, so field signals, friction and the prep brief attach to it across the week. Upcoming prepared meetings are listed in the card to reopen or discard, and Start Meeting on the day promotes the same meeting instead of creating a second one.
• The connector can prep the Signal meeting end to end. A headline added with no meeting running is held on the project and moves into the next meeting when it starts; an agent can set issue priority and order, assign owners by name, and read scorecard metrics. A person still presses Start.
• An overnight agent pairs transcripts with recordings. When a meeting transcript you uploaded sits in the Signal Library beside its transcribed recording, the agent scores the pair on shared phrases, length, title and date and writes one proposal with quoted evidence. It never attaches anything itself; a card on both rows lets either inspector accept, and accept folds the two into one grounded recording.
• ICP team review settles split votes. The Review lens shows how much of the ICP each teammate has reviewed and which items are agreed, contested, decided or not yet reviewed. A PM breaks a contested item's tie with an AI recommendation and the case against it in front of them, first click wins, and the lock snapshot records how much of the ICP the team actually agreed on.
• Review nudges for the ICP. Request review asks the teammates who have not voted, once per ICP per week, and reports who was asked. A vote that leaves an item contested drafts the recommendation in the background and sends each PM one tie-break ask per ICP per day. An @mention in a comment reaches exactly the named teammate.

Improved

• The connector no longer reads individual performance or HR data. Team development, assessments, coaching, calibration, recruiting, and a person's own development matrix, work style, growth and merit reviews are gone from the public connector; course catalogs, community templates and job postings stay.
• Adding a lead or moving a contact to prospect over the connector no longer triggers third-party enrichment, so the connector's no-enrichment claim is true in code. The app's own dialogs are unchanged.
• Generate SAFE drafts only the option the investor has already chosen, instead of every option the round offers. The signature email now CCs the admin who sent it, and the company signer prefills as the CEO; every signer field stays editable.
• Newer-model findings arrive as a decision, not an alert. Each one carries an AI recommendation that a devil's-advocate pass has stress-tested, and the admin adopts the candidate or holds the finding for thirty days.

Fixed

• The Messaging credibility review headline counts the distinct decisions left, not the raw number of claims. Similar claims were already grouped so that one thumb rules the whole group, but the headline hid that; it now says how many decisions remain and that a vote on a group rules every claim in it.
• Connecting an integration from Settings returns you to the Integrations pane. Before, HubSpot, Figma, Canva, GitHub, LinkedIn, X and Google Slides connects landed on the Members tab.
• A retired background job's leftover monitor no longer reads as a job that stopped checking in. The alert now names the exact credential scope the cleanup needs instead of asking whether to revive it.

Under the hood

• The ChatGPT plugin package carries its demo recording and a new version number, so the directory accepts a resubmission.
• A new milestone vesting letter joins the legal templates, and any legal template can now be rendered and sent from a reviewable values file with the private details kept out of the repository and out of every log.

New

• A complete ICP goes quiet. Once an ICP is locked or scores 100, it stops sprouting inline cards, strips and nudges. New signal, proposed edits, merges, estimates and evidence re-checks collect behind one Update ICP button, and you accept, decline or snooze each one at a time; Later brings it back in seven days. The background build now parks new rows as proposals instead of writing them into settled work.
• One count of ICP updates waiting, read from one place: the Awaiting-you line in Agent Activity, a dot on each ICP in the switcher, and the connector can list them.
• Upload a meeting transcript instead of recording the room. The Meeting recording panel on a note takes a VTT, SRT, TXT or DOCX; speaker names and timestamps are kept, and it grounds action items, the follow-up email and Complete my notes with no audio pass. The same door sits on a Signal Library recording: an uploaded transcript becomes the governing one, the AI transcript stays, and Revert to AI transcript is one click.
• Copying an ICP to another organization now carries its personas. An Include personas box, on by default, copies each linked persona set into the target as a draft tied to the ICP copy; hidden personas stay hidden, and votes, feedback and the lock stay with the source.
• Accounts issued a password can sign in with it, from a quiet link under the magic-link form on the login page. Magic link stays the default; the password route is rate-limited and bot-checked.

Improved

• The ICP lock review narrates each named step with the time elapsed, including the final check of every item against your source signal, and says the lock is still open while it runs. No more opaque spinner.
• The ICP slide deck quotes the finished ICP's own numbers. Live items go into the deck first, so segment sizes, triggers and deal size match the module. A new Market Size slide draws nested TAM, SAM and SOM rings with per-ring cards, computed the same way as the ICP's Market Size canvas.
• The Claude connector can set a draft module's title and description, archive and restore a module instance, and runs long actions under the same ten-minute budget as the app, so a build that used to die with no message now completes or fails with a reason.
• The ICP help text now says where Lock lives: at the right of the action row under the content, with Build my ICP under More actions.

Fixed

• The Content Studio publish button says a piece is live only when the public page answers. Before, it showed Published while the page still returned 404.
• Paid marketplace package orders email the agency's owners and admins once per order, and the buyer gets a confirmation. Previously only the first of several orders in a day got through, and the buyer heard nothing.
• 7 automatic fixes from the app's own repair loop.

Under the hood

• Every AI call is now attributed to the person and connector token that triggered it, including the background jobs a connector call fans out. An hourly detector flags spend that passes a floor and four times that subject's own recent normal, and the team can acknowledge alerts and tune thresholds without a deploy.
• The AI tiers behind every routed call moved to the current frontier models from Anthropic, OpenAI, Google and xAI, each move backed by an evaluation sweep with no hard failures.

New

• ICP items are reviewed per person. Each member votes thumbs up or down and leaves a note, blind until they have voted; the item then shows the named tally and reads agreed, contested or unreviewed. A pain's verdict follows the team's tally instead of whoever clicked last, and a hand park is never undone.
• ICP edits carry a name. Every change to an item is logged with its author and can be reverted from the item, two people editing the same field get a conflict notice instead of a silent overwrite, and the lock diff credits each change to the person who made it.
• Copy an ICP into another organization you run. The cards travel with their evidence dates and hidden state, the market segment is reused or copied, and provenance says where the foundation came from. Votes, feedback and the lock stay home, and you must be an owner, admin or project lead on both sides.
• ABM list sources: My connections. Upload your LinkedIn connections export or sync your first-degree network from your linked seat after consent; the people feed scoring and contacts, visible only to you.
• ABM title databases per persona. The AI drafts the titles a buyer actually carries and the near-misses to exclude, grounded in your signal and ICP and checked by a devil's advocate; you accept per row, and people searches then use the saved set.
• Call transcripts from your notetaker now link to the deal they belong to, matched by participant emails, so call signal counts toward pipeline instead of floating unanchored.
• Conversation trends: recurring themes across calls, competitors, objections and asks, are counted per segment per week with quotes. A theme that crosses the threshold you confirmed, which the AI recommends first, becomes one content idea.
• A monthly Marketing Intelligence Briefing with receipts: a win log of content touches on closed deals, a before-and-after timeline for each trend you answered, and every number opening to the touches, deals and calls it was counted from.
• Organizations get a backup owner. The backup holds every owner power except transfer and takes the seat automatically if the primary owner leaves; only the primary owner can name or clear them from Settings, Members.
• Signatures appear on the executed PDF. Type your legal name in a script face, draw with pointer, touch or pen, or upload a signature image; it lands on the template's signature line, and the typed name stays the evidence of intent.
• Visual Brand gains a sonic brand: moods, tempo, instrumentation, bans and an audio logo, drafted by the AI and locked by you.
• Video Studio composes three AI music beds per cut from the sonic brand and the storyboard, each with a one-line why. Keep, swap or mute; the kept bed is mixed under the voice, faded at the tail, normalised for loudness and closed with the audio logo on the end card. A cut cannot be saved or published unless every cue carries a cleared licence.

Improved

• The Agent Activity panel on a project page is folded by default, with the activity count and the decisions waiting in the folded line, so the tasks you came for lead the page.
• A candidate who signs in only to sign an agreement sees the document, its trail and the sign panel, not an empty workspace; any other page sends them back to their agreement.
• A support request the coding agent reviews and finds needs no code change now closes as "no code change needed" with the agent's reason, instead of reading as failed. A request already fixed elsewhere resolves once the merged fix is confirmed.
• Fewer, truer emails. Engineering alerts mail on a new failure, a changed cause or a recovery, not on every rerun; a daily brief with the same asks as yesterday is not re-sent; recipients are chosen by role.
• The Claude connector no longer offers the platform cost tool to accounts that cannot call it, and an argument outside the allowed values is refused with the list of valid ones instead of a raw database error.

Fixed

• Clicking Publish in the Content Studio no longer shares a piece to your LinkedIn and X when someone else is the byline. Shares go out only when the publisher is the author, or the byline is empty.
• ICP: the two sizing buttons that did the same thing are one, the yellow "criteria changed" bar no longer returns right after a re-size, the SAM card lines up with the others, and every row in the Accounts lens has thumbs again.
• 1 automatic fix from the app's own repair loop.

Under the hood

• Every tool the connector advertises was exercised on production before the directory reviewers try them; the two defects found are fixed above.

Published

• Journal: No Export Button Is a Spec, Not a Wall
• Journal: You Cannot Throttle What You Have Not Named

New

• ICP → Size gets a Market Sizing canvas: TAM, SAM and SOM drawn as nested circles with step cards, each stage marked done, draft or not yet sized. A stage with nothing behind it says so instead of showing zero.
• SOM becomes real names. A Sales Navigator sub-search can score its CSV export, or run live on your own LinkedIn seat after a consent dialog that states the page count, pace, daily limit and the risk, narrating each page as it goes.
• SOM counts only accounts the scorer or a person judged qualified, so the ICP canvas and the ABM header report the same number.
• Budget Mix: a screenshot or phone photo of a budget table, dropped in the AI Targets box, is read for its numbers. The Targets suggestion cites them, an unreadable cell is flagged rather than guessed, and HEIC photos are accepted.
• Budget Mix: Suggest mix builds the channel mix from the committed lines in your uploaded budget files, each note quoting the line it came from, instead of re-pricing from benchmarks.
• Swipe file: your filed cold emails now ground every outbound draft for the persona they sit under, across relay, sequences, the ABM copywriter and campaign copy. The model matches the register and never copies wording, claims or numbers.
• Swipe file: forward a cold email that got you to open to your org's capture address and it is filed with the reason you forwarded it. Your own hooks that got replies are proposed as swipes for you to confirm.
• Quick add captures glints: a pasted quote, a lone link or an observation files into Signal on the glint lane, where Content and Campaigns can draw on it.
• CRM win patterns say which way the won mix moved, upmarket, downmarket, sector shift or stable, and list what the drifting segment keeps asking for and why it lost.
• Visual Brand: a brand-systems agent sweeps for drift between the locked board and the exported style files, classes each finding major or minor, and flags duplicate assets in the library.
• Academy admins can trim a lesson video in-app: set keep-from and keep-until and the lesson swaps to the clipped copy.
• Legal: offer letters go out from the app with a cover note, a six-event trail from created to completed, and signer-only logins for candidates on personal addresses. Restricted stock purchase, board consent and share surrender templates follow, with a sensitive-field flag that keeps a tax id out of titles and file names.
• Capital: an executed agreement files into the minute book in one step, its signed PDF and certificate checked against the fingerprints recorded at signing.
• One daily "Needs you" email replaces five daily senders and three per-event notices: new items first, at most five lines, nothing sent when nothing is new. Every email now names its kind, so digests can be deduplicated, rate-limited and switched off per person while a password reset always gets through.

Improved

• Help: the ICP section says where the Lock button and the Build my ICP door sit, and the Presentations section says where an exported deck lands, in all seven languages.
• Every agreement names its version: a version chip beside the title on admin rows, in the void confirmation, on the superseded pill and on the signing page, which now fills a wide screen.
• Agency listings: the Edit package link opens in a new tab so the listing dialog keeps your unsaved edits.
• Account lists reuse companies already resolved for another workspace, so the same company is not looked up twice and lists fill faster.

Fixed

• A client member who was not also a member of their agency could not open Marketplace People or spend from the agency's credit pool. Both now read the agency's plan and balance, and a refusal names its reason.
• An agency owner on a comped plan who signed up before the role question had no New Client button on the home page.
• "Sort into ICPs with AI" failed with a database error whenever no customers were selected.
• Opening a video edit whose render had died showed "rendering" forever. It now says failed and offers a retry, and two-camera rendering passed production acceptance.
• The journal series no longer posts first-person shares to the authorizing account's LinkedIn and X when an entry publishes.
• The signing certificate overprinted hashes when a long browser string wrapped. A newly minted signer login no longer opens on the workspace tour over the document.

Under the hood

• Engineering alerts fire once per incident: error tracking runs on production builds only, monitor alerts need three consecutive failures, and preview builds no longer deploy on every push.
• Live-database checks in CI retry connection drops with backoff and still fail closed when the database is unreachable.

New

• Max remembers the person. What you tell it about yourself or how you work is kept and rides every later thread, with a Remembered chip you can undo, a Manage panel under account settings, and a switch to turn memory off.
• Max keeps the whole conversation. Long threads replay within a token budget and fold their oldest turns into a rolling summary, shown as a "Conversation so far" header, so a late answer still knows the opening brief.
• Edit a message or try a reply again. A pencil on your turn opens an inline editor and Send replaces everything after it; Try again discards the last reply and asks once more. Both wait until Max has finished streaming.
• Max opens on what is waiting for you: your oldest overdue task, the next open journey module, this screen's last thread, and how many module changes landed since you last talked. A refused step names where it happens instead of retrying under another name.
• Video Studio cuts two cameras into one edit. Add a second camera angle and it is aligned by the transcripts both recordings already carry, with the evidence shown and a named reason when a pair cannot be matched. The AI proposes the angle switches, a devil's advocate flags the weak ones, you flip any shot, and the render cuts the second camera in over the main recording's sound. Proven on a real two-hour call.
• Video Studio: Replace the original swaps a recording for its rendered recut while keeping its tags, folder, name and id, with Undo. A Delete recording door sits beside it. The transcript now sits under the inline player, and clicking a line seeks to it.
• A swipe file for outbound email. Quick add in the rail takes a cold email that hooked you, as text or a screenshot, pulls the subject and opening line, drafts why it worked, and files it under the persona and pillar it fits. The ABM Swipe file tab lists them, starts with fourteen research-backed B2B patterns, and Draft for my persona turns a pattern into five openers with a critic's flags.
• ABM list building, first three waves. Every account link names its source and carries a qualified, no fit or unclear verdict you can overrule, so a below-floor company is no longer silently dropped. Add a list turns a CSV, spreadsheet or event attendee list into scored, deduped accounts. Per ICP, a Sales Navigator search is drafted from the ICP's own criteria with a reason per filter group, and its count becomes the SAM figure.
• Budget Mix: drop a spreadsheet or picture into the AI Targets box and the targets are derived from it, cited in the why, with the file shown as a chip. Workbooks are reduced to a digest first: summary sheets in full, raw sheets as headers and totals, contact columns never sent.
• Customer Interviews: Interview with this script seeds a Signal note under the script's own questions, and the note files as a response to that script, feeding the Interviews tab, Findings and ICP.
• Fundraising for admins: an executed SAFE now matches its incoming wire from the bank feed, with Confirm funded and Not this one, and the Form D clock counts from the first countersignature. Loan facilities drawn in tranches show each advance's accrual, and the signing flow gains a reusable offer letter where every candidate term is a field.
• AI operations for admins: the model catalog drafts new releases and retires dead ids on its own; a real-cost ledger stores what vendors billed against what was metered and proposes repricing; the prompts page shows each model's acceptance, health and cost per accepted output, plus an eval coverage meter weighted by spend.

Improved

• The ICP Pains detail panel gains a widen toggle that stays wide across pains and sessions, so the statement, buyer quote and cost basis are readable where they are judged.
• Connector setup docs and the account connections panel now include Perplexity beside Claude, ChatGPT, Gemini and Grok. Connector tools report Destructive only for delete, remove and merge, and every parameter is typed.

Fixed

• Mobile email-code sign-in works again: the app passes the human check the sign-in gate started requiring, and the sign-in email carries the 8-digit code the app asks for.
• Auto-refill cards saved before the billing account move get one re-entry notice and a pause instead of three failed-payment notices, and Add card opens checkout instead of erroring.
• Renaming speakers in a recording now re-scores Video Studio's "What it could become" card, so it speaks in the names you gave instead of raw speaker ids.
• Two-camera renders finish in production. The encoder reads sources over a local connection, and a render whose worker died now reads as failed with Render re-enabled instead of "Rendering…" forever.

Under the hood

• The React Compiler lint rules are adopted across the codebase, fixing about forty render-time ref writes at the source.
• T2D3's own assistant seat accounts no longer draw retention nudges or the beta weekly update.
• Verify and rotate on the credentials panel now authenticates payment keys before writing them, and refuses a live key in a sandbox slot.

Published

• Journal: A Pin Is a Floor, Not a Ceiling

New

• Max narrates its tool calls: one line before it reads, a bridge when direction changes, the answer after the last call. The status line names the running tool, and thinking shows as one collapsed row you can expand.
• Max's answer length follows the question. A lookup gets a sentence or two, a comparison a table, a draft arrives whole. It no longer introduces itself unprompted, and it says what it drew on.
• Video Studio takes raw footage directly. Drop a video on the canvas, it lands in the Signal Library as an ordinary upload, and transcription starts on its own. The row becomes editable when the transcript is in.
• Every recording is scored in the background for what it could become, weighed against your locked ICP, personas and value props, so the studio shows ideas without a click per video.
• Speaker names in a transcript: the AI applies only a confident guess and shows the rest as a proposal you confirm or edit inline. A name a person typed is never overwritten.
• Buying Triggers gains Auto-trim: merge near-duplicates first, then take off only the triggers no data source could detect. The list never drops below three, and a person confirms every take-off.
• Content Studio review columns: the pipeline board and the Backlog, Drafts, Calendar and Published tables show each piece's Value Rating and five axes, weakest tinted. Published cards open the editor first, and a channel filter narrows the view.
• Customer Interviews: the Interviews tab becomes a column set with Status, Quotes, Insights, Coverage and Foundation, so an interview that gave nothing to the foundation no longer hides in a list of names.
• Agencies get View as buyer on their roster listings: your live packages the way a client sees them, with the buyer price and the margin named.
• Admins can send a verified investor both SAFE tiers as Option A and B drafts, preview them, and send for signature. The template panel renders a sample with placeholders before any field is filled and names the source form and its counsel review date.
• Autopilot policies, wave one: steer proactive agent work per capability, module or project as auto, ask or off, with snooze and consent. Proactive jobs check the policy before they run.

Improved

• Back walks the whole journey. An open detail panel is a step, so Back closes it and a copied link opens the same row. Account, organization and billing settings open over the page you were on, and legacy links carry their panel and tab through the redirect. Checked in production.
• Quick add on Todos reads spoken priority and focus: "priority one", "urgent", "star it" and "add it to the focus zone" set the priority and the star instead of landing in the title. The create panel gains a Today focus toggle.
• The marketplace quality check is advice, not a gate. A refused listing keeps its findings and gains Publish as-is, which sends it to a human reviewer with the findings beside it. Reviewers can send a refused listing to the queue anyway, and the roster folds.
• Admin: pricing plan cards list every door and sub-door the navigation renders, and the email templates page folds the authentication previews so the sequences sit above the fold.

Fixed

• A plan-version pin is a floor, not a ceiling. Organizations pinned to an older plan snapshot now receive every feature the plan grants today; a buyer had lost the Marketplace People tab to it.
• Video Studio's New edit had failed for every entry since a migration made the creative format mandatory. A cut run now names its format.
• Publishing a listing that had gone live and was then unpublished was refused as "nothing to review". A re-publish is judged as a go-live, and only the first one waits for a human.
• Unpublishing a listing keeps the edits queued while it was live, applied to the draft, instead of dropping them without a message.
• Pushing Visual Brand to Google Docs told project members Drive was not connected when it was. The push uses the card's connection lookup, and a refusal offers Connect or Reconnect.
• The welcome email no longer promises a dated beta window. It says open beta and lets the launch date update itself.
• SAFE documents no longer print a draft stamp on every page now that the templates are counsel-reviewed.
• 3 automatic fixes from the app's own repair loop.

Under the hood

• Billing moved to a new payment account. Every price was re-created from one idempotent script and production rows were repointed in one transaction, so buyers saw no change.
• The bank write-back worker went live and booked its first batch. An under-scoped token is reported as out of scope instead of dead, and a rotated secret reaches the fleet workers without a hand restart.

Published

• Journal: The Status Code Answered a Question We Did Not Ask
• Journal: Nobody Decided That Back Should Go Home

New

• Email sequences for admins, built in four steps in one day. Duplicate any built-in email template into a custom copy that starts disabled and attached to nothing. A Sequences tab draws every series as a flow, left to right: the moment that starts it, how long it waits, which email goes, and what ends it. Each template says in a sentence what sends it and where it sits in its series.
• Sequences start from app moments, not the calendar: a sign-up, a locked module such as the ICP, a finished Compass. When the moment happens, the person is enrolled and the steps go out on their own schedule, and an exit event takes them out again. The built-in onboarding and feature-discovery series stay read-only, because their order is code.
• Draft with AI fills an empty event sequence from what that moment actually unlocks, in the house voice, with a critic's review shown beside the draft. Every drafted email lands inactive and attached to an inactive sequence. Switching mail on stays a human act, and each step reports its own sends and opens.
• Browser Back now walks the way you came. A picked tab, lens or section is a step in your journey, so Back undoes it instead of throwing you out of the module, a refresh keeps it, and a copied link lands on it. In-app Back links go where you actually came from rather than a fixed parent page.
• Agents working over the connector can suppress any ICP item, such as a trigger, criterion or pain that is not your ICP, with a reason, and restore it later. Suppressed items leave every downstream read, nothing is deleted, and a compact titles-only read lets an agent scan the whole module at once.
• When a bug reported in the Slack feedback channel is marked resolved, the app replies in that thread and mentions the reporter, so they know to try again.
• Admins can open any team course as a student would, or watch the whole course as one playlist, from the Academy admin page or the course editor. A learner who has not enrolled yet sees Enroll and watch all beside Enroll and Start.
• A video in a learn article's body plays inline with its poster, and the page ships a VideoObject description so answer engines can index the film.

Improved

• Team-interview drafts no longer land silently in a locked ICP. They arrive as an improvement proposal the owner accepts change by change, with a restore point as the undo, and a proposal that is already open is refused before any model call.
• The guided "Rank my claims" entry on a draft Value Props set shows on every lens, not only Themes, and a dismissal survives switching tabs.

Fixed

• When one AI provider's account hits its spend cap, generation jobs now fail over to the next provider and page the team, instead of dying as a bad request with the other providers idle.
• Normalize audio in the course editor works again on basic video assets; it had rejected every one with a deprecated-option error.
• Visitor analytics on www.t2d3.pro and the app report again. A build setting had left the tag off for every real visitor since the weekend.
• 1 automatic fix from the app's own repair loop.

Under the hood

• The product demo film has a music bed: two generated cues under the narration, ducked beneath the voice and normalized for streaming. The demo organization's reset now works from a marker written after each clean run instead of a hand-kept list of what every scene writes.
• The work ledger re-reads every open block against the systems it names and clears it only when all cited items happened after the block was written, and a launch that died no longer counts as launched.
• The bank write-back worker runs again on the fleet host, with a test that keeps TypeScript out of the worker's import graph.

New

• The home page of www.t2d3.pro opens with "Own the signal. Rent the models." and a band naming the three things T2D3 OS accumulates for you: a graded corpus, the human vote, and the loop that distills votes into rules. The engine counts on that band are read from the product at render time, never typed in.
• A new /agents page speaks to the other audience: your Claude, your GPTs and the answer engines your buyers ask all start with no context, and a locked foundation is the context they lack. Agents gather and draft; locking, voting, sending and spending stay with a person.
• The long module lists are searchable by what they say, not only by name: ICP customers (company, products, why they bought, churn reason, contact), Website Audit findings, named accounts and value props. Select all while searching selects exactly the matches, and the label says how many.

Fixed

• A background review job re-read a capped list of its own past verdicts, judged hundreds of items as never reviewed, and redrafted them every fifteen seconds. It now reads only the open items, so it stops spending the AI budget on work already done.
• 4 automatic fixes from the app's own repair loop.

Under the hood

• The product demo film is captured: twelve scenes at 4K, composed into a four-minute narrated cut and uploaded to the Signal Library as a recording Video Studio can cut from. The staged opening call is labeled as a dramatization with AI avatars, and film mode keeps the ICP economics test and the Anti-ICP lens off screen while recording.

Published

• Your team's disagreement is your best GTM signal
• Journal: Replaced Is a Fact, Not a Title

New

• Accounts belong to the ICP they were scored for. Each ICP's Accounts tab shows only its own list instead of one shared list under every ICP: a pull links to the ICP that ran it, deleting from one ICP unlinks it there, and an account can sit on several lists. Accounts this ICP never scored get a "Score them" offer with progress narrated.
• Account votes belong to the ICP they were cast on, so "Refine ICP from feedback" learns only from that ICP's thumbs. In the ABM module every account carries a chip per ICP list it is on, the facet rail gains an ICP filter, and territory coverage counts the ICP's own list.
• Example customers belong to ICPs too. The Customers tab lists this ICP's customers plus any nobody has placed, an upload stamps the ICP it ran on, a row can be moved or added to another ICP, and unplaced rows sit behind a banner where the AI proposes a sort for you to confirm. A customer can belong to several ICPs, never to all by default.
• The three-dots More menu is on every module. The shell draws the rows every module owes, Print / PDF, Share, Generate slides, the branded export and the lock controls, and each page adds its own above them. Ten pages lose their separate PDF, share and export buttons, and the seven modules that had only base rows gain their own: Ask Max, Build with AI, Export .md on Manuscript, Download brand kit on Visual Brand. Rows appear only when there is output to act on.
• A marketplace listing carries one reference link. Put the deck or document a buyer should open in its own field, and the card shows "Open deck" or "Open doc". The quality floor no longer reads a document link as off-platform contact and never asks what the deck shows, since it cannot open it.
• The quality floor judges agency packages fairly. A package carries its own role instead of inheriting its billing unit's, price is compared only against listings for the same kind of work, and a missing package block is offered as a fix rather than held as a refusal.
• Publishing a listing narrates its check on the button, and the floor's verdict stays on the card: each reason and fix on its own line, with Edit listing and Re-check. The toast is one line and points to the card.
• A SAFE generates from an investor's interest row in the Founder's Cockpit. The round's terms are data on the round, so a tier edited once changes the admin strip, the public invest page and its checks together. The agreement is built from Y Combinator's own form into a PDF and a Word file and hands off to the signing flow. The button stays locked until someone on the company's side records how the investor's accredited status was verified. A fully signed SAFE marks the row committed and drafts the cap table entry.
• An executed agreement that a later agreement replaces says so without being edited: a write-once "Superseded by" pointer, with a banner on the old record and "Supersedes" on the new one. Nothing is deleted or renamed, and a chain of replacements cannot loop.
• The company books write back to the bank. A confirmed booking's treatment, category and note land on the Mercury transaction itself, and a vendor cost can become a payment proposal a person approves at the bank. The bank answers one address only, so a fixed host carries a queue across every fifteen minutes.

Improved

• A locked ICP stops nagging. The "N pains owe a vote" strip gains Snooze for a day and Dismiss, and the header pill goes quiet with it. The subtitle still states the owed count, and the strip returns only when something new owes a vote.
• The library offer on ICP, Personas and Value Props lives inside the Pull from Signal fold rather than above it. Folded, the strip's subtitle is the verdict; when the module already has enough signal the offer is one line with a Review link.
• The ICP Overview buttons say what they do. Recompute is now Rescore and Re-distill is now Re-rank & rewrite, and both tooltips say that neither reads new files and that Pull from signal is how new signal comes in.
• Returning to the ICP paints at once from the copy you left, then revalidates behind it, instead of a cold reload. The Value Props lens no longer loops into a full reload.

Fixed

• The "your transcript is ready" email opens the recording's inspector in GTM Signal instead of the customer interview list.
• The AI team's own accounts no longer draw the customer onboarding emails.
• 2 automatic fixes from the app's own repair loop.

Under the hood

• The ICP reads from a third less data: the curation blob loads only when it changes, and a 143 KB fingerprint is a short digest.
• The company's finance pages read the in-app books; the QuickBooks integration is removed.

Published

• The 30-60-90 GTM plan you can start with zero inputs
• Journal: The Tick Is Not the Stamp

New

• The Agency Supplier Agreement and the individual Worker Agreement state the dispute process in full: a seven-day window for a client to accept, request a revision or dispute an order, the three grounds for a dispute, the response and decision windows, and the possible outcomes. Both documents carry a new version, so anyone who accepted the shorter text is asked again rather than bound by a clause they never read.
• Writing Habits shows the sentences your dials refused, struck through, each with the habit that ruled it out, so you see the rule at work on real text instead of trusting a slider label.
• Approving a content idea now also plans the charts and illustrations for its body sections, and the idea page shows them as they arrive. A decorative-only pass is pruned before it reaches you.
• A newsletter edition can be previewed in the app exactly as the reader's email will render, and editions appear as chips on the content calendar beside the pieces they carry.

Improved

• The ICP's "Beyond the brief" list keeps its answers across regenerations. Voted, dismissed and snoozed traits are never re-proposed, the model is told what the draft already says, and restatements of decided traits are dropped. The list folds to one line, opens by itself only when a run adds something new, and offers bulk dismiss and snooze. A snooze lasts until the next run.
• A lock on an ICP is a stamped version, not a freeze. The pill names the version, the lock date and how many changes arrived since; edits and votes stay open and ground downstream right away, and re-locking stamps the next version. The copy that claimed otherwise is gone.
• The ICP Overview's five tests live in one place: larger clickable pills in the scorecard band open each test's verdict and rationale, with a link to the tab where the fix lives. The duplicate five-card grid below is gone.
• Suggested target phrases that the workspace measured in the last 30 days reuse that reading, so only new phrases are paid for.
• The archive-sweep note is now called Signal library upkeep and sends from its own address rather than the digest's. An organization's first sweep is silent by rule. The note says how many verdicts a person cast and how many a rule cast, and carries a one-click way to put the whole sweep back.

Fixed

• A brand that uses one typeface family for both headings and body no longer fails the guided brand wizard's identity check with "no typeface has role body". The two brand wizards now cover the whole viewport instead of letting the page show through at the bottom.
• Inviting a whole bench to the marketplace at once no longer loses a share of the invites to the mail provider's rate limit. Every outbound fan-out is paced, and the roster says "N sent, M could not be delivered" instead of reporting only the successes.
• An audiogram for a brand whose logo is an opaque app icon shows the mark as a reversed tile instead of a blank white square.
• Mission Control no longer marks an item "Queued for fixer" when the nightly fixer cannot take it. It says why in place of the queue button, and an order whose fix already merged no longer stays stuck on queued.
• 2 automatic fixes from the app's own repair loop.

Under the hood

• The nightly signal evaluation no longer stops at the five-minute mark.
• Session memory and seat state for the AI builders moved from per-machine files into database rows, with the files kept as mirrors.
• Demo film: one capture script per scene reads its numbers off the screen during the take, and the illustrative recordings come from the real transcription and clip pipeline, flagged as synthetic wherever they play.

Published

• A website audit scored against your buyer, not generic standards
• Journal: A Mirror You Still Write by Hand Needs a Way Back

New

• The ICP's Trigger tab becomes Reach: when (buying triggers) and where (watering holes) together, and Density becomes Size. The Trigger test reads "wired" from the integrations the workspace has connected, says why it is yellow, and no longer nudges toward fewer triggers.
• The Pains tab votes in one place: a one-line strip whose Review button walks only the pains owing a vote, and "Move to…" is back for a pain that belongs elsewhere.
• A locked ICP applies the changes you accept from a proposal, re-locks, shows the Five Tests before and after, and keeps the previous version to restore. A card built on older evidence flags the files uploaded since, and one click re-checks it and proposes, never overwrites. The Build-with-Max door stays on a locked ICP and asks before unlocking.
• Documents the app wrote itself no longer count as full-weight signal. The library detects AI-written text on every intake path, halves its weight once, shows a reduced-weight chip, and leaves the Signal, Noise or Authoritative call to you.
• Any member can connect an integration that is not yet connected. Only the organization owner can change or disconnect one that is.
• Suggested target phrases arrive with search volume and difficulty. A long question with no volume is marked as an answer-engine target rather than left blank.
• Brand Voice can blend the best of both A/B rewrites into one sample, with notes on what it took from each. You edit, keep or discard it, and the distiller learns from the verdict.
• Max's voice is yours to pick: six voices with a preview each in Settings, under Voice and dictation.
• An idea draft in Content Studio can show its generic twin beside it, the same prompt with every grounding block removed, with a difference score, four quality scores, persona fit and lineage.
• A newsletter edition in review can unlink one piece from its slot without discarding the whole edition.
• Sending a survey or a Brand Voice invite can pre-authorize one automatic reminder to each silent invitee inside the final 24 hours. The checkbox is on by default and says what it approves.
• A mastery ladder, in beta. Points come only from real work, short lessons appear at the step you are on, and a decision untouched for three days gets one nudge where "No follow-up needed" is a valid answer.
• The Website Builder treats kept GTM Diagnostic prescriptions as must-address criteria: each is placed on a page the critic checks, or declined with a reason.
• An Academy course plays as one playlist on a phone. Signed agreements too large for one upload, such as a scanned contract, go straight to private storage and are re-checked by the server before being recorded.
• The public team page lists founders with one-word roles, all seven AI colleagues and a launch partner. A new people colleague handles candidates, paperwork and onboarding.

Improved

• The Signal library's Table layout shows each file's parse status, and the funnel counts move the moment a document is voted.
• The member roster's GTM-role suggestion is one line, role plus Confirm, and "Head of Customer Success" now gets a suggestion.
• The This client, All clients and Prospects switcher sits top-right on every scope.
• Mission Control applies a decision brief's recommendation after a 24-hour window for a founder to steer it, and a request that ends without a pull request shows one state everywhere.
• The marketing site counts a sign-up start even when the sign-in redirect comes first, and the GA4 panel shows key events per channel.

Fixed

• Voice mode listens again after Max finishes speaking, instead of saying "Listening" while hearing nothing.
• An agency with the unlimited-client-workspaces override can reactivate a parked client without a subscription or add-on charge.
• A HubSpot connection scoped to one brand no longer receives portal-wide authors, tags or campaigns, and never publishes to a guessed blog. Connecting HubSpot completes on whichever app host it started on.
• Background imports and interview processing check that the source upload belongs to the destination workspace and honor private files, closing two privacy gaps found in review.
• Mail to the AI colleagues' addresses was dropped for about ten hours after their handles were reserved. Delivery is restored and the dropped mail was replayed.
• 4 automatic fixes from the app's own repair loop.

Under the hood

• The repair loop's bookkeeping was audited and repaired end to end, and a nightly watch names any item stuck between stages with nobody assigned.
• Build fleet: a rebooted host returns itself once its memory check passes, and required builds no longer queue behind browser jobs.

Published

• Build a marketing budget that argues with your pipeline math
• Journal: A Score Nobody Can Turn Green Is a Nag, Not a Measurement

New

• Talk to Max. Tap Voice, speak, and the answer is read aloud sentence by sentence while it still streams. Speaking over Max interrupts him, and the status row says what he is doing. A spoken "yes" never approves a confirm card; destructive and outbound actions still take a tap.
• Content attribution, end to end. Every place a piece is published gets its own tracking code on the links that leave it. Deals from HubSpot, a CSV import or hand entry land as one deal book, and the Performance tab shows, per piece, the deals influenced, the pipeline and the won revenue. After each sync the app mines closed deals for segments that win well above the baseline and for drift from the locked ICP, and proposes changes, each with a devil's-advocate objection. A shared HubSpot portal can be scoped to this workspace's brands.
• The marketing site records a buyer journey in Google Analytics 4: the tag runs only on production hosts, each stage has one event, and the server reports sign-ups and payments so the channel gets the credit.
• Agreements are managed in the app. A legal admin area, granted explicitly and never by group, opens three doors: create an agreement with signers in a signing order, record one already signed on paper or in another tool, and void anything unsigned. No Social Security number can enter the capital record, and a holder can file the Certified Mail receipt for an 83(b) election on its page.
• Two more AI colleagues join the virtual team. A pricing and packaging lead owns the feature matrix's pricing ladder and says which value metric and documents it read. The GTM lead proposes one scoped re-draft when a plan drifts from the company's stage or ARR target, and ends a workshop with an owner-and-date map. Both only propose.
• An ops colleague turns stalled programs, open pull requests and standing alarms into one ranked list of asks, each with an owner. Every AI colleague distills its week into at most three lessons, and a lesson that keeps holding up is proposed for its charter.
• The AI colleagues are two-way on Slack. A reply in their thread, a direct message or a mention reaches the colleague, and it answers in the same thread under its own name.
• The app header shows a live AI credit ticker that counts up when a charge lands, with the month's use and the last charge on hover.
• A blog draft in Content Studio shows how well it fits its persona, scored blind against what the same model writes from the topic alone, so the value of the company's signal becomes a number.
• The Writing Habits dials in Brand Voice rewrite a live sample paragraph as they move, with each change named.
• Accepting an ICP improvement proposal applies the changes you kept and rescores the ICP in place. Profile criteria, triggers and watering holes show the quoted line they were pulled from, as pains already did.
• Newsletter sign-ups are double opt-in and land only after the inbox owner confirms. A bot check gates the newsletter form and the calculators' "email me this" captures, and junk-flagged contacts never receive a send.

Improved

• Publishing to HubSpot from Content Studio offers the portal's own authors, tags and campaigns, and a campaign can be linked to its HubSpot campaign. Publish live stays the default with draft as the opt-in, and the Website Builder can save pages to HubSpot as drafts.
• The Signal document-type list is one grouped picker: two new collateral types for long-form content and sales assets, and five near-duplicate types merged into their neighbours.
• The Ore tile on the Signal funnel opens the named list of files still being worked on, with the step each waits on, and the stuck ones lead.
• A document's customer segment shows in the document panel and can be changed from the review queue's bulk bar, and a same-conversation strip has Merge all.
• The confidential-documents banner names the documents "Make all private" will act on before anyone clicks.
• The credentials admin page gains search, filters and a sortable table in triage order with a colour per row, and a card says what is wrong with a credential's setup and the fix.
• Mission Control's "decide" link opens only the decisions, the page speaks one state vocabulary, and the Owner column says whose move it is.
• Audiograms render in the organization's own Visual Brand palette and heading font instead of T2D3's.

Fixed

• Accepting a long ICP pain no longer fails on a label-length cap, so the verdict is captured.
• The Brand Voice source picker says what it lists, names the newest documents, and can add any Signal document.
• A newsletter intro keeps the paragraphs, bullets and bold labels its author wrote, so the inbox matches the preview.
• The Company Library's "waiting for a vote" count matches the review queue: a failed parse no longer inflates it, and a recording still awaiting a vote under an already-voted note is its own row.
• 1 automatic fix from the app's own repair loop.

Published

• How to turn an ARR target into a portfolio of growth bets
• Journal: The Seat That May Only Ask

New

• T2D3's AI colleagues now remember. Before a run, a colleague recalls its open commitments, the decisions already taken, its standing lessons and what it knows about the sender and subject of each message. After the run it writes down what it communicated, decided, learned and promised, so the next run starts where this one stopped. A colleague whose memory is unreachable says so and carries on rather than failing. The marketing lead's written record is now an export of that memory.
• Two more AI colleagues join the virtual leadership team on the marketing lead's recipe. A finance colleague reads bank, cash, AI margin, recurring revenue and the billing ledger, and every figure in its brief names where it was read and as of when, or is marked unavailable with a reason. A support colleague is measured on first response time, time to resolution, reopen rate and backlog age over a 28-day window, each with its source and window named. A silent zero can never reach a reader. Both are built and stay inert until a person switches them on.
• The AI colleagues attend the weekly Signal meeting. Each files the issues it sees and a headline before the meeting, is listed beside the people, and afterwards remembers what the room decided on its items. A headline filed before anyone has started the meeting is held on the project and joins the meeting the moment someone opens it.

Improved

• The marketing dashboard reads Google Analytics 4: sessions, users, new users, engagement, key events, a daily series, and top channels, landing pages and sources. It also checks whether the tag every page sends belongs to the property being read, so an undercount is told apart from a traffic drop. When Google refuses the read, the reason names the actual fix, the Analytics Data API switch or the property grant, instead of always blaming the grant.

Fixed

• When an AI colleague answers an email, the T2D3 colleagues who were copied on the thread stay on it. The reply used to go to the sender alone. Addresses are still taken from the stored message and never from the email's text, and addresses outside the company are still left off.
• The support reopen rate could only ever read zero, because leaving the resolved state cleared the resolved stamp and a reopened ticket looked like one never resolved. Reopens are now counted, and the report says from which day the count begins.
• 2 automatic fixes from the app's own repair loop.

Under the hood

• One script creates an AI colleague's production identity end to end, idempotently, and each run of it is recorded in the operations ledger.

Published

• The Verdict Was About the Whole Library. The Read Stopped at 200., Wren's weekly.
• Journal: The Safe Reply Answered One Person

New

• Core Signal opens on four stage lenses, Needs Review, Signal, Noise and Glint, each with an honest count, and the lens survives a refresh or a shared link. A to-do banner says how many documents are new, how many need a person and about how long the sitting takes, with two actions: confirm the confident ones or review the ones that need you. The Examples tab is now Inspiration.
• One "Add signal" button names every way to bring signal in, grouped as Bring files, Bring words, Record and Connect, and each row says what happens next.
• The Signal lens carries a module map. Pick a module and the list shows the documents that feed it, and each refined draft reads Fresh, Locked or May be stale with what it was built from.
• The Audit lens is now the one place for what is wrong with your signal. A badge on the tab counts the required sources still missing, a Fix first rail names each problem with the door that fixes it, and the health cards and agent reports that sat above the library live there.
• While documents wait on a vote, the vote is the screen: the document list and housekeeping strips fold, say why, and open again once the queue is clear. A Settings & integrations disclosure gathers folder rules, the learned pattern, the review goal and the connect doors in one place.
• The Noise lens groups documents by why they are out, and every row offers Back to review or Keep as Signal. Noise older than a threshold your organization sets moves each week into an Archive group with a Restore button, and the owner gets a note naming what was archived.
• Confirming a slate of recommended verdicts now has a way back. A strip names the batch and Undo returns every row to review, and an undone confirm no longer counts as agreement with the recommender. In the reviewer walk, S, N and G cast a verdict and advance, and U steps back.
• Conversations voted to Noise before anyone mined them get one look. Scan them from the Noise lens, ten at a time with names shown while they are read, and any that hold a claim are offered Restore as Glint. Every library row gains Ask Max, which opens the chat on that document with the question already asked.
• An inquiry you save as a document now counts as Signal and grounds the modules. A receipt names the document it became, the modules it feeds and links to it.
• Your organization sets its own Signal rules: how many documents a review sitting shows, how much human review it wants, who may vote, and whether the learned pattern may sort on its own or only propose.
• Triage rules cast Noise on arrival. State once that a kind of document is Noise and the rule files each new match, never casts Signal, never overrules a vote and never counts as reviewed. A panel shows what your rules did in the last seven days, with Undo on each run.
• Every classification now reads a document for confidentiality. Pay, NDA and personal conversations the classifier is confident about are made private at once, and only the uploader sees why and can share them back in one click. Lower-confidence cases are a proposal to the uploader alone, and the document holds out of org-wide grounding until they answer.
• In the ICP density lens, the list of documents that could feed ICP can be dismissed, and new uploads bring it back. Accepting a parked pain un-parks it and moves to the next one, and a settled pain reads Accepted.

Fixed

• Verifying a document as still true now lands on the row instead of vanishing on reload, every discarded state reads as Noise, and a dismissed inspiration stays Noise when it is re-profiled.
• Mission Control no longer keeps merged work in its review queue after a rate-limited read, and the Requests table keeps its pills on one line.
• Closing a Brand Voice dialog no longer leaves the page unable to scroll.
• Disconnecting a contact center connector and deleting a library document each leave an audit entry naming who did it.
• 1 automatic fix from the app's own repair loop.

Under the hood

• The web framework moved up two releases, closing a published security advisory.
• The connector for AI assistants can read a long transcript in pages instead of only its opening.

Published

• How to turn a feature matrix into packaging, pricing and positioning
• Journal: The Read Stopped at a Thousand Rows and Said Nothing

New

• A module now grounds in its best documents across the whole Signal library, not the best of the newest 200 uploads. Each document is ranked per module on the classifier's confidence, its outcome history in that module and, once the foundation is locked, how closely it fits your ICP, personas and value props.
• A "Feeding" lens in the Signal library shows the documents the next generation of a module will read, the ten on the bench behind them and the ones kept out, with a reason line per row and Pin, Exclude and Reset on each. Rows that entered the window since your last visit are marked, and a window-health score says how graded, verified and human-decided the evidence is.
• Once a module's library passes a threshold an admin can tune, a weekly steward proposes what to pin and what to keep out, one reason per line, with a devil's-advocate note on each before anyone sees it. Accept or dismiss each line in the lens.
• Tagging a document for a module now makes that module read it, however old it is. Every reader asks the tag first and the recency window only decides what else rides along; the Signal pull, the Voice Kickoff inventory and the Brand Voice build share one membership rule instead of three.
• ICP, Personas, Value Props, Messaging and Brand Voice each say whether the library holds enough for them, judged over the whole library. When a module is short, it names the documents already uploaded that look like they belong, pre-selects the strongest and tags them in one click. Brand Voice gives one verdict instead of a green line above an amber one.
• Transcripts name their speakers. Every completed transcription proposes names where the conversation gives clear evidence, never overwriting a human rename, and a "Name the speakers" button in the Signal inspector fixes the rest. Renames reach the transcript text itself, so copy, download, search and every grounding prompt quote the real names.

Improved

• Duplicating a set warns when the Signal library holds files newer than the evidence the set was built on, so a copy no longer looks fresher than its source. Recompute and Re-distill carry tooltips that say what each does.
• The Brand Voice by-format habits grid reads as editable: the fixed identity rows sit under their own header with the reason, inherited cells show in normal ink with a key, and cells that do not apply say why. Casing rules get a guide to the vocabulary and can be scoped to an element such as headings.
• The readiness matrix counts a document a person tagged for a module, not only the classifier's guesses.

Fixed

• Documents voted Noise or Glint, and superseded duplicates, no longer ground Engagement Health or the Competitive scan.
• Calls that arrive through a contact center connector are credited to the connector, not to the person who connected it.
• Accounts that never confirmed their sign-up no longer receive the welcome and onboarding emails, are left out of the signup counts in the digests and the admin dashboard, and are purged after 48 hours instead of a week.
• 11 automatic fixes from the app's own repair loop.

Under the hood

• Two dependencies with published security advisories were updated.

Published

• Three ways to tell AI what on-brand means for your visual brand
• What it means to work at T2D3 OS: 13 expectations to settle before you join
• Journal: The Gate Was in the File. It Was Not on the Read.

New

• The Founder's Cockpit keeps the company books on real cash. An hourly Mercury bank feed brings in accounts, balances and transactions; the AI proposes a bookkeeping call for each, a devil's advocate challenges the calls that move margin or acquisition cost, and the founder confirms. Each correction becomes a rule for that counterparty.
• Confirmed bookings roll up into a monthly cash-basis P&L and cash flow, tied to the bank's own closing balance with the difference shown. Closing a month freezes the statement; a month that does not tie out closes only with a note. The QuickBooks mirror is retired.
• A contact center connector brings recorded calls into the Signal library on their own. Org Settings provisions private storage per org in one click, each finished call is recorded, paired with its audio and sent for speaker-separated transcription, and arrives tagged "Automatically added recordings". The card shows the last call, counts by state and minutes against the monthly cap, and an admin can mint and revoke a view-only link for people with no account.
• A merge of two ICP items can be taken back within 24 hours, votes and comments included, by the person who made it. "Keep separate" is remembered across re-runs of Find similar, and a Recent decisions panel lists every recorded change with Undo where it still applies and the reason where it is withheld.
• The Founder's Cockpit gains a Board records tab: a decision register with signed files, append-only once signed, where a later decision supersedes an earlier one. The documents library groups paper by holder, a ledger row can attach its document, and a buy-back right that lapses on milestones is modelled separately from vesting.
• A public Trust & Compliance page gives procurement teams the DPA terms, sub-processors, security controls, AI posture, incident response and end-of-engagement data handling in one place.
• Agency operators get a per-client monthly usage statement of billed AI calls, as CSV or JSON, and Agency Pulse shows AI cost per client. An operator's own framework names replace T2D3's on every page their clients open.
• Self-serve deletion of an org enters a 30-day grace window instead of erasing at once.
• Competitor cards in Value Props lead with a one-line "N of M claims matchable" verdict, sort by threat and show only the red claims by default. The admin users list shows when each person was last contacted.

Improved

• Every overlay closes on Escape, and a stray backdrop click no longer discards typed input.
• One red per screen: selected toggles, chips, checkboxes and tabs go charcoal or underline, so the remaining red is the primary action.
• The GTM Diagnostic re-run reads your live site instead of a days-old crawl, and a recommended move can be marked done so completed moves stop reading as open.
• The Signal page's Ore count includes only files that will be graded and says why it differs from the Library's "parsed" count. Market-label proposals arrive in under a minute instead of two.
• An agreement opened from Admin leads back to its list, and a viewer who cannot sign is told which account the signature awaits.
• When a mail gateway holds a magic link, the "Check your email" screen offers Microsoft or Google sign-in instead.
• A long competitor URL truncates in its chip so the remove button stays reachable, and "Adjust this map" edits the saved positioning map in place with live scores.

Fixed

• A reviewer's slug and meta edits on another project's piece no longer fail with a 500.
• "Not a duplicate" clears the badge even when the duplicate scan times out, and near-duplicate detection no longer times out on libraries past 1,500 files.
• A timed-out membership read no longer bounces a project manager off the team work-styles page.
• Locking an ICP over several persona sets runs in a handful of statements instead of one per row.
• Ten regressions caught by the nightly run are closed, each with its test kept as a guard.
• 3 automatic fixes from the app's own repair loop.

Under the hood

• Connector AI spend is tagged on every call record, and tools that spend AI budget get their own 20-per-minute limit.
• A locale's translations load in one round trip instead of about 22, and batched AI jobs stop abandoning their batch when their grounding drifts between polls.

Published

• The "only we do this" claim that didn't survive the competitor audit
• Your brand voice should be researched, not asserted
• Journal: The Cap Table Has No Balance Column
• Journal: When the Corrections Disagree, Learn Nothing

New

• The virtual team takes on more of the writing. The presentation specialist argues the client report and QBR deck with an action title and a so-what per slide, checked by a critic that drops any title it cannot defend; the deck page shows what changed, and the client share link shows none of it. The same specialist drafts the locked-plan document when a GTM Plan locks.
• The campaign strategist drafts the Budget Mix channel mix when targets land, refuses to draft until at least one target is set, and marks each card she drafted. The conversion copywriter authors the Pain-Claim-Gain cells of the Messaging Framework under his own standard and byline, with every human edit kept as a diff, and answers a locked Website Audit finding or a competitor's copy move with a grounded rewrite proposal.
• The outbound copywriter challenges persona pains and triggers against what buyers replied, counted on the persona brief. The verbal-identity specialist diagnoses every piece that drifts from your brand voice: the passages, the rule each breaks, and the case for leaving it alone. The voice-of-customer researcher reviews the interview synthesis before it reaches the ICP, signs the scripts she drafts, and turns unevidenced persona claims into a research agenda.
• Agent suggestion panels open folded. Each one is a single line naming what waits, with Show, Snooze a day, or Dismiss after a confirmation, across the ICP checklist and improvement proposal, every virtual-team proposal list and the module-owned suggestion cards. Module content stays above the fold.
• The Founder's Cockpit gains an equity and capital ledger. Every number is derived from dated, document-backed entries that cannot be edited once posted; the cap table shows both dilution bases, vesting, plan headroom and a compliance panel. A company's history seeds from a file, a signed document drafts its own entries in one private call, and 83(b) elections are drafted from the ledger, reviewed, and emailed only on an explicit confirm. The old cap-table tab is retired.
• Video Studio Motion formats, a product explainer, a kinetic-type teaser and an animated stat, draft their own script and storyboard from your locked foundation. A scene editor re-renders only the scenes you changed, in three aspects. One set of screen takes becomes an edited demo plus reels, with every private region blurred until a person rules on it.
• Onboarding and engagement emails arrive in the reader's language across six locales, and an English edit flags the translations it outdates. The email template editor gets an explicit Save, a numbered version archive with Restore, and an A/B test that compares journey progress per version.
• Admins can delegate every area except admin permissions, including Reply-To per purpose, onboarding email routing, prompts and the API docs. A View only preset sees every delegatable area and changes nothing.
• The Growth Stage interview docks beside every My Journey tab, and the numbers the team gives in conversation land on the Calculator.
• The feedback form's "I need a response" box now means it: unchecked, no follow-up emails arrive and the item closes on the fix.
• A managed client on a client-pays agency package pays for itself once it subscribes; until then the agency pays, and the engagement card says so.
• The Engineering Intro course gains 27 narrated lesson videos, each with a diagram still.

Improved

• ICP Density has one sizing button, a funnel that says how each number was counted, live totals beside the criterion toggles, and a territory plan that quotes the page's own numbers.
• Signal library strips name the files they auto-filed or left untyped, share one neutral color, and a Segments filter sits beside Tags. Proposals now read "market segment labels".
• The GTM Diagnostic's grading rubrics are editable prompts, so an admin edit reaches the next assessment.
• A Visual Brand foundation build finishes even if you leave the page.
• An AI Build call in Mission Control builds itself: one "Build it", coding agent first, nightly fixer as backup.
• Every public surface, legal page and agreement names T2D3 Inc., a Delaware corporation in McKinney, Texas, and admins re-accept the agreements against the new party.

Fixed

• Running the campaign strategist no longer times out halfway through her brief.
• The Content Studio review shows the byline a piece publishes under, and says "Not live yet" until its post has deployed.
• The erasure Confirm button, and an ICP save button, were invisible because a brand-red class painted nothing; both are back.
• Copy validation cells no longer run out of room on the model's own reasoning.
• A Website Builder build no longer dies in visual QA after the site is built, and a cancelled build reports failed instead of showing an ETA forever.
• Mission Control's parked items get their decision briefs again after weeks without one.
• 7 automatic fixes from the app's own repair loop.

Under the hood

• AI call records keep each resume count and the prompt behind every usage row, and the nightly canary watches the top-cost lanes per model.
• A build or inference host that goes dark now pages a phone within minutes, and the runners tab says when its picture is partial.

Published

• Journal: The Guards Watched the Decision. Nobody Punched the Ticket.

New

• Max AI becomes a way to run the app. It runs the same governed tools the connector uses, moves your screen to the table you describe, filtered and sorted, and runs any module action you could click yourself. Publish, export, send, an invite or a role change stop at a card you confirm; a delete shows what it removes and runs only from the click; every write can be undone. It files documents into Signal, moves Audience contacts between stages, proposes a newsletter send naming the recipient count, and launches a guided skill in the thread. A file dropped into the chat lands in Signal by default or is read once and kept nowhere.
• Content Studio is five lenses in the order the work happens: Ideas, Drafts, Calendar, Published, Performance, landing on Ideas. The old Inbox folds into a Backlog chip where every idea says where it came from and shows its Create, Refine or Skip verdict, with validation in place. Performance links each target query to the piece aimed at it and what it earned, and offers to draft an idea for a gap. Every verdict in Validate ends in a next step: draft it with the feedback in the voice you pick, save it, or start over. Compose gets a searchable value-prop picker, and search and filters move above the Drafts list.
• The connector reaches every door in the app, with new read tools for the inbox, Relay, meetings, dashboards, campaigns, competitors, decks, Talent, the Partner Hub, Level Up, agents, skills, audio clips and Video Studio. Locks, votes, sends and spend stay in the app. Every module action is ruled dispatchable or withheld with a reason.
• Newsletter editions can be approved and scheduled: pick a live audience and a time, confirm the count, and the send goes out in chunks at that hour. A planned-issues backlog holds the editorial calendar, and a reader who replies "unsubscribe" is opted out like the link.
• The Growth Stage interview works for a leadership team: it asks each person their role, hands off between teammates, records a disagreement as a question for an owner instead of overwriting, and keeps one organization-wide stage that members propose and an owner confirms.
• Four more virtual-team members file proposals: image briefs, a pain slate quoted from real buyer questions, a survey base-quality verdict, and book concepts. The campaign brief starts from what the strategy workshop settled, and the deck recipe names who proposed it and learns from what you adopt.
• Video Studio reviews each take as it uploads: a critic reads sampled frames against the step's brief and says what to re-record.
• Website Builder sites gain a post feed section and case-study pages. Agency packages record who pays for T2D3 OS, the agency or the client.

Improved

• The ICP lens row follows the order of the exercise: Customers, Pains, Profile, Triggers, Overview, Density, Anti-ICP. Value Props gets its six-step tracker back, and the lit market reads its brief back under the pill strip.
• The signal library speaks the refinery vocabulary throughout: Ore and Concentrate lenses, documents assayed, Waste, and the Lode as the asset you own. Prompts grounded on the library now also receive your confirmed signal, the market read, your SWOT and your own competitor edits.
• The menu opens on the surfaces your signup role works in, and the command palette finds surfaces by the names Interface v2 moved away from.
• Admins get one saved Advanced switch that restores the dense Users table, the home counters, the full domain rail and every Feature Status column. The clean screens stay the default.
• A signup refused for a personal email address is kept as an audience lead and told so; its leftover account is proposed for erasure instead of counting as a user.

Fixed

• A tab left open across a deploy no longer stops navigating silently; it reloads onto the new build.
• AI visibility probes that paused mid web search came back empty; the turn is now resumed. The citation audit keeps its Sources list and the content planner stops thinking aloud, the last two cut-offs after the cap raise.
• Confirming an ABM jurisdiction re-arms only that program's sequences; the competitive digest goes to team members only; one confirm buys one batch of brand example images; an AI style-rule draft no longer rewrites a locked brand voice until a person applies it.
• Ask and Compose grounding no longer compares a company with itself, and a value-prop or persona change marks the Messaging Framework stale as it should.
• 10 automatic fixes from the app's own repair loop.

Under the hood

• Ops alerts split into three owned lanes with one daily digest each, and a persisting condition escalates to a named owner.
• The coding agent lands its own tooling changes, runs on the fleet's own subscription seats, hands a red check back to itself once, and closes a request it proves already fixed.
• Two signal crons dark for weeks are back, and a second host serves the mid-tier model.

Published

• How to build B2B personas with dollarized pains that sales will use
• Journal: One Green Run Is a Sample, Not a Proof

New

• Admin Studio completes its redesign. Feature Status states, for every feature, who it reaches and how many organizations that is, proposes the one beta promotion worth making, and a flip that reaches live users names its consequence before it commits. A delegated admin sees only the tools they can open, with a reason and a request path for the one action they cannot take. The command palette on admin pages becomes a tool picker that shows how much each tool is used. Under 900 pixels the rail becomes a sheet, the queue a single column and the users list a set of two-line cards.
• Visual Brand compares two directions as a state of the page, not a modal. Both directions show at specimen size with their own fonts, palette and logo, the page says what the choice decides before the click, and one Choose Direction button follows whichever direction is focused. Back returns to the kit at once.
• Visual Brand's Examples view is a reading desk. Each reference is one row naming what it taught, what it informed and who judged it, with the references still waiting on a verdict listed first. Accept or Dismiss per row, or accept all, and the strip updates in the same frame.
• A stakeholder opening Visual Brand now sees the whole locked kit instead of a Forbidden page. A chip names their role and only the verbs that would write are greyed.
• Content Studio gets a visible Newsroom door in its header, and the Ideate view opens with a Targets lead-in, so you pick the queries you want to win before you ideate against them.
• The signal library renames its bottom tier. Waste is the discard: ignored, superseded and outdated documents. Tailings is the non-consensus pond: documents a human marked as a glint, kept and searchable, weighted zero for foundation modules and weighted up for content. Content and Campaigns generators now receive that pond as a separate outlier band in their evidence, marked never to cite as consensus, and it reaches glints anywhere in the library rather than only the newest two hundred uploads.
• The output cap an admin sets on a prompt in prompt administration is now admin-owned. It used to share a column with the production seed and was silently reverted within minutes of the next deploy.
• The public site says what shipped: no hard count of systems, a reframed line about decks that assemble themselves when you need one, and Brand Refresh listed in beta.

Improved

• The curriculum generation wait estimate says about nine minutes, measured from a real production run instead of a development-queue figure.
• Auto-refill treats a 3-D Secure challenge as a request for your confirmation, not a card decline, so a good card is no longer paused after three challenges. Paused refills are visible in admin, and the existing guardrails are pinned by tests: pause after three consecutive declines, at most one charge per day.

Fixed

• Several AI outputs were still being cut short after the recent cap raise. Four prompts whose code-side floor sat below their output contract, including ICP distillation, are sized from that contract; three newer critics get the same answer floor as their peers; high-effort thinking gets its own reserve so reasoning no longer consumes the answer budget; and the Messaging copy critic stops paying for its grid cell twice.
• 5 automatic fixes from the app's own repair loop.

Under the hood

• A new merge gate binds every prompt's registry output cap to the JSON shape its prompt declares, and a daily canary watches production for answers that stop at the length limit, so a cut-off that arrives without a code change is seen the same day.
• The dead-code check is now a required check before a change can merge.
• The agent fleet's own bookkeeping tightened: a crashed unattended session is closed by a daily reaper so the ops dashboard counts sessions that exist; a wave marked ready starts itself; every handoff must name one host machine and a handoff prompt can be consumed once; date gates written as plain sentences are honored; a stopped runner is named at session start only when work is waiting on it; and a finished program removes its own row from the shared memory index.
• Test runs on the shared build hosts stopped flaking. The job serialization gate survives a directory owned by another user, the memory watchdog no longer freezes a test runner mid-run, and a test whose fixtures aged against the real clock is pinned to a fake one.
• A transitive dependency of the website-crawling SDK was updated, closing twelve security advisories on a path that fetches untrusted customer websites.

New

• Ideal Customer Profile matches its approved design in every state. Criteria read as facets with real counts, accounts and pains as dense rows, and one action row ends in a single Accept and re-lock. When an upstream module re-locks, the ICP stays locked and a banner proposes the refresh. A never-locked ICP teaches what it claims and names the signal it will draft from, and a scoring run narrates itself instead of spinning.
• Value Props matches its design. Every claim carries its evidence and the competitor it beats in the row itself. Accept, Edit, Dismiss, Combine and Park sit on the surface with one Lock and one Push.
• Home and the navigation rail match their designs. Home keeps four bands: where you are in the journey, one next move, what the agents ran, and eight doors. The rail reads Home for everyone, names the active project and groups the Doors.
• Content Studio matches its design across all eight frames. One piece set shows through four lenses, Calendar, Drafts, Ideas and Published, with counts that cannot disagree. List lenses are a stage rail beside a dense table with bulk actions; the Calendar lens is a month grid with drag to reschedule. The selected piece opens as a judged card showing its brief, what it is grounded in and how it was judged. Checking undrafted ideas and pressing Draft N writes them in one batch, with a narration line naming what is in flight. An unlocked foundation teaches the dependency instead of showing an empty calendar.
• Website Builder opens on the site itself. A standing chip (Published, Draft, No site yet, Build failed) replaces the lock chip, the canvas lists every page with its copy, images, publish and judgment columns, and a stage line names the next step. Setup dialogs move out of the way and open by name from the command palette. A failed or stale build outranks published, so the one red button becomes Rebuild.
• Visual Brand matches its design. Identity opens on a specimen sheet: each colour, type scale, logo size, icon and photo treatment is shown as a real specimen with a judge band under it. The header reads which version is locked, what a draft changed against it, and what would go stale. An upstream re-lock shows as a staleness flag with one Re-lock brand action and a was-this-right seam. An empty kit offers two visible paths, import or guided build, and writes nothing until one is chosen. A failed import is named once above a kit that keeps rendering.
• Graphic Studio matches its design. The library is a table showing where each asset is used, how often it is reused and its state, with a kit-version facet. The strip measures whether generated assets get used and judged, not how many were made.
• Admin Studio gets its own chrome. A charcoal rail lists tools in usage order, and the topbar shows which environment you are on and which role you hold. Home answers what needs you and whether the estate is healthy with a queue of at most ten rows. User detail puts the person at the head of the page with one verb card and four sourced numbers, and the users list leads with who is slipping. The inference-fleet page opens with any anomaly written as a sentence: which server is down, since when, and what took over its traffic.
• Curriculum playbook generation runs as a background job. The request returns at once and the studio narrates the run from its draft row: the step, a sections counter, a bar and a clock, all surviving a reload. A ready draft says how many lessons were written. A failed one shows its error and confirms the credit was refunded.

Improved

• Agency playbook generation runs in about half the time. Content is written in parallel chunks instead of one call per module, and the phase is measured. A long run now stops itself before the platform kills it and alarms at 70 percent of its budget, and the curriculum wait estimate states what was measured rather than a bound.
• Prompt administration is one capability. An admin granted prompt access can run benchmarks, generate samples, save a prompt and create eval cases, instead of seeing a Run button that returned a permission error.

Fixed

• Thirty-four prompts whose output was being cut off at a stale ceiling now run to completion. Seventeen of them were critics, judges and devil's advocates returning truncated verdicts. The raise was first reverted by the production seed and then landed through it, verified live.
• 2 automatic fixes from the app's own repair loop.

Published

• A Status Is a Claim About the Past. The Live Row Is the Fact.
• Journal: Zero Red Is a Decision, Not a Gap

New

• Segments get a brief. The segment dialog opens with a brief the AI drafts from the documents and survey evidence labelled for that segment, with a devil's-advocate critique under it, and a person edits or accepts. Every ICP creation door shows "This ICP focuses on…", and ICP, persona and value-prop generation read the brief as their lens, with a cited beyond-the-brief pass for traits the evidence supports.
• Survey respondents sorted by market. One click has the AI label every unreviewed respondent against the segment briefs with a one-line reason, a person can correct any label, and a correction is never overwritten. New surveys can ask which market a respondent serves, results split per segment, and a segment ICP drafts from its own respondents instead of a blend of every market's answers.
• Suggested segments. When the evidence holds a group of buyers no segment names, a card arrives with a drafted brief, why it is distinct, and the documents, accounts and respondents behind it. One accept creates the segment. When the buyers look like one group, the card says that instead.
• A segment's own documents and interviews reach its ICP even in a large library. Documents labelled for another segment stay out, and unlabelled signal is ranked by how well it fits the brief rather than excluded.
• Buyer Personas matches its approved design in every state. The draft set reads as one dense table with Accept, Edit and Dismiss on the row, one action row ending in Lock personas, and a five-tile strip. Sharpening, the real-people lens, the empty set and the locked set render through the same shell, in light and dark.

Improved

• The November 1 launch date now drives the remaining places that still keyed off October 1, and trials that ran to the old date were extended to the new one.
• An admin's per-prompt fallback-chain pick in the prompts panel now reaches the model dispatcher. Before this it never did, for any prompt.

Fixed

• A due date a person types stays the date they typed. The scheduler recomputes only dates it derived itself, whether the task was edited in the app, in chat, or by an agent over the connector. Action-item priorities in chat use the same four levels as everywhere else.
• Two editors locking different Website Builder sections no longer lose one of the locks. A write against a stale version is refused and asks for a reload.
• A stale read can no longer paint over data you just changed. The shared data cache commits only the request that is still current, and the uploads list now runs on it.
• A judgment module drafted from chat leaves the same AI baseline as the other paths, so its lock shows what the human changed.
• Send-once onboarding emails dedupe on the date they quoted, not on bare existence, so a moved launch date gets its correct reminder instead of being blocked by the old one.
• Non-admins calling the admin refund and glossary routes get a refusal before any read. Dead exports and about 250 duplicate prompt copies that could drift are retired.
• 11 automatic fixes from the app's own repair loop.

Under the hood

• Fewer round trips on hot paths: a leaner evidence corpus for signal digests, one read fewer in daily lead scoring, a lighter background-job tick, and one auth hop fewer per page render, with identical output.
• Mission Control's Ledger tab shows which session started which. Chains nest as a tree with depth, and a launch whose child never opened is called out in red.
• An audit-row write that died in transport no longer counts as a request failure. The write is kept alive until it lands, with its own time limit.
• The merge gate no longer turns red for every pull request when a wave was completed in the work ledger before its plan row caught up.
• The nightly suite gets read-only database access for its gate sweep, its proof screenshots no longer block the publish, and two specs reach their intended state.
• The recurring deploy death is instrumented: a per-job network probe and build-output forensics name the broken file when it happens. The production migrate job that never ran green is removed, since production receives migrations by promotion.
• The approved designs for Content Studio, Website Builder, Visual Brand, Graphic Studio and Admin Studio are committed as the bar each surface is held to, as Buyer Personas was.

Published

• How to build the buyer-question list AI assistants should answer with your name
• Journal: A Date Nobody Signed Belongs to the Scheduler

New

• Feature Matrix per market: a company with several markets can keep one matrix per market, and each plot shows that market's buyers and competitors. A market matrix can start as a copy of the org matrix and pull title changes from it later, and a cross-market view points at the features positioned differently in two markets.
• The matrix plot carries only top-level capabilities. Sub-features nest under a parent, duplicates merge and can be restored, and an AI consolidation proposal is applied by a human. The seven AI actions on a market matrix reason from that market's ICP and value props, and say so when they fall back to the org's.
• A feature sheet with one tick column per line of business imports in one pass, each column routed to its market's matrix. Unmapped columns are reported, never guessed. Over the connector, agents can rank, retire, restore and move features too.
• Per-member module reveal: a project manager can show a client team member only the modules the engagement is on. A grid on the team page sets it per member, and a hidden module disappears from the sidebar, the journey, the map and the direct link.
• Search a data-heavy list, then act on exactly those rows. ICP triggers and pain points, value props, customers, personas, competitors, the presentation library and website audit findings get a search box, and Merge, Move and Delete work on the matches only.
• An archived ICP can be restored from the same menu that archived it. Accepted ICP suggestions now name their tab and jump to the row, an applied row leaves the list for good, and Re-distill asks before replacing saved work.
• Graphic Studio drafts the week's graphics before anyone asks. A "For you this week" row ranks the three to six pieces the org needs, draws each in the locked brand, says why it was picked and what grounds it, and keeps the critic's case against one click away.
• Prompt Craft in Graphic Studio: 150 terms across render, shot, lighting, lens, mood, motion and audio, most with a without-and-with example pair, as chips that append their fragment to the brief. A Sharpen toggle, Compare styles and a Learn drawer sit beside them.
• Team interviews are readable. The response count on an interview is now a door into a Responses lens, which opens first when responses exist.
• Auto-filed notices in the Company Library can be kept, not only undone.
• A pasted Google Drive link reads through the org's connected Drive account, and a folder link opens the Picker inside that folder instead of an error about sharing.
• Admins see AI providers by modality, priced in the unit each model is bought in. Image and vision prompts are offered only models of their own modality, and can be benchmarked two at a time on judged quality, cost and latency with a recommendation to confirm. Image, vision and video calls meter a unit beside dollars.

Improved

• The ICP page puts the data above the fold. The market and version bars ride the header, and the scorecard collapses into one band with its failing tests first.
• The growth calculator slows its growth rates year over year, and the curve ends on the ambition the card recommends instead of contradicting it.
• The beta launch moves to November 1, and the public copy reads one date everywhere.
• Erasure handles the remaining cases: a user's sole-member personal org goes with them, a legal signer can be erased with the signed record kept, and a provider's erasure no longer blocks on other buyers' bookings.

Fixed

• Org admins can create playbooks in Playbook Studio again. The policy named a role that cannot exist on an org.
• Project owners and admins can read their project's own modules and sections instead of seeing the template library with no tasks.
• The foundation counts on the signal page move the moment a file is classified, with no reload.
• Pain-Claim-Gain reserves thinking room per pillar, so a long ladder no longer saves a cut-off matrix. The ICP devil's advocate and the Messaging grid's copy validation get the room to finish, and a truncated prompt retries with more.
• "View as regular user" applies from the first paint, not after a client refetch.
• In dark mode the Focus Zone, due-soon rows and the Awaiting-you strip stop washing out to beige.
• Typed text in the wrap-up and notes editor survives a cold load, and a names-only customer import shows the new rows in place.
• The second site address redirects to the main one, and the old-design switch is gone.
• Six automatic fixes from the app's own repair loop.

Under the hood

• Mission Control's Ledger tab lists every open session on every machine, stranded first, and shows what is waiting to start.
• A prompt approaching its output cap raises a warning before it truncates, and each capped call site is bound to the output shape its prompt declares.
• The nightly suite stops filing a database stall or a harness artifact as product failures, and its stateful phase has the budget to finish.

Published

• A discarded draft is a training example
• Journal: The First Runner With Room Takes It

New

• More than forty screens move onto the one five-slot grammar: a heading that matches the sidebar door, a row of lenses, one red action that names the next step, and the standing after the canvas. Lists share one shape: a facet rail with exact counts, dense rows, and the picked item in a side panel.
• On that shape now: customer interviews, the competition monitor, competitive analysis and the signal library; team tasks, action items, the inbox, Relay threads, all projects, presentations, the product roadmap and the partner hub; the team hub, skill library, Academy courses, the public job board, recruiting, team member OKRs, My OKRs and the talent desk; the agent roster, the slide library, the relevance gate, ABM accounts, campaigns and audience.
• Instruments read as one graded result with stages in lifecycle order: Website Audit, Team Assessment, Work Styles and the GTM diagnostic. Scorecard, Engagement Health, Foundation Dashboard, My Dashboard and My Growth open on their standing as the first band of the canvas.
• Video Studio and Audio Clips make the library the canvas and the picked item its preview. Naming candidates and the brand voice guide render as specimens at the size they will be used: a wordmark, a name inside a sentence, a voice as a paragraph.
• The Signal meeting, the coaching 1:1 and the live workshop room follow one ritual grammar. The single red button is always the meeting's current step, from start to next item to end.
• The Growth Journey and Role Growth show one "you are here" on a progression band, in order and never red.
• Three more virtual team members draft for review and spend nothing themselves. The ABM strategist files a weekly tier slate where every Tier 1 promotion names the demotion that paid for it. The motion designer proposes a hero loop per piece with an in-browser animatic and a price, and Accept is the only way to a render. The campaign art director files one creative concept when a campaign goes active, with renders left to a human click.
• Sources say who gave them. The lineage chip on a grounded output names the person and role behind each source, and each foundation row on the signal page names who contributed what.
• Marketplace reviewers see the full offer, T2D3's margin and comparable live listings before a go-live, and can ask the submitter a question by email.
• Feature Matrix imports tell the truth. An import over the cap is rejected whole, parked, or truncated by a stated rule, and the result says what landed and what was parked. A dry run returns the same report with no writes.
• When the database is unreachable, the app shows a branded outage page within seconds instead of a platform timeout.

Improved

• ICP market pills draw as one bordered pill, and dragging the grip sideways rearranges the strip.
• The critic that reviews each virtual team member's draft has room to finish, so a run no longer ends in a cut-off verdict.
• The app's short name in the installed-app manifest reads T2D3 OS.
• An empty "Ask the team" now says where go-to-market roles are set.

Fixed

• Human work is off the product catalog. An agency package no longer appears as a "Free" product that could be claimed without payment.
• A locked website still publishes from chat, and go-live never marks a site live on a host that does not resolve.
• Share previews for the login and GTM diagnostic pages point at the T2D3 site image again.
• Account and organization erasure no longer stops halfway. Child rows are deleted before their parents, email-keyed signups are erased too, and a second erasure no longer collides on a unique key.
• The GTM diagnostic's delta chip no longer throws a formatting error on every report.
• A journal entry queued for auto-publish no longer waits behind stale slots from other series.
• Six automatic fixes from the app's own repair loop.

Under the hood

• Self-hosted models are gated by task family. A local model cannot take over the prompts it serves without a committed passing evaluation, and every local tier now has one.
• A second self-hosted model shadows the signal triage lane before the verdict, recording confidence and agreement without acting on anything.
• Agency HQ is ruled a second shell like the admin studio rather than forced into the five-slot grammar, and three of its largest files were split.

Published

• How to run a content queue that arrives already prioritized
• Build a POV bank so every article starts from an owned stance
• Journal: The Gate Refused the Cure
• Journal: The Path That Ends the Work Needs an Epitaph

New

• Simple view is in beta. Home opens with a greeting, one ask box, AI-ranked next-move chips built only from your real state, a foundation pulse line and a "Needs you" card. An At a glance column shows today's meetings with a drafted prep brief, yesterday's notes and the signals they produced, and threads to continue.
• For review: one cross-module list of everything the app drafted for you, including virtual team member drafts, with Approve, Edit and Reject. A rejection needs a reason, and that reason feeds the next draft.
• Routines: save a question on a schedule and an agent drafts the answer before you ask. Each run lands in Needs you.
• The mobile app's home becomes the same simple centre: a composer that opens the assistant, ranked chips, the pulse line, a native For review queue with reject-with-why, and a door to past chats.
• Collect signal by role. Invite a teammate with their go-to-market role and the welcome email asks for the signal that role holds. "Ask the team" turns each ask into an owned task with one email per person. A reply to the signal mailbox lands in the library tagged with who sent it and which ask it answered, and the ask closes itself. The library gains a Contributed-by column with contributor and role facets. After seven silent days the app proposes a reminder but never sends one on its own.
• Capture a video you own into the Signal Library. Confirm your rights first, then record from the video's own browser tab or hand it to a background capture engine. The recording is transcribed like any other upload.
• Book Radar reaches the human: a proposal card with the case for the book, the case against and what the radar read, readiness dials, and a "Consider a book" card on Home. "Wrong type" asks why and retires that kind of book for ninety days.
• Five modules move onto one positioned canvas where every item sits in its cell and an empty cell says what the gap means: Feature Matrix as the 2x2 its scoring already makes, Budget Mix at funnel stage by payback band, the Message Map at pillar by persona, Dev Matrix at performance by enjoyment, and Calibration at merit tier by team. Crowded cells scroll inside themselves and narrow screens keep the positions.
• GTM Plan and Book Studio keep their structure on screen while a draft runs, with a hold line counting progress and the selected part previewed beside it. GTM Survey and Team Interviews adopt the shared lens, stage and facet grammar.
• Roadmap suggestions publish themselves to the public board in neutral words once the redaction rewriter is confident, so votes can order the review queue. Founders get a product planning page that ranks every open suggestion by weighted votes, the panel's verdict, impact, risk and effort, refreshed nightly.
• ICP grouping is a drag: drop an ICP onto a market pill to label it, and with nothing lit every ICP shows with the primary leftmost. The standing "Company-wide" pill is gone.
• The Website Builder opens without a prerequisites modal. Unmet foundation shows as a band on the Site lens, and Pages, Images and Activity become filterable collections with an inline page card.

Improved

• Virtual team members draft before being asked, and their new drafts show up in For review instead of only on each member's card. Agency operators see which member spent what on which client.
• Engaging a user from the admin console previews the rendered email before anything sends, including the NPS request and free-text notes.
• One approval channel: a yes in the working session is the acceptance, and website copy proposals are decided in Mission Control or by replying to the mail.
• The careers, jobs and architecture pages have one owner on the www host, and every site page carries a canonical.

Fixed

• AI credit metering and signal search now admit the same callers the app admits, so a superadmin or project collaborator working in a client organization no longer trips access errors on every AI call.
• Virtual team members running at the highest effort no longer run out of room while thinking and return nothing. The reserve scales with effort, and a truncated slate fails loudly instead of closing as "nothing to propose".
• The Runners admin panel never serves a job history older than five minutes, so overnight failures no longer read as live trouble.
• The weekly workshop steward runs each of its bodies in isolation under a time budget, so one failure no longer skips the whole run.

Under the hood

• The five-slot module grammar is now a gate, not prose: every conjugated surface declares its archetype and slot evidence, and a lost slot fails the build.
• The Website Builder view shrank from 2,350 lines to 611 with no behaviour change, and a round of table, overlay and colour-name debt was paid on the Interface v2 surfaces.
• Founder-band pull requests get their band stamped from the changed paths, and a production sync that carries the fix for an error-budget breach no longer waits on that breach.

New

• Simple view: a quieter lens on T2D3 OS with one ask box on Home, a full-page chat with Max that shows its sources, a composer that can attach a file into Signal or take dictation, and Recent threads in the rail. The Simple view / Full view switch sits in the rail footer, the command palette and Account → Appearance; following a link onto a full-only page shows a "Back to simple" chip. Behind a feature gate while it finishes.
• Website Builder: typing "publish" with no destination now asks for one in the thread, with the chooser in the card, and the same thread carries on to Preview & publish. A lead capture section with three variants turns every form submit on a client site into a lead in your own audience, forwarded to HubSpot when connected. A seventh template, Platform, for multi-product companies that sell through a demo, scored against its reference sites. A product loop section plays a muted clip in a browser frame, with scroll-reveal and count-up numbers.
• Content Studio: at the last drafting step, try another locked voice or a stack of writing filters, read the rewrite beside the current draft or as a paragraph diff, then use it or keep the current one.
• Book Radar: the app reads your library and everything it has produced against thirteen book archetypes, shows which books the signal can already carry, and proposes one weekly with its evidence and a counter-case. Founder intake conversations count toward the founder chapters. Available where Book Studio is.
• Video Studio: a rendered cut keeps its own transcript, so it is searchable and clippable the moment it exists, with no second transcription.
• Agency packages name their unit: hour, day, month, quarter or project, with the unit pre-picked from the proposal's cadence.
• Requests: a failed request can be closed as already shipped, no longer needed or a duplicate, and the queue retries a transient failure once, closes requests whose fix landed another way, and surfaces timeouts.
• Interface v2 reaches Value Props claims, the content library, the compose flow and idea drafts, the Graphic Studio library and Visual Brand: a facet rail beside dense rows with an inline inspector, one red primary per screen, and Visual Brand's thirteen tabs become four views.

Improved

• A persona or value-prop set pinned to a market's ICP now grounds in that market's documents only; unpinned sets still read company-wide.
• Google Drive: a watched folder offers "Pick files in this folder", and the card says plainly that watching a folder does not reveal its files, since Google grants access per file.

Fixed

• A Growth Stage checklist tick followed by a quick tab jump no longer vanishes.
• A watched Google Drive folder ingests the documents in it, not only recordings.
• Merging two accounts keeps the signed agreements of the merged account.
• The made-here media lane in Content Studio lists and attaches again.
• 22 automatic fixes from the app's own repair loop: dark-mode primary buttons are legible again, an accidental AI merge of buyer pains can be undone from Versions, a pasted Drive link opens through your connected account, a dropped zip gets a readable refusal, and the ICP page no longer re-renders on load.

Under the hood

• Preview builds moved to our own build servers after the hosted builder ran out of memory, a new build host joined the pool, a database-policy check that had been silently skipping now runs on every change, and the repair loop no longer re-reviews a fix the panel already reviewed.
• Two scheduled jobs that ran on two schedulers run on one, ending the doubled admin email on a new product-qualified lead.
• The operations inbox closes an item once its fix merges and sends the drafted reply after the veto window, and stopped AI fixes get a recommended call to apply in bulk.
• House corner, elevation and spacing scales now back the card tokens, pixel-identical in the app.

Published

• How we build T2D3 OS: fourteen principles that govern the code
• How to audit an AI-built codebase: the full program, every brief, every gate
• Journal: A Template Is a Teardown, Not a Copy

New

• Newsroom: the bench grows to sixty-seven journalists, each in a publication's tradition under a fictional byline, and every organization seats its own nine. The desk proposes a panel from your ICP, personas and pillars with a rationale and a critic's note per seat; accept it, keep yours, or swap a seat from the bench drawer. You can also add your own journalist from a short brief.
• Three guided front doors shaped like the persona builder. Value Props drafts and scans the Great list, then deals the claims for you to rank on the Great / Better / Only ladder. Visual Brand asks what you have and what you like, then shows a board to react to. Brand Voice reads your own writing and deals the voice decisions as one-tap cards.
• Agency packages are buyable end to end: publish from Team, show in a rail on Marketplace → People, and a client pays in dollars by card rather than credits. Reviewers are emailed when a listing waits, and the seller sees "Awaiting marketplace review" instead of Draft.
• The Community library opens with a Templates tab: free playbook templates listed by operators and agencies, peers first, with search, sort and a "Use this template" door.
• The Video Studio records a demo in the studio: one take per brief step with a teleprompter, resumable upload, and re-record any single step.
• Every content list shows when a piece reached its step: "In review since Sep 25", "Scheduled for Oct 2".
• Glints: a new conversation is scanned for claims on arrival, private uploads are never mined, and a kept claim can be promoted to the content queue or to Company beliefs as a draft.
• Roadmap voting: an upvote can carry a 1 to 3 intensity, a paying, active customer's vote weighs more, and superadmins manage the weights from the admin console.
• Agency staff operate client organizations at their agency grade rather than as owners, the agency owns each new client, and an organization admin or the project's PM can start a Human Signal Panel on the shared pool.
• Interface v2 on Home, ICP and Personas: Home's next move names what an open foundation gap blocks and offers "remind me Friday", and ICP Accounts and the persona set become a facet rail beside dense rows with an inline inspector.
• The Website Builder groups pages into dropdown menus and footer columns: the strategist proposes the grouping with the sitemap and you reorder it in a Navigation panel. The T2D3 house template is re-curated from the current public site.
• Two more virtual team members: a brand strategist drafts the messaging How/Why ladders, swap-tests them against two curated competitors and signs each; a creative director grades every visual asset family against the locked brand board and names one thing to stop making.
• Google Drive connects through Google's file picker with per-file access instead of read access to the whole drive.
• The public About page explains how the OS is built, and "Book a walkthrough" on the cost and ROI page opens a form that emails an introduction to the team.

Improved

• Library counts keep their promise: the review banner separates pieces awaiting approval from pieces blocked by the quality gate, and facet counts respect the search box.
• The brand voice performance card and its "By format" chips reach the home page everyone opens, and a refreshed landing page stays a page idea, not a blog idea.
• An archived client's domain offers Restore instead of a dead end, guided onboarding from Agency HQ creates the client workspace first instead of looping, and a package's minimum commitment follows its billing unit, hours on an hourly rate.

Fixed

• Tenant isolation tightened in five places: no signed-in user can write votes, comments, assessments or usage rows into another organization; anonymous callers cannot draw on an organization's credit pool, and credits cannot be added with a non-positive amount; the AI connector refuses arguments naming two projects; a mailbox connection must belong to its own organization; and a template clone checks who owns the template.
• The mobile app launches again after six weeks of package drift, Academy progress reads back, and dictation uses the server recognizer.
• The Friday beta update went out twice because two schedulers fired the same job; it now runs on one schedule and claims the week before the first mail leaves.
• The Pain-Claim-Gain matrix no longer fails as "unreadable" on a large value-props ladder, and a quick reload of the ICP page no longer reports "action already running".
• 25 automatic fixes from the app's own repair loop: RTF documents upload again, drafted ICP rows can be edited inline, a wrong duplicate flag can be ignored, the Company Signal tab survives a refresh, and a role refusal names who can act.

Under the hood

• Code-quality baselines are compared against the merge base so a change cannot raise its own allowance, every plan with open rows must be a registered ledger program, and program triage runs on the server so the queue drains without an admin tab open.

Published

• Programmatic SEO without the spam: the information-gain gate
• Journal: Only Part of a Voice Should Bend

New

• Writing habits know what they are writing. Five identity habits follow the brand into every format; the other eleven are form habits a format may bend, across seven content registers, so a landing page is no longer told "no bullet lists" beside "scannable sections". Writers, the lint and the library grader read the piece's own register, a "By format" grid under the Writing habits panel shows every effective setting and why it was bent, and each verdict on a voice pass feeds the per-format recommendation.
• The Video Studio's new Demo walkthrough format drafts its own recording brief the moment it opens: the job to be done, who it is for, the value prop it proves, three to five steps with what to click and say, and a set-up checklist. It is grounded in the locked value props, personas and brand voice, and a devil's advocate challenges the draft. Under it, a headless motion engine turns a scene into one video part with entrances, staggered reveals and camera moves from the locked motion tokens, so a video no longer needs source footage.
• Run reports on many websites. On the Audience list, agencies select leads and queue one GTM Diagnostic per lead as background work, with limits per batch and per day and progress that names the sites in flight. The Website Audit becomes multi-site, with a switcher to add, rename or remove sites.
• The Newsroom's bench grows from nine to thirty-six members, each modeled on a publication's tradition under a fictional byline. The new members are catalog only for now.
• Agencies can disconnect a wrongly connected team payout account and set it up again; an account still holding money is refused with the reason.
• The guided persona builder's Exit becomes "Don't ask again" and remembers it, so it stops re-opening for a user who declined; the ICP builder's button says the same.
• Roadmap votes can carry intensity, with a weight computed on the server that no voter can forge, and public roadmap copy records who reviewed it and when.

Improved

• The public roadmap no longer shows a reporter's own words. An item with no curated public copy is rewritten before it reaches the board.
• A glint's echo test says whether its zero was measured. A verification that failed or never ran no longer reads as a genuine isolate, and the review queue labels it "Echo test did not run". The devil's advocate's read on a claim is kept, so every reviewer sees the same critique.
• A 1:1 in the coaching lane is private by construction. The database raises privacy in the same step that attaches the feed and holds it there, and an owner who tries to share one is told why it stays hidden.
• An ICP built for a market through the agent door now carries that market's label and shows up in its market pill.
• The primary-ICP recommendation is cached once per organization, so looking around the ask panel no longer re-spends a model call, and a failed confirm shows its error beside the button that was pressed.
• The demo sandbox is marked "Demo" in the workspace switcher, so it is no longer mistaken for a client project.

Fixed

• The rate limiter shared one bucket across thirteen routes, so a couple of chat calls could make the GTM Diagnostic say "You're doing that too fast" on the click that should have run. Buckets are now scoped per route, and the diagnostic charges runs rather than requests.
• The /invest form on the public site returned the marketing page instead of recording interest, and the /work and /jobs forms had the same hole. Every API path now passes the marketing rewrite.
• A client's order for a named agency member landed in the agency's package queue, where it could be reassigned. Only a package takes a name; a client's named pick is never reassigned.
• A vote or comment on a module item must belong to the current project; a foreign item returns not found.
• The team readiness report no longer quotes buyer interview marks as what the team said when there is no team signal.
• Copying a value-props set onto a different ICP reverts the copy's label on screen if the re-point fails, and long inclusion links on an agency service package wrap inside the card and are editable again.
• 10 automatic fixes from the app's own repair loop: page text is capped without splitting an emoji so Compass saves stop failing, a scorecard on a deleted ICP says so instead of erroring, a module action on an instance deleted mid-flight stops instead of running against nothing, and a run of by-design warnings stop being filed as errors.

Under the hood

• The invariants gate reports failure when it aborts part-way instead of a green check for rules that never ran, and acceptance probes whose clicks or fills failed no longer count as passed.
• Hand-run production data writes must name their authorizer and land in the operations ledger, checked in CI.
• A twenty-five-finding audit by a second model family was confirmed and its severity calibration recorded.

Published

• Count Before You Rule: Three Queries That Beat Three Opinions
• Journal: A Fallback Is a Publishing Decision

New

• An organization has exactly one primary ICP. Where the choice is arithmetically clear the app elects it and backfills organizations that had none; where several ICPs compete and none is primary, owners and admins are asked instead. That ask does its homework first: a grounded recommendation, its reason and a devil's-advocate challenge sit beside the choice, nothing is pre-selected, and the confirm button says why it waits.
• Markets are labels, so nothing is inherited any more. "Move to another market" on the ICP canvas becomes "Belongs to markets", a multi-select, and "Company-wide" is now simply the name for an ICP wearing no label. The inherited chips, Override and Hide-here controls and the inheritance health line are gone; a labelled set owns all of its own items. Duplicate copies a market's labels, Merge and Delete move labels rather than sets.
• A signal document can carry market labels too, from a Markets verb in the Signal library. An ICP inside a market grounds in the documents that share one of its markets plus every unlabelled document, and a company-wide ICP is never narrowed.
• Locking one market's ICP stales only the work that read it. A personas or value-props set pinned to a different market's ICP is left alone instead of being marked stale across the whole organization.
• Every value-props set names the ICP it promises to. Sets that named nothing were linked to the organization's primary, and an organization with no ICP gets its first one, empty, the moment it writes a promise.
• A mixed team interview splits in two. The market half grounds the ICP as before; the team half becomes its own document in the people lane, covering capability gaps, ownership, morale and strategy buy-in, and the team readiness report now reads it. The private coaching lane stays out.
• Agencies can sell through the marketplace. Earnings carry the payee recorded when they were earned, so money stays where it was earned when a worker changes agencies, and an agency earnings statement shows it. An agency can list a package, capacity with nobody's name on it, and assign a member once the order lands; packages carry the agency margin only, with no take rate. Before collecting money an agency accepts its own supplier agreement, now in force, and the team page shows a readiness strip naming the one next step between the agency and its first supplied order.

Improved

• Account-based marketing sends use the consent jurisdiction the human confirmed. Every send path had been assuming an unknown regime and refusing all touches, including the drip cron; an organization with no confirmed jurisdiction is still refused, but the message names the decision and how to recover.
• A Google Drive pull in Explore that finished with some files failed can be pulled again for exactly those files, and the card now tracks the retry rather than the first attempt.

Fixed

• The Import from links dialog on the Signal page painted underneath the library rows, with no dimmed backdrop. Dialogs now render at the top of the page wherever they are opened.
• Retention nudges judged most active users as never active. A single capped read of everyone's activity kept only the busiest users; each user's last activity is now read on its own.
• Two nightly jobs that were being killed at their time ceiling now finish or save their progress: support triage defers what it cannot reach and picks it up next night, and the content performance rollup saves its place after every organization. The competitive scraper caps how many pages it reads per site.
• 29 automatic fixes from the app's own repair loop: the guided persona card sort no longer says "Sub-item not found" when a card is dealt, the ICP Profile Criteria panel and the Disqualify ledger show the same criteria, a Drive recording is claimed before it is downloaded so two runs cannot ingest and transcribe it twice, long recordings no longer abort at a 30-second upload limit, an agent passing a shortened id over MCP is told to send the full one, and a run of by-design warnings and missing-row lookups stop being filed as errors.

Under the hood

• The production release gate speaks up. A hold that stands for hours while content waits now pages a human instead of reporting success, a dropped deploy upload retries, and only a failing production canary can veto a sync, not a failed build.
• Dozens of acceptance probes that had reported passing for weeks without executing now run against real fixtures, a spec that skips everything fails unless its quarantine is ledgered, and the intake panel stops re-reviewing bugs the repair loop was already ordered to fix.

Published

• An LLM should never grade its own citations
• Journal: Inheritance Is a Coupling. A Copy Is Not.

New

• Market segments grew up in one day. A market can now hold a full ICP set instead of a single instance, an ICP or a value-props set can move to another market from its row menu, and the market strip's menu gains Rename, Duplicate, Merge and Delete, each stating what happens to the work inside before it happens. By evening the model settled further: a market is a label on an ICP, not a container around it, so one ICP can carry several markets and the organization keeps exactly one primary ICP that decides the homepage and the default audience.
• Brand Voice's "Use this voice" menu exports the locked voice as an Agent Skill, a zip that Claude, Claude Code, Codex, Cursor, Gemini CLI and Copilot load on their own whenever you draft something the brand will publish. It sits beside the five existing paste packs.
• The Signal pack card scopes itself. It carries a free-text search over document titles, a multi-select facet per attribute with live counts, a minimum signal score, an upload date range and a named preview of what goes in, and the copy uses the full width of the card.
• A meeting is one signal document. The review queue folds a recording into its typed note, and a bulk-bar Merge joins two to twenty rows into one, with the app recommending which row stays and the confirm saying what happens to the rest.
• Persona sets have a real Delete. The row menu had offered Archive under a Delete label in six languages; the permanent delete now asks you to type the set's name and is checkpointed for support.
• Team interviews and one-on-one notes get their own signal lanes. They feed the team development and coaching modules instead of grounding the ICP, personas or value props, and the reviewer sees the lane as a pill on the document.
• Every glint claim card shows an AI case against the drafted consequence, with up to three objections, before the reviewer votes Keep or Kill. The gate also shows why each document counted as an echo, and a passage only counts when it asserts the claim outright rather than presupposing it.

Improved

• The ICP module has one door for all four ways to build, with the guided walk as the top choice. The walk opens on a project's first ICP only instead of every visit, the "start from project context" path is gone, and the drafting screen settles when the build finishes rather than spinning. A guided walk on a project with no signal now says so before it starts instead of bouncing you back to the chooser.
• Persona tier and relay role are one question. The first persona is the initiator, the second the decision maker, the third the blocker, so the separate Relay role dropdown is removed and the tier labels carry the meaning.
• Client sites built in the Website Builder take their corner radius and shadow from the locked Brand Visuals instead of fixed values, in preview, publish and the deployed starters, and the Website Styleguide gains a Shape group naming the source of each.
• Five canvases lost their second red. Brand red now marks the primary action only, so a red badge or link no longer competes with the button that acts.
• The Lingo integration is retired. The two import doors in Visual Brand and the credential card are gone, since the API sat behind a plan our customers do not buy and failed identically to a stale key.

Fixed

• Google Drive ingest had been dying at its memory limit since mid-September, refetching the same large recording every run. Recordings now stream straight to storage, documents stop at a size cap, over-cap files are skipped from their listed size, and a Drive pull in Explore that died mid-way now says so and can be pulled again instead of reporting itself as done.
• 11 automatic fixes from the app's own repair loop: missing-row lookups no longer file as errors, a watched Drive folder holding only chat and notes files no longer stalls, a Growth Stage interview that replied "Set" without saving now shows a confirm button and the true saved state, and a text paste that hit a transport blip asks you to retry instead of reporting failure.

Under the hood

• A time-boxed capture of full requests for eight prompts feeds a blind, cross-family judge panel comparing a candidate model against the incumbent, so a provider move rests on real inputs rather than fragments or self-graded claims.
• Four acceptance probes now prove, in the running app, that an ICP and a value-props set can change markets from the screen, that a market's primary sticks after the server answers, and that kept glint claims lead the Ideate picker. The pre-merge preview wait was lengthened to outlast the build it waits for, so those probes actually run.

Published

• Why our content critic never sees the writer’s reasoning
• Journal: A Missing Row Is Not an Incident

New

• Claude Opus 5.5 is now the default model behind T2D3 OS. Every prompt and agentic route that ran on Opus 5 moved to it, and a reviewed set of tasks that had been routed to the heaviest tier moved as well, where its judgment was not the product.
• The copy-paste prompt library has a permanent door on the Signal page again. The prompts that let you extract a persona or an ICP in your own Claude, ChatGPT or Gong and bring the result into the library used to be reachable only while the library was nearly empty or a checklist row was unmet.
• The Signal library's selection bar gains a bulk "Only me", so you can mark several documents private in one action, with an undo on the toast. The private flag is also available as a table column. Your own documents flip either way; someone else's can only be restricted, and only by an owner or admin.
• Compose shows one workspace-level notice when the organization has no locked brand voice. It names what the app cannot write in yet, links to where a voice is created, and disappears once one is locked.
• Visual Brand now holds all seven design-token categories: colors, typography, spacing, corner radius, shadows, components and assets, with a grounded recommender for the spacing and type scales. The exported tokens file carries every category and imports back without loss, so a brand designed in Claude Design comes home whole.
• Uploads that no GTM document type should ever cover, such as images, font packages, licence files and empty extractions, no longer reach the classifier. They receive an honest type up front instead of a wasted model call and a permanent "unclassified" row.

Improved

• Everything the app writes now carries the organization's governing brand voice. The remaining module families, from ICP and personas through value props, growth matrix, feature matrix, website builder, interviews and visual brand, were brought under it, along with the client progress update, recovery brief and early-warning note in engagement health. Prompts that extract, score or argue rather than write stay voiceless on purpose.
• A divergent claim now counts as echoed by the market only when a model confirms a passage asserts the same claim, not merely discusses the topic. Hedged or unparseable verdicts go to a human, and the Feature Matrix market-interest axis reads from the verified count.
• The Growth Journey page fixes its two critical accessibility violations: the plan collapse button has a name for screen readers, and the glossary trigger declares itself a button.

Fixed

• The default "Organize my signal" press failed every time with a generic error when no focus text was typed. It works again.
• The duplicate finder in the Signal library grouped different meetings as copies when one side was an empty recording with nothing to compare. Same topic is no longer treated as the same document, and an empty recording cannot chain unrelated notes together.
• 30 automatic fixes from the app's own repair loop, most of them transport blips and false alarms that had been filed as production errors, plus batched writes where a drag-to-reorder or an admin analytics read had issued one query per row.

Under the hood

• The six website starter templates now share one source for their common files, and the section prompts derive from the section registry, so a variant change is one edit instead of six.
• The internal daily digest names any document the classifier gave up on, so a document that grounds nothing costs someone a glance instead of looking classified forever.
• The nightly interface test run stopped spending its budget on waits that could never resolve and now seeds the fixtures its probes depend on, so more of the suite runs each night.
• The app's own repair loop now defaults to Opus 5 and must write a reason to step up to the heaviest model, and a reviewer's objection arrives with its full argument so the founder can rule on the record.

Published

• Wow, How, Now: the editorial spine that survives channel trends
• Journal: Not Every Prompt Is Writing

New

• Claims the agent extracts from voted-out signal now reach a human. The Glints queue under Signal shows each divergent claim with the consequence it implies, and you keep or kill it on that consequence, not on the quote. Keeping demands a written thesis, promotion is limited to three per organization each thirty days, and a promoted claim lands on the strategy matrices as a marked bet that a human accepted, never as evidence and never as a lift to a quality score.
• A kept claim is now idea fuel. The Ideate tab's glint group offers kept claims first, one sentence each with the reviewer's thesis as its detail line, ahead of the per-document chips that used to fill it.

Fixed

• Review duplicates in the Company Library could show a copy with no verdict and still report the group as reviewed. Every copy now carries a verdict chip or an explicit "Not compared yet", the bar claims groups were reviewed only when at least one verdict exists, and a recording's take time is labelled instead of sitting as a bare number beside the play button.
• The rotation guidance for the Lingo integration on the credentials page linked a settings page that does not exist. It now points to Lingo's own help article, names where the token is minted inside a space, and notes that the tab exists only on Lingo's Business and Enterprise plans.

Under the hood

• The nightly interface test run reported an empty Agents page as a defect. It was a test-environment switch left off, not the product; the run now resets those switches before it starts.
• The session banner that lists pending database changes now separates "applied to development, awaiting the production sync" from a genuine gap, so it stops asking for changes that already landed to be applied again.

New

• Market segments: an organization that sells into several markets can hold one ICP per market. A Market strip above the ICP switcher lists Company-wide plus each segment, and a segment's ICP shows what it inherits from the company-wide profile as read-only rows you can override or hide, beside what it adds itself. The connector works on the same model.
• A brand voice and a set of brand visuals now say who they speak to: the whole organization, one ICP, or named personas inside it. The org brand is always kept, an ICP-tagged visual brand sits beside it as a variant, and the audience a piece is for picks its voice and brand.
• A guided buyer-persona builder opens the first time you arrive at an unstarted persona set. It drafts three tiers from your Signal, asks you to name the three seats, then has you place twelve cards with one tap each before handing you to the full persona pages.
• Graphic Studio composes a quote card and a stat card from your locked value props and signal evidence, shows the reasoning and the case against, and typesets them against your locked brand so a figure renders exactly as typed. Carousel, framework diagram, infographic, ad set and event banner follow.
• Video Studio's six cut objectives become a format catalog with a gallery front door: the job each format does, the brief that works, the mistake people make.
• GTM Survey tells its two failures apart: people who never opened the link are a distribution problem, people who stopped partway are an instrument problem, and the read-out shows both. A pilot state keeps rehearsal answers out of the data, and a refined instrument can be saved as a template for the next client.
• Approvals has a door in the rail under My Work, badged with what is waiting on you. A pending proposal from your AI team reaches every owner and admin through the bell, and where Slack is connected it can be decided from the channel with a button or a reaction.
• An edited recording carries its own transcript, projected from the original along the cuts you kept, so a recut saved to the Signal Library has words in it without a second transcription.
• Certification badges in the Academy are drawn per track instead of one generic icon for all.
• The connector clears the bar for listing in connector directories, and the help explains how to connect T2D3 OS from ChatGPT, Gemini and Grok.

Improved

• The Signal library table has one horizontal scrollbar instead of two, a "What it is" column that says in a sentence what each document is, both dates, and a copy-transcript action.
• Quick add moved out of the left rail to the top bar beside Take notes. Developer settings folded into Connections, so API keys and connector tokens live with the other credentials that act as you.
• The survey preview's "Open live survey in a new tab" works before the survey opens, so a project manager can check the real participant view during design.
• An ICP criterion states each fact once: values that differ only in case, spelling or a slash are merged, and the lock gate shows all of them. The persona side panel no longer asks for a relay seat it already knows.
• Links the AI teammates hand you open the exact review drawer, and an agent asking about a module gets the tabs it actually renders.

Fixed

• A brief typed into Compose's "Tell Max what you want to write" arrived in the Max panel as an empty greeting the first time Max opened in a tab. It now lands as the first message.
• Opening Max beside Compose squeezed the editor to one character wide. The rails now fold before the editor shrinks.
• The English translation set outgrew a row ceiling, so the last namespaces rendered as raw keys after a cache miss. The whole set is read.
• The Signal drafts card showed an error while the AI batch behind it was still processing. It now waits.
• Nine scheduled jobs, including onboarding and retention mail, had silently stopped being invoked by the hosting platform for five days. They now fire from our own fleet.
• OKR suggestions drafted in the background fell back to keyword search instead of vector search. They are grounded again.
• A demo whose project was deleted restores itself on the next open, and a prospect who clicked a demo link no longer vanishes from the commercial digest.
• Forty automatic fixes from the app's own repair loop.

Under the hood

• The AI marketing teammate can read the production go-live clock, ask for a sync, and has a per-capability track record counted toward graduating from proposal to auto.
• Nightly acceptance runs had been cancelled inside an unbounded step every night since early September. Each step is bounded and the reporting tail always runs.

Published

• The analyzer with no tab: When a feature becomes a reflex
• What an AI crawler actually sees on your website

New

• Graphic Studio can edit an asset you already have instead of regenerating it from the brief. On the asset review panel: Reframe to one of seven aspect ratios, each tooltip naming the channel it is for; Cut background; Sharpen 2x, which invents nothing and is safe for logos and screenshots; Enlarge with detail, which re-renders and is never for identity assets; and Fix a spot, where you drag a box on the image itself and say what changes inside it. Every edit lands as a new asset with its source recorded, so the original stays the master.
• Formats that carry words are now composed rather than painted. Text is laid out from your locked Visual Brand tokens and a layout template, fitted to the frame deterministically, and every color pairing is checked for contrast before a template may use it, so a light brand color can no longer render label text unreadable on white.
• Draft image renders run on an open-weight model served on T2D3 OS's own hardware. Three candidates were measured on one fixed prompt set and the one that both spells rendered text and follows framing won; it now serves the draft tier for hero, illustration and photography renders, while final renders keep the metered chain.
• A GTM Alignment Survey's confidentiality promise is now a control the product enforces, not a line in the invitation. The owner picks attributed, confidential or anonymous; scores and verbatims below a minimum group size are withheld under the two stricter regimes, and the choice locks the moment the first respondent has been told what it is.
• Authoring a survey no longer starts from a blank screen. A fresh survey names its three doors, write questions here, import a questionnaire already in use, or let the agent draft a first pass, each question shows its required toggle, a routing summary states which of the six respondent functions sees what, and a codebook ships with the responses so every figure names its base.
• The AI marketing lead can read T2D3's own marketing numbers through a read-only tool: traffic split by site, search clicks and queries, the signup and paying-workspace funnel, and form conversions, with any dark source named. Only internal staff may call it, and contact-level rows are left out.

Improved

• Per-file usage instructions in the Signal library now override the document's literal text when a foundation module reads it. A transcript that mistranscribes your company name, corrected once in the file's instructions, stops leaking the wrong spelling into the ICP.
• The Graphic Studio library grid mounts only the rows on screen, so a library of hundreds scrolls without stalling.

Fixed

• The personality test typed nearly everyone as an intuitive-feeling type. Each trait is now scored against its own population norm rather than one shared midpoint, so thinking and sensing types are reachable and the percentile bars read true.
• A failed certification exam attempt no longer returns the answer key. Correct answers and explanations appear only after a passing attempt, whatever the exam's teaching-aid setting.
• Review duplicates opened on a group of one with nothing to remove. The kept file now sits in its own cluster beside its copies, so the group can be acted on.
• The admin billing page showed an organization on a retired plan as Free and over its limits, and buried the credit pool total under a long member list. The plan it is actually on is named, subscription controls appear for it, and the usage lists start folded.
• 81 automatic fixes from the app's own repair loop, most of them closing support tickets and error reports across the ICP, Personas, Value Props, Compass, Content Studio, the Academy and the Max chat.

Under the hood

• The self-hosted image endpoint closes the connection after an error reply, so a refused request can no longer scramble the next one on the same connection.
• The repair loop's wrap-up drain steps past items it skips instead of stalling on them, and the inbox counts a fix as shipped only once its change has merged.
• A flaky interface test that had been knocking unrelated changes out of the merge queue now asserts the settled state.

Published

• Your marketing system should draft before you ask
• Six Months of Shipping Software With Claude Code: Our Field Notes
• How to make every background AI job accountable to a human owner

New

• The new website design, built on the preview host, is now the marketing site on www.t2d3.pro and every T2D3 host. Pages outside the main site tree, such as the ROI calculator, render in the new design too instead of dropping back into the old branding, and the legal pages no longer print their footer twice.
• Atelier is now Graphic Studio, its own door in the rail beside Video Studio, named for what it does. Old Atelier links and the Atelier tab in Visual Brand send you to the new page, and series and medium links come along.
• Graphic Studio opens on a format gallery instead of a bare medium dropdown: eleven formats grouped as Brand imagery, Channel-ready and Brand system, each card with a real example, a one-liner and the channels it suits. Picking Icon set, Graphic element set or Social profile kit opens its tools inline under the recipe. Visual Brand keeps directions, identity and the logo, and points to Graphic Studio for the rest.
• The onboarding interview board fits on one screen: all five hypothesis cards sit above the fold as one-sentence summaries, each expandable to its evidence, with a ruling bar that counts what you have confirmed and the interviewer chat docked at the bottom.
• The settings window is the only settings interface and is live for everyone. The old account, organization, billing and settings pages redirect to the matching pane, including their section links, so no bookmark breaks.
• The Foundations certification exam reads as its 25 grouped questions instead of 85 numbered rows, and every question that asks about a table or diagram now draws it above the answer options. Checkbox questions score the way the original exam did, adding points for correct choices and never subtracting for wrong ones.

Improved

• A project member manages ICPs and persona sets the way a project manager does: rename, delete, merge, reorder, and lock or release a lock, including entries they did not create. The manager role is for the people side of a team, not a gate on GTM work.
• The Content Studio's To approve column no longer drops pieces silently past twelve. The header count is the true total, a piece filed in the last day always gets a seat, and a footer link takes you to the rest in the Library.
• Closing a GTM survey no longer seeds ICP, personas, value props and the Growth Matrix on its own. Close-out records the aggregates and parks the seed plan, Apply is a human click behind a minimum response count, and the invitation no longer promises anonymity the survey does not offer.

Fixed

• A video that had sat open for a while no longer fails with a "link may have expired" error on Play. Every video player quietly fetches a fresh link, returns to where you were and resumes, and only shows an error if that fails twice within a minute.
• Graphic Studio no longer shows a red error when you reopen it while its automatic scene draft is still running, and its title no longer appears twice on the page.
• A Learn article that carried an editorial comment in its source no longer prints that comment to readers.
• 13 automatic fixes from the app's own repair loop.

Under the hood

• The AI model catalog watch now flags new releases from providers that date their listings, stops calling an older model newer, and can tell an unset provider key from an empty one. The models found in this month's library scan are catalogued and selectable.
• A task handed to the AI marketing teammate on a day her brief already exists is worked rather than skipped, and her reply opens its own pull request instead of landing on a branch nobody sees.
• The check that keeps production secrets away from agent steps in automation now scopes to the step that runs the agent, so a secret in a fixed pre-step no longer fails unrelated changes.

Published

• Your AI teammate has a spiky profile. So do your best people

New

• Settings open in one window over whatever page you are on, from the gear in the rail, Cmd/Ctrl+, or the command palette. Closing it returns you to the page untouched, and a search box reaches individual settings, so typing MCP lands on the tokens card in both the window and the palette.
• Your own settings are regrouped into eleven panes named for what they hold: Profile, Appearance, Menu, Notifications, Voice and dictation, Connections, Developer, Mailbox, Plan and usage, Programs, Account. API keys moved out of Connections into Developer, the Mailbox has its own pane, and Notifications lists which projects send you a weekly digest.
• Organization settings, billing and the old settings page are one Organization group in the same window: Members, Integrations, Clients, Brand, Profile, Plan and billing, Data and privacy, and a Danger zone for owners. Each pane shows only what your role allows, and the old addresses open the window in place.
• A merge in the Company Library can be taken back for seven days. The library shows what was merged into what, how long is left, and an Undo that restores the merged files, their classification, the overwritten fields and every reference the merge repointed.
• Notes take pasted, dropped or attached files: a screenshot or document pasted into the editor uploads and appears as an image or link. The Library detail panel lists a note's recordings, lets you move one to another meeting note or open the note in the notes editor, and rows whose note has audio show the audio icon. Merging a recording with its meeting note keeps the audio, the transcript and the note.
• The in-app Foundations exam asks the questions of the real certification exam, so a badge earned in T2D3 OS means the same as one earned anywhere else. The answer key never enters the codebase.
• The new-design preview of the website now covers every page, with a pricing page drawn from live plan data and a Cost and ROI page that prices the operating system plus the Navigator who runs it against a traditional team. Both website versions read one pricing source so they cannot drift.

Improved

• The guided ICP wizard's last screen invites the two tests it skips: Deal economics, which checks whether this ICP is worth selling to, and Market density, which counts the companies that match your profile and pulls a sample of real ones. The ICP saves either way.

Fixed

• An action item added under one project task also appeared under the next task you opened, and deleting it seemed to remove both. The panel now starts clean per task and drops any read that arrives for a task you have already left.
• The AI teammates' mailboxes read only the oldest twenty messages, never reported the backlog behind them, and turned a failed mark-read into a green run. The check now sees the whole queue, fails when marking fails, records every hand-sent email on its thread, and reads forwarded messages instead of showing them blank.
• One automatic fix from the app's own repair loop.

Under the hood

• Shared interface widgets are placed by family in the change-authority policy, so a two-line widget fix no longer waits on founder review while navigation, onboarding and the collection views still do. A test now fails when any policy path matches no file, after two were found matching nothing.
• The Search Console sign-in client is named in the credentials ledger, and the sweep that keeps the ledger complete now sees credentials that code reads by name rather than spelling out.

New

• The Signal library's table is now the default view and the place where the work happens. Each row carries Signal, Noise and Glint buttons and keyboard shortcuts, so a library of hundreds can be sorted without opening a document. Two cleanup columns show which files duplicate another and which are near-empty. A List choice you make from here on still sticks.
• The library now records who a conversation was with and where in the relationship it sat, separate from what kind of document it is. Both are classifier verdicts you can review and pin, each showing the transcript line that decided it.
• The library can say which buyer conversations point at one buyer and which at several. It first resolves the corpus into accounts deterministically, then groups them, and reports usefulness and coverage beside the result. It hands the grouping to the ICP module and never names an ICP itself. It also spots when an audio file and a document are the same meeting and proposes the merge.
• A guided wizard walks a new ICP one test at a time. It opens on an ICP that has never been finished, names the flagged documents it will build from, drafts every section while narrating its progress, then asks one keep-or-drop question per test. The existing module is unchanged; a Build your ICP button reopens the wizard any time.
• Team Interviews scripts open on the Scripts tab, and Interview with this script starts a meeting note with the script as its template: questions, why each is asked, and follow-ups, titled for the role and filed with the team's interviews.
• An AI teammate on the campaigns team drafts the campaign's copy package the moment its briefs land: one line the campaign says, a first pass for every undrafted piece, and a critic against it. Nothing is written until a person accepts, and every generated piece then descends from that one idea.
• A second, more visual homepage is running beside the real one at www2.t2d3.pro. Every stat on it is read live, it links back into the existing site, and it stays out of search.

Improved

• A saved meeting is one library entry. The recording folds onto its note as a chip, is named after the meeting instead of a timestamp, and is renamed when the note is. A Meeting notes filter sits first in the chip row.
• On an admin user page the actions on a person now lead the page: view journey, impersonate, grant credits, comp a plan, send a sign-in link, and delete, set apart at the far end. The delete confirm counts what it removes, in sessions, uploads and memberships, and what stays.
• Mission Control's ledger shows one row per initiative with its waves instead of one fragment per working copy, and judges completion for the initiative as a whole.

Fixed

• An agency owner working inside a client organization saw an empty portfolio while the header counted managed clients. The portfolio now scopes to that client's parent agency.
• Facet menus in Browse all templates, the admin playbook lists and the marketplace did not close when you clicked beside them inside the panel.
• Wheeling over the right rail of a note now scrolls the rail, and My notes lists up to twenty-five notes instead of stopping at four.
• A member saw holes where the integration cards should be. The cards now render with a note saying who can connect them.
• Seven automatic fixes from the app's own repair loop.

Under the hood

• The support diagnosis report falls back to the profile record when the account store is unreadable, and names the facts it could not see rather than going dark.

New

• Bugs and ideas posted in a connected Slack channel land in the feedback queue. An emoji or a leading word such as bug or feature request sets the type; otherwise an AI classifier decides. A bug becomes a support ticket with a response deadline, a suggestion becomes a roadmap draft, and an unknown poster gets a thread reply.
• Naming finishes the job. Make it the company name asks you to type the name, then renames the organization and its starter project, and one confirmed click reverts it. The organization and project pills in the top bar pick up the new name without a reload, and each naming session keeps one leading name.
• A survey knows the decision it informs. On first open the agent proposes the decision frame, who reads the result and why, and you confirm or edit it. Drafted questions tie to that decision, item rules flag agree/disagree scales, hypotheticals, missing units and open text placed too early, each with a fix, and a survey estimated over fifteen minutes cannot be sent until it is shorter. The survey page header also shows again on the project route.
• Five AI teammates join the virtual team, each filing at most three proposals you accept, revise or reject, none acting on its own. On the Marketing Scorecard one checks how every metric is captured and proposes a source for each gap. On Resonance one reviews the sending estate for authentication failures, bounce or complaint breaches, warm-up schedules and consent gaps. In Content Studio one annotates every chart with a title stated as a claim, unit and denominator, a does-not-show caveat and a source line, every number re-checked against the citation and editable inline. On a locked ICP one plans the territory with segment slates, denominators and a stop call. On Visual Brand one maps an icon metaphor to every uncovered value prop and feature.
• The AI marketing lead can act in the app. Mail from a known colleague wakes her inbox check within the hour instead of waiting for the next scheduled slot. Six connector tools give her publish, mail, calendar and newsletter hands, each tool resolved to auto, propose or deny from a policy on her own credential. A proposal becomes an approval a human decides in the app or by replying to an email with a short code, and the approved action runs as that human. Newsletter sends can never run on their own, and she can now propose changes to the public website.
• Collection surfaces get a shared kit of facet rail, rows and detail panel, and module headers can show a human-validation chip that counts how many items a named person has validated.

Improved

• The admin home is a triage queue of at most ten rows, each one object, one verb and one reason, linked to the place that can act on it. Per-user tools fold onto the user record as a Journey tab, and the old routes redirect.
• The engagement table shows signup date, active days and sessions in the first days after signup beside the last days, with window buttons above the table. A user's features-used count now counts features, not event names, and the health score follows.
• The density control leaves the navigation rail and lives under Account, Preferences, beside the light and dark toggle. A change there repaints the open shell in place.

Fixed

• A recruiting candidate admitted through a magic link was refused on their next login and then hidden from every signup report. The admit decision now reads the account type stored on the profile.
• In the preview environment an AI job queued behind older ones never ran and its button spun forever. The preview now drains its own queue the way production does.

Under the hood

• The fast model tier routes to a different self-hosted server, and the self-hosted model stops reasoning by default, so short structured calls no longer time out and fall back to the cloud.
• Build hosts can be drained, powered off, resumed and resized from the server stats page, with capacity handed to another host while one is down. A runner that loses two jobs is quarantined and reinstated on a pinned build, orphaned builds are reclaimed under memory pressure, and a watchdog that was corrupting runner checkouts now works in its own directory.
• Server stats name AI calls and GitHub Actions jobs apart, explain each cloud fallback, and the runners view reads recent history: job size, durations, failures and attempt number.
• Shipping a capability files its companion article as a Content Studio draft for review, one act instead of a committed markdown file and a publish task.
• The AI coding partner works support mail: it looks up the sign-in and delivery record for the address in question and hands customer success the findings plus a draft reply.
• The schema lease guard gates on what a statement does, not on the tool that runs it. Three noisy CI schedules run less often and a per-day failure rate lands on the weekly ledger.

Published

• The agency I would build if I started from zero today
• An AI Teammate Should Tell You What It's Waiting On
• Renaming your company is a decision, not a field edit

New

• Content Studio keeps every version of an article. A rewrite, a writing filter, a manual edit or Regenerate no longer loses the text that was there: a History button lists each earlier version with what replaced it, and restores it in one click.
• The content planner does the missing half of the job. Beyond recommending a pace, it ranks the queue against the open slots and marks each idea make, park or drop with a fit grade. When nothing fits it names the real reason, such as a full window, and offers the lever that would clear it. Plan the Queue and Plan the Calendar are now named apart.
• The Calendar tab proposes the team's planning rhythm: publish days, drafting days and a thinking window, each inferred from the team's own calendar history with its evidence shown and one click to change. The next plan run packs against what you confirm.
• Customer interviews tell evidence from noise. A competitor the interviewer named first is tagged prompted and never raises a tally, compliments and future intent do not count, and an anchor log records what the interviewer said aloud. Synthesis reports how many independent customers back each pattern, checks whether the sample is saturated, and flags led witnesses. The agent recommends the interview type, discovery until patterns can form, and the drafted script follows the type you pick.
• Team interview scripts get a devil's-advocate critic and a check that marks leading questions and offers a past-event rewrite, on generated scripts and on questions a teammate types in. Scripts follow a timed 45-minute spine.
• Brand Refresh reads the Naming module instead of running its own name rounds. A card shows the leading name and the contenders with their votes, and earlier Brand Refresh candidates moved into Naming with their votes intact.
• ABM account scoring is calibrated against real outcomes. An AI teammate measures tier precision and meeting recall on accounts scored at least thirty days ago, proposes at most one constant change with a critic against it, and files a holdout when the cohort is too small.
• Market tests can be decided on qualified leads, not clicks. Newsletter links and social share drafts carry their experiment arm, an organization can connect WhatConverts, and an hourly pull brings back calls and form fills so the experiment reads a qualified-lead rate.
• Administrators get a live purchase ledger: every charge in a chosen window with who bought what and a receipt link, plus the subscription book, searchable and exportable, read from Stripe on every load.

Improved

• Admin Studio shows one rail instead of two. On admin pages the app rail steps aside, and Back to app returns you to the page you left rather than a fixed index.
• Contributors can start the first persona set on an ICP that has none. The empty state showed them no action while the New set button further down already worked.
• Mission Control's ledger lists sort by start, activity, machine or status, each item's next step is an AI recommendation with a devil's advocate and a person who confirms it, and sync times show in the viewer's time zone.

Fixed

• Accepting one drafted retrieval fix in Content Studio wiped the other four. The remaining fixes now ride along with the accept, re-verified against the text as it stands.
• A kept on-camera recording placed on a Website Builder page published as the waveform audiogram. The publisher now prefers the recording and falls back only when there is no cut.
• A signup refused for using a personal email address still appeared as a new user in digests and the admin dashboard. The refusal is stamped on the account and every cohort read excludes it.
• Fifteen automatic fixes from the app's own repair loop.

Under the hood

• Changes carry an authority band read from the files they touch. A change can raise its own band but never lower it, and the band gate became a required check on every merge.
• Production builds only on our own fleet, so a release no longer builds twice. Builds wait until the host has memory, run only where they can finish, and a hung test suite fails loudly instead of silently dropping the pull request from the merge queue.
• A merge burst no longer starves the preview environment of every deploy, a stale preview turns a check red until the commit is live, and the runner watchdog no longer kills live jobs on GitHub's word alone.
• The production sync gate stops counting a held draft as content waiting, sees the heartbeat canaries that walk production between pushes, and promotes schema changes itself.
• The work ledger rescues edits stranded mid-rebase, names machines by seat, and records the clean machines it looked at.
• The AI marketing lead's daily brief reaches the team by email instead of ending in a merged file. Both AI teammates read their mailboxes three times a day and can answer a message, only ever to the address on the thread they are answering.
• Fleet hosts can be drained and safely powered off from the server stats page.

Published

• Your AI Makes 159 Judgment Calls. Ours Has to Declare Every One in Writing Before It Ships.

New

• Naming module, four waves. Every candidate shows its votes, you keep a shortlist, and a devil's advocate argues it on collision, sound, meaning and longevity, then names one recommendation. The module checks domain availability itself, a likely-free domain registers from its card with a live quote and a confirm before checkout, and a registered domain becomes working email and your Website Builder site's address in one click.
• Signal library Table view. A List | Table switch shows one row per document with source, type, Signal score, ICP and persona presence, the modules it feeds and tags. Every column sorts and filters, and a row still opens the detail panel.
• Content Studio takes many files as material. Drop up to eight documents and images on New idea or onto the Compose empty state, and say how to use them in an instruction that reaches every draft.
• Schedule an approved piece. Pick a date under a web channel's Publish menu or set its calendar slot to Scheduled, and it publishes on that date.
• ICP decks know the moment. Before the lock, Create Presentation Slides builds a working-review deck of what is ranked, open and to be decided; after the lock it presents the settled ICP. Generated slides save as one dated deck that opens straight in the viewer.
• The ICP density screen shows one market funnel, universe to serviceable to obtainable, with a caption saying how the hard and soft criteria cut each stage, in place of three rival account numbers.
• Website Audit opens on the Report tab and turns findings into work: accept the AI-suggested next steps or file everything as project tasks.
• The 26-member virtual team is visible inside the app under a Team lens on the agents page, with each member's mandate, autonomy tier, modules and run history. Two members now work: one files ABM account-coverage proposals for a person to accept; another turns the website audit's failed checks into a ranked fix set.
• Customer interviews follow a method. Nine named evidence traps feed the script critic, a coach card precedes every call, and drafts have a timed spine at 30, 45 or 60 minutes. Hypothetical and stated-price questions no longer ship.

Improved

• Every Compose drafting path applies the whole locked brand voice, never truncated, and the output is enforced in code, including the no-em-dash dial. Rewrite in this voice now rewrites your edited draft instead of regenerating from the original inputs.
• ICP after a live review: the economics fields take a typed number; members can pull the account list; the Distill review says Not saved yet at the top; anti-ICP rows are two lines and the receiving bucket glows when you file one; custom dimensions get a name and a search box spans the Profile tabs; an item moved to Profile arrives with an AI-recommended category to confirm; and triggers added by Pull from Signal wear a New chip.
• Module switchers refresh live: a teammate's new or deleted ICP appears or leaves your tab strip within seconds.
• Brand Voice lab: drag a file onto the upload zone, and the checklist no longer shows doors a guest cannot open.
• Transcripts download as Word or PDF from the signal inspector.

Fixed

• The foundation checklist counted a company's own deck and press release as competitive intel; it now scores on what a document is.
• Deleting a company's only project no longer brings it back on refresh. New signups no longer receive a duplicate personal workspace, and the executive dashboard stops flagging free-tier and personal workspaces as comps.
• A hand-set task status survives bulk edits and the to-do sync. Value Props locking flips the body with the header, with no refresh.
• Compose: pain and trigger options are readable and wrap in full, and the Image ideas dialog no longer lets the article paint over it.
• The Growth Stage interview's Send button was invisible in dark mode, a failed save there now shows a red Not saved chip, and the Personas header no longer collapses to one letter on a phone.
• Activating a Competition Monitor showed raw database text; it now says plainly that nothing changed. The ICP prioritization ledger certified the 90% gate on criteria that no longer existed; it now re-reads after merges and deletes.
• Seventeen automatic fixes from the app's own repair loop.

Under the hood

• Production sync moves from once a day to an hourly check with a gate: it syncs when content changed, 25 or more commits wait, or the oldest has waited six hours, and holds while production is red. Mission Control shows the next sync with a Run sync now button.
• Mission Control: items waiting on the review panel read Panel reviewing instead of Stuck, Send all ready replies wraps up panel-decided items, and ops alerts land on the errors page with a recommended action.
• Eight of the largest handlers and routes became folders with one file per action, thirty-six loose library files moved into their domain folders, and the refactor program closed.
• Two critical framework advisories closed by moving to the patched release. Three never-scanned database indexes retired.
• A rejected changelog entry now gets a revision pass instead of a dark day, and unfilled author prompts send a piece back to review.

New

• The Brand Voice lab is public. Anyone with no account can open it, name their brand, paste or link something they have written, or upload a PDF, Word, PowerPoint, Excel or text file, and the production Brand Voice module runs for them inside a private sandbox. The voice is never gated; the take-away asks for an email, and the work becomes theirs when they sign in. The free tools now sit in the Learn menu, their own footer group and the features page, and the ROI calculator is indexed.
• Presentations has one home for everything deck-shaped: saved decks grouped by project with open, present, share link and delete; the slide composer; and a read-only browse of the slide templates. Saved decks had no list at all before.
• Signal library folders can file their own arrivals. After you apply a folder proposal, each qualifying folder offers a rule with Turn on or Not now. Once on, a matching upload lands in that folder and the library says how many files were auto-filed, with Undo. The proposal itself is now drafted unasked once an organization has ten or more unfiled files.
• Outbound market tests, end to end: ABM relay content and designed sequences emit two or three variants on one declared dimension, a sent touch records which variant went out, and the test is decided on positive-reply rate with the required volume and detectable lift stated before the send. An under-powered test is reported as directional and kept out of the learning loop.
• ICP Focus says when the niche slate was proposed and whether anything in the evidence has changed since, so Propose again is a considered choice rather than a guess.
• Every organization member sees the LinkedIn sending card in Organization Settings and can link their own LinkedIn account. The card now explains that sending is 1:1 outreach from a person's own account while LinkedIn social publishes to the company page.

Improved

• Feedback you file gets an answer sooner. Reviewed items that were decided but never closed now get their reply, long-parked items receive a recommendation with the strongest case against it side by side for a person to confirm or overturn, a suggestion that is really a bug is refiled as one, and colleagues on company mailboxes no longer wait for an approval click to report a bug.
• GPT Image 2.5 is the default image model, and the reference-image edit path uses its edit-precision tier so a recolor stays a recolor.
• Two identical reads of a module no longer collide: a page remount used to turn a harmless double fetch into an "action in flight" error.

Fixed

• Slide exports failed on decks whose footer logo declared its own size; the logo is normalized on ingest and at render, so every layout exports again.
• The board PDF showed zero interview and survey evidence for every organization; it now counts the real items. Key-result links on action items never persisted because the backfill was refused by row security; it now writes.
• Meeting routes answer a database failure with a server error instead of blaming the request. A failed close no longer leaves two meetings in progress, and a scorecard metric can no longer be imported twice.
• Creating a personal OKR cycle without dates failed; it now takes the current quarter. An OKR update with no id no longer reaches the database as an undefined filter.
• The Brand Voice lab's AI steps failed on their first real run because the public route spelled the action field differently from every client; both sides now share one envelope. The lab's abuse caps no longer count signed-in T2D3 staff testing it.
• The content editor's grounding read comments on a misspelled column and threw the whole build; it reads the right one.
• Nine automatic fixes from the app's own repair loop.

Under the hood

• The typed database client migration is complete: the untyped client is deleted, every route, connector tool and helper downstream of auth is checked against the schema, and the defects that surfaced were fixed rather than cast away.
• Four of the largest module handlers, personas, visual brand, customer interviews and ABM, became folders with one file per action, every line moved verbatim. Nearly two hundred tests moved beside the modules they cover.
• A work ledger records what each session and each program set out to do, re-derives every completion claim from git and GitHub daily, and reports to a Ledger tab in Mission Control, so a reboot or a finished session cannot erase a plan.
• The daily production sync now promotes schema changes itself and pages a person only when promotion fails. The production build waits for idle cores as well as memory, and the repair loop resumes a fix when its credentials lapse mid-run instead of failing it.
• The routing audit runs on a schedule for the first time, the prompt-proposal inbox clears in bulk, and all 43 skill briefs were reviewed against their real handlers.

Published

• Reviewing 600 Files Without Reading 600 Files
• Autonomy Is Earned, Not Configured
• Your Brand Voice Is a Contract, Not Three Adjectives
• What an AI Agent May Do When Nobody Is Watching

New

• The connector, four waves: it can curate the signal library (triage, verify, re-parse, organize, delete), read every one of the 23 module types with a curated projection, answer the reads that were missing (brand assets, content pieces, the scorecard, ABM measurement, audience contacts, your own usage, agency clients), and run a module's own actions behind an allowlist that can never include lock, send, spend or delete.
• Brand Voice ships all twenty-nine writing filters, not six. Filters also mount on the lock step and the examples playground, protected phrasing keeps the words a brand never rewrites, and every rule you switch off is read back as "what this brand is not".
• Content Studio idea page: brand voice and the brief's primary fields sit at the top of the left rail, the reader detail is collapsed, and writing filters apply per piece with a rewrite.
• Signal inquiries take a pasted screenshot or an attached file. The attachment lands in the library like any other document and is scoped into the run automatically.
• The signal inspector states where a document came from, names the door it arrived through, and offers Confirm. Delete lives on every row, the bulk bar paints above the inspector, and the Company Library shows how accurate the agent's folder proposals have been.
• Signup: the company name is asked up front and names the workspace, the plan follows the role, every new user starts on a trial whose length is set in admin, and the card ask says what it is for.
• Personas: starting from scratch lays down all three slots at once (the user, the decision maker, blockers and sponsors). On a fresh project the recommended entry path is the one that can actually run.
• ICP: a locked Overview reads as a summary rather than a workbench, Customers absorbs Import & Learn into one door, the third-party data pill sits on each criterion row, and unsubmitted estimates say so.
• Scorecard and GTM Plan: the long generations run in the background and survive switching client or closing the tab. A returning page finds the running job and keeps the button busy.
• Organization Settings renders one section at a time, most-used first, so a collapsed card no longer leaves a dead half-screen.
• You can lock your own module set. Naming the set that speaks for the whole organization is reserved for its owners and admins, and a refusal says why.
• Book Studio's title page names the author and the organization's own imprint, never ours.
• Deck builds record which documents grounded them, and a reviewer sign-off is dated.
• Budget & Channel Mix files a variance readout 30 and 90 days after a lock. A metric with no capture path is reported as a blocker, never estimated.
• Resonance's door gate is scored against actual reply rates, with never-sent drafts held out of the count.
• The public team page lists the people and the two AI colleagues, one of whom writes this page.

Improved

• The diagnostic is the GTM Diagnostic everywhere: journey, activation checklist, prompts, help and every locale.
• The Growth Stage focus-mode interview has a gutter and shares the row with the diamond map instead of pushing it below the fold.
• The signal-foundation popup names the sources without counting them, so the live count comes from the checklist itself.
• GPT-6 Astra is in the model catalogue for evaluation, and the highest reasoning-effort setting is honoured instead of silently clamped.

Fixed

• Opening a project from the signup journey brings the menu, nav bar and chat panel back on every route.
• A published Content Studio piece no longer shows Approve and Reject, and no caller can walk a live piece back to draft.
• Long-form pieces are no longer blocked by a cost guard that charged already-paid research against the writer's budget.
• Pieces marked published in the app were not live on the site; the publish step is repaired and the production sync now runs on a schedule.
• Test signups from our own acceptance drives no longer send welcome emails.
• A market signal with no competitors no longer crashes the nightly backfill.
• The team-interview alignment critic clips disputes on a word boundary.
• Five automatic fixes from the app's own repair loop.

Under the hood

• Every module handler now has a contract test beside it, closing a program that started at 74 untested handlers.
• Several hundred API routes now use a typed database client, and the type errors that surfaced were fixed rather than cast away.
• The build fleet's admission gate counts cores as well as memory and spends one shared budget per host, the stuck-runner watchdog can see workers again, and the production deploy no longer runs a second typecheck.
• The prompt-library review is standing rather than a campaign: a model tier change now requires a behavioural re-audit in the same change.
• Groundwork for a public Brand Voice sandbox: a token-scoped guest API mirroring the production module.

Published

• The Lock Button Should Start Work, Not End It
• One Number for "Is Our Marketing Grounded?" — The Signal Quality Score, Explained

New

• ABM LinkedIn step, two halves. First a manual assist: the module resolves the profile and drafts the text, and the touch stays queued until you press "I sent it". Then bring-your-own transport: an organization connects its own Unipile tenant under Integrations, each user links their own LinkedIn, and an approved touch sends from that user's account.
• Content Studio: pick a different locked brand voice on the article screen or in the review drawer and redraft in it. A voice change never rewrites silently; the piece says the draft is behind the choice until you redraft.
• Compose: the rails drag wider, the Context panel names the voice and lets you switch it, the brief pre-fill reads as prose, and the empty center offers a prompt box that hands your intent to Max.
• Calendar reads as part of 2 · PLAN beside Ideas, and a search box finds any piece across the whole calendar, past dates included.
• Help search answers natural questions instead of matching a page verbatim, and Max reads the same index, so it cites and links the help section it quotes.
• Atelier gains a media lane: keep your team's own photography and campaign files, foldered, without them becoming brand input. Media grounds nothing and ships in no brand kit.
• Signal → Add from website has two new boxes: the website of a company that is your ideal customer, and a market or analyst view. Neither is read as your own messaging any more.
• The connector applies the app's plan gate to every tool by declaration; a tool absent from the map is refused. It also gains a feedback seam that lands in the same distiller as in-app feedback, a redacted lineage read, and a feature catalog derived from the module registry.
• Signal meeting prep reads the meeting-tagged uploads on its own, so board letters and investor updates reach the draft before anyone opens it.
• Website Audit keeps a member's filed finding through a re-score, Website Builder keeps each generation's quality measurements, and workshop pre-reads name the documents that grounded them.
• Marketing clips cut as audio-only from a video source offer "Add the video clip", re-running the render for the picture without discarding the audio or captions.
• ICP: the disqualify ledger says Hard filter and Soft filter, so "signal" means evidence quality everywhere.

Improved

• A second AI colleague, an AI CMO for T2D3's own marketing, gets a daily run: reconcile yesterday's promises, read what published, write a brief with named asks.

Fixed

• Brand Voice: Regenerate carries the rulings you made in the devil's-advocate step instead of reverting them, and Apply retries a parse miss.
• Uploaded interviews no longer sit untranscribed for over an hour: the job is recorded before the audio derive starts, and one that dies mid-derive heals on its own.
• An empty ICP Economics tab is an unmet input, not a failed draft, so a fresh ICP is no longer thrown away over numbers nobody was asked for. The growth calculator likewise keeps a draft when a metric is missing.
• The Customer Interviews quality check graded a shape the module stopped writing and always scored zero. It now reads what the module actually produces.
• Max: a suggestion chip no longer replaces a draft you pasted, and a failed skill run renders red instead of green.
• ABM compliance: an unknown jurisdiction is a hard deny, never a pass under a law that does not govern it.
• The Signal meeting's extraction and metric suggestions stop reading documents the team voted out, and the signal backfill drains its backlog oldest-first instead of re-serving the newest uploads.
• Signal Quality Score: first-party measures whether the buyer spoke, not whether a human authored the document, so seller-side decks no longer score a perfect first-party share.
• Journey: Growth Stage is step one on every surface, a step ticks when its module is locked rather than when the page was opened, and the signal-foundation popup counts eight sources like the checklist does.
• Content visuals: a chart's unit is verified against its citation, and large numbers are grouped in the reader's locale.
• Six author placeholders that were live in published posts are filled or removed, and a check keeps the next one from shipping.
• One automatic fix from the app's own repair loop.

Under the hood

• The type-escape burn-down ran three waves: website pages, chat messages, OKRs, API routes, client views and workshop items read through schemas or the generated database types instead of assertions. Open escapes fell from 498 to under 200, and the live defects the assertions hid were fixed on the way.
• Every build was tracing the whole repository into every serverless function because of misplaced test files. Fixed, and a check now catches it in a second.
• The quality-check packs are proven to read item types and fields the modules actually write, and the module catalog stops declaring types nothing produces.

Published

• Journal: Agents lead, humans steer: what proactive-by-default means for a small marketing team
• The Empty Input Field Is a Confession
• Journal: The Gate Graded a Shape Nobody Wrote
• Journal: An Unmet Input Is Not a Failed Draft
• Journal: A Placeholder Is a Promise

New

• Signup and first run, three waves. The wizard is full-screen, the role question is required and never defaulted, and it ends in the Growth Stage interview. A skipped role no longer drops a founder into Agency HQ, and an operator setting up a client is asked about the client, not their own company.
• Growth Stage focus mode: the interview opens the tab, and Continue stays disabled with the reason until the stage is confirmed. Later tabs are locked and say why.
• The five "What brought you here today?" options are curated by an admin, and the GTM Diagnostic opens with that choice selected. A competitor named in the interview chat lands on the competition card.
• Signal page: the foundation checklist leads the panel, and the drop zone with every import door moves into the folder column.
• Brand Voice writing filters: six named bundles of writing rules with an intensity dial, applied after the voice locks. A filter can change later or apply to one article only.
• Brand Voice: each voice column in the Try-it playground picks its own model, so an A/B test can compare models as well as words. The claim policy drafts itself from the voice guide and the target ICP.
• ICP: a proposed improvement can be accepted change by change instead of all or nothing, and the Distill panel reads as a decision rather than a report.
• Interviews: a story promised on the content calendar with no story on hand earns a question in the next script. Multi-step builds run a spec interview first: at most three questions, a one-screen spec, then the build runs with no further questions.
• Leads: a flat-fee data provider joins the enrichment waterfall, and account lists expand into named people with graded emails.
• Early Access: beta ends October 1. Today's prices become the Early Access rate, locked while a subscription stays active, against a published list price.
• The tagline is The Living GTM, on the login card, the site, the welcome email and the machine-readable surfaces. The public team page publishes the full T2D3 OS team, including the virtual team of Navigators, Sculptors, Engineers and Scribes.

Improved

• The prompt library, rewritten for the current models: twenty-nine system and contract prompts gain one paragraph on who consumes the output and what a wrong answer costs. Contract evals moved from 23 of 30 to 28 of 30. Fifty-eight house skills become briefs that state the decision they serve and when they are done.
• No blue. The blue token family, the journey-map rainbow and the deck palette's blue are retired, and every cool utility hue now paints charcoal or brand gold. Yellow is a sparing accent, not a secondary colour.
• Book Studio is in beta, and a feature rollout flip leaves an audit trail.
• The privacy policy and terms say what was already true: one policy covers the app and the marketing site.
• Quality frameworks run as written: authored thresholds are the ones the verifier uses, check ids resolve, and a skill that promises a locked upstream stops when it is absent.
• Skill-built personas take the shape the Personas module saves, so they render downstream. Generated content is English only.

Fixed

• A chart no longer zooms past the page it sits on; labels and strokes keep their size at every width.
• The three highest-traffic prompts had been sending literal template slots to the model, and an override identical to its seed had been masking every later improvement. Both are fixed, shadow overrides are retired, the admin editor can re-base a shadowed override and records why, and the seeder no longer re-activates an override an admin parked.
• Brand Voice: the team vote can be closed with the live ballot count, a deadline change after invitations persists, the Distill footer button says "Reset to the AI suggestions", and the calibration recommenders no longer spend their budget on a preamble.
• The value-props lock gate can see attached proof, and a long claim is no longer cut mid-word. The ICP win/loss validator no longer argues for deleting the exclusion criteria the loss data supports.
• A scheduled journal entry no longer returns a missing page the day it comes due. A magic link completes on the host that asked for it.
• Nine automatic fixes from the app's own repair loop.

Under the hood

• Prompt text is off the schema stream: all 725 prompts live as files with a hashed seeder that can also target production, so a copy edit is a readable diff instead of a migration.
• Three new repository gates: the shared-code root is closed, every module handler gets a test beside it, and a skill brief may only name tools the skill can call.
• Thirteen waves of type-escape removal put the typed database client back on across routes and handlers; the write errors that surfaced were fixed, not cast away.
• Web search and fetch tools are versioned to the model that runs them, and the pull-request queue re-runs stale verdicts instead of rotting on them.

Published

• The Eight Sources Every B2B GTM Foundation Needs
• Journal: The Verifier That Verified Nothing
• Journal: The Column You Write But Do Not Own

New

• This page. From now on, the app's AI partner reads what merged each day and writes it up here. The Labor Day journal entry explains why.
• Proactive-by-default, waves P2 through P4: the Growth Matrix drafts its portfolio when the ICP locks; slides generate on their own lock; a website audit chains its report when the crawl finishes; the lock critique pre-runs at six lock gates; seventeen surfaces that already fired on open are declared so; a new on-write trigger; scorecard targets, ICP focus pains, Reply Studio drafts and Resonance qualification all run before you ask.
• Organizations have exactly one owner. Ownership is transferred, never assigned, with a confirmation, and the database enforces it.
• Interviews: a shared interview contract fed by signup answers, goal and boundary prompts at high effort with an A/B replay harness, the script as the contract with verbatim storage, and a capture protocol with a three-question debrief.
• Executive dashboard: churn and downgrade reason capture, a Stripe reconciliation tile with a data-health strip.
• Signal library, waves C and D: a folder-structure proposal card (propose, rename, skip, apply), a review slate first, one edit model in the inspector, near-empty files flagged for deletion, and a bulk bar that sets Feeds, Tags and the verdict.
• Compass: the journey trail chip, a skip-ahead door and a chaptered report.
• Newsletter: monthly by default, graded content selection with evergreen filler, a rest-of-community segment and a delivery readout.
• Mission Control: a filed request records who asked; a merged coding request lights its inbox item; a pull request nobody merges nags and stays mergeable; one priority, one state, one owner, with draggable columns.
• Content review: the Retrieval-ready checklist drafts its own repairs, sourced.
• Uploads: "flag as orphaned" is recoverable, with restore and purge, and an acceptance test proves the blast radius.
• AI credits: background AI spend is attributed to the user who asked for it.
• Sales texting: the opt-in screen and full message list are published for carrier review.
• Interface v2 phases 5x, 5y, 5aa, 5ab: AI Agents, Team, Engagement Health and Book Studio in the five-slot grammar. The whole set is conjugated.

Improved

• A voted-out document never grounds anything: Noise feeds nothing, a Glint feeds Content.
• Prompt review: one per-provider effort knob, and the house agentic-posture prompt rewritten (why, act then stop, continue).
• Actions that are unavailable say why instead of hiding. Inline micro-labels become chips.
• Every admin page states its own width; the fleet page uses the whole screen and charts the CI runs.

Fixed

• A scope-less Google Drive connection shows one reconnect notice, not a flashing toast.
• Direct membership never demotes an agency operator; which client organizations count toward billing is right; in-place Unarchive.
• View-as is read-only for uploads, and "make private" names whose eyes it closes.
• The human-voice pass reads the whole draft and refuses partial revisions.
• Media failures in Video Studio say so and offer retry; a content fingerprint needs enough content to mean anything.
• A Sentry title is an error class, never an incident identity, so alerts stop folding into the wrong request.
• Eleven automatic fixes from the repair loop.

Under the hood

• The bare-fetch sweep is complete: 403 bare client fetches to zero across eighteen batches, and two more route-validation batches. The refactor program's waves 3 and 4 are next.
• The pull-request queue watcher runs every fifteen minutes and rescues unarmed pull requests that fell behind.

New

• Content Studio: the Newsroom. Nine scribes on three desks research beats and file stories; an editor argues with them. The content calendar becomes a lens of the studio, and share posts show up in the Library under their piece.
• Signal library: the Export button. Your evidence, tags and provenance leave as a file you own. Wave A and B of the agentic library: Organize runs the recommender on the same click, today's slate with reason chips, and a "Your pattern" card that applies how you file things.
• Visual Brand: a logo design journey that briefs, explores and refines with you instead of asking for your taste. Generated image sets pin one provider and anchor on the brand style.
• Executive dashboard: leadership, functional and investor views, a "Share with investors" panel with a public link, "What changed this month" commentary with a devil's advocate, cost categories and a monthly cash close, board plan versions with a headcount roster, pipeline in Audience with snapshots.
• Compass diagnostic: a "Grounded in" line on the cover and buyer briefs; quick take versus full diagnostic with a depth gate; the model that actually scored is stamped.
• Messaging: the content audit persists and picks from Signal; approved interview clips ground every messaging draft.
• Brand Voice: a "What this voice was built on" evidence card, named-voice routing with lineage on every piece, and a claim policy (stance, proof style, never-assert phrases).
• Growth Stage interview: the organization's shared thread, persisted and resumable.
• Marketplace: worker supply in three phases (one schema for profile and listings, admins publish listings, a review queue for buyer-visible edits); reviewers see what changed since a template was listed.
• Projects: move a project to another organization from its settings.
• Audio clips: optionally keep the video cut from the original recording.
• Agency HQ: commercial terms on the 9am page, and a steward that drafts the monthly client report.
• A second demo company beside the first, with a 71-document corpus and twelve gated modules codified at beta.
• Interface v2 phases 5k through 5v and 5z: Website Audit, My Journey, Core Signal, Growth Matrix, Feature Matrix, GTM Plan, Budget & Channel Mix, Marketing Scorecard, ABM, Resonance, Website Builder, Presentation Library, Workshop and Competition Monitor in the five-slot grammar.

Fixed

• The Ideas row sits on one line at one type size.
• Audiogram captions center in the empty band, off the title.
• Newsroom "Run now" enqueues through the right client, and the search seam falls back cleanly.
• The ICP type module is React-free, so API functions stop tracing an icon library into their bundle.
• Fourteen automatic fixes from the repair loop.

Under the hood

• The refactor program's big day: thirteen batches of the bare-fetch sweep and fourteen batches of route-body validation; the seven value-import cycle groups broken; three oversized module handlers split into per-action folders; a library-to-components layering rule with pure siblings; the OKR domain moved into one service with contract tests, and its API and agent tools onto it; read routes onto the row-level-security client in three batches.
• Every AI call logs its thinking tokens and effort setting. Background jobs drain from the fleet; the hosting-platform fallback drainer is gone.
• Admin can flag the company's own organizations internal, so usage reports separate internal from customer spend.

Published

• Your Signal Library Belongs to You. Here Is the Export Button.
• Three Ways to Be a Navigator, a Scribe, a Sculptor.
• Journal: Nine Journalists Who Never Sleep, and the Editor Who Argues With Them.

New

• Site: "While you were sleeping", the growth labor calculator, is the third lead magnet. Learn cards get a clean share-image URL.
• Archetypes: twelve named ways of doing the four Syntropy roles, resolved in the app and on the career pages.
• Executive dashboard: the metrics engine (an org MRR rule, the bridge, net and gross retention, cohorts), a monthly snapshot engine with a weekly steward and backfill, brand-palette chart primitives and a shared sparkline.
• Agency HQ: the 9am page (production queue, read queue, pulse), an agreements route with a team payouts card, Team & Talent listings that pin the agency, an engagements ledger with a renewal recommendation, white-label eligibility in one definition, a capacity and billing card.
• Brand Voice: a "Use this voice in Claude / ChatGPT" export pack and a solo-first vote; the app says when the primary voice is not locked.
• Marketplace: edit a service listing; a human-review queue for community templates where the AI recommends and a reviewer decides. Owners edit their own playbook templates.
• Max: saves what it drafted as a Content Studio piece, and writes the blog from your brief instead of a wall of messaging cells.
• Signal: the Feeds pills in the inspector header are the editor; the duplicate banner's Review opens the grouped review.
• Admin: a user's detail page shows every agreement pending or signed.
• Journey: the onboarding tracker carries past the payoff into visual brand, site page and content.
• Interface v2 phases 5f through 5j: Brand Voice, Messaging, GTM Survey, Team Interviews and Customer Interviews in the five-slot grammar. The borderless brand card lands as one utility across the app.

Fixed

• Diagnostics run by an operator scope to the client's organization, not the operator's own.
• An empty Team OKRs tab is never the landing tab.
• The pencil on a published piece opens its body.
• A resolved voice tension carries its kept side into the voice and the A/B test.
• The nightly triage loop reads screenshots and never parks a screenshot report.
• Ten automatic fixes from the repair loop.

Under the hood

• The schema snapshot and database types scope to the branch's own migrations.
• An orphan reaper for the self-hosted runners, fleet-wide.

Published

• Buy the Hour, Keep the Insight.
• Three Ways to Be a GTM Engineer.
• The Week T2D3 OS Stopped Waiting to Be Asked, Wren's weekly.
• Journal: Four Months of Letting the Machine Fix Its Own Bugs.

New

• Signal: the first foundation distill starts itself when the corpus is ready. Clip suggestions are drafted at transcription finalize, before anyone opens the recording.
• Navigation: demoted rail rows reveal by name ("Show 3 more" lists them) instead of a bare count. Skill Library became "Max's Skills" and moved beside Automate.
• Agency HQ: one operator identity, a routed hub with one route per section, "Portfolio" as the operator's Home word, and an Engagement Health roll-up on the overview. Ledgers for engagements and rate agreements underneath.
• Admin: a per-user Beta switch; credentials gain a mint link, a "Check readiness" preflight and expiry detection.
• Content Studio: the Target queries table is resizable, sortable, with choosable columns.
• Site: a page on what T2D3 OS is not: a portal, a framework community, a content chatbot.

Fixed

• The templates gallery opens above the Create Project dialog, not behind it.
• Credential expiry badges no longer wrap onto two lines.
• Self-rotation of the hosting token authenticates with the new token.
• Sixteen automatic fixes from the app's own repair loop.

Under the hood

• A second machine joined the self-hosted CI fleet with four runners; browser workflows pin to it.

Published

• A Growth OS Is Not a Consulting Portal.
• Taste Is Not Evidence: Why Style References Must Be Hard-Excluded From Your ICP.

New

• Website builder: a conversation mode. Describe the site you want; the builder plans the structure, writes every section, and offers a T2D3 house template as a sixth archetype. Journey steps scroll to their section.
• Navigation: the ratified eight-door rail is the default for new accounts, and the command palette indexes the rail's own sections.
• Home: the Home artboard, a journey band with one next move.
• Interface v2, phases 1 through 5e: the rail's visual conformance, the five-slot module grammar (brief, evidence, draft, decision, trail), and the first five modules conjugated in it: ICP, Buyer Personas, Value Props, Content Studio, Visual Studio.
• Brand Voice: "Tell me what I sound like" reads the voice back before you vote, in a dialogue you can argue with, plus a writing-model picker.
• Agents: locked Syntropy role avatars, drawn by the app's own image engine.
• Proactive-by-default: a locked ICP gets an arrival-driven improvement proposal; content-target search phrases are drafted before anyone clicks Suggest; the 103 click-to-start surfaces are frozen so the count can only fall.
• Every logo gets an app-derived light and dark variant, picked per background; a white-label logo knocks out to white on the rail.
• Account: section-navigated settings, an AI usage card that breaks spend down by usage class.
• Notes: pick the organization a note is saved to; the project is a tag, not the home.
• Projects: project-scoped task search on the Timeline, shared with the Board and List.
• GTM survey: a deadline-closed survey re-opens from Extend on the same link.
• Admin: a CI runners dashboard as a Mission Control tab; separate marketplace margin and worker take-rate knobs; Human Signal Panel spend ceilings as knobs; credential rotation that finishes itself (redeploys production and retires the old secret) and writes GitHub Actions secrets too.
• Training: a 62-lesson core curriculum along the journey, five minutes each.

Improved

• The heavy AI tier moved to the newest frontier model, with the prompt sweep it owes in the same change.
• Background AI jobs route through a batch API, at half the price, with no change to the output.
• A fleet-side cron dispatcher runs sub-minute triggers, off the hosting platform's hundred-cron ceiling.

Fixed

• The weekly content sweep skips internal organizations (demo clones, QA, canaries).
• Refused personal-email signups get no lifecycle email.
• The piece editor no longer crashes when a channel has no base URL.
• Project members see "+ New voice"; every admin write to a plan-info input drops its cache.

Under the hood

• Barrel type cycles broken, with an import-cycle ratchet so they cannot return. The refactor program's first pull request.

New

• Notes: per-item action-item destinations, a partner-meeting type, "Start another note" always visible, and a disabled Save that explains itself on hover.
• Content Studio: studio-wide text search; the review dialog shows author, voice and channel with AI targeting suggestions; promoted pieces arrive at review lint-polished, aimed, with meta and slug.
• Proactive drafting primitives: a readiness-gated drafting helper and a review banner. This is where proactive-by-default starts.
• Brand Voice: tensions and theme variants are A/B-testable in "Try it on real copy".
• Admin: merge two accounts of the same person; edit or delete worker profiles; Team Member Agreement tracking; a period picker for model cost rates.
• Marketplace: an agency team payment profile with an org-level Stripe account; workers confirm or set up.
• Talent: the OS avatar becomes the worker photo automatically, and the profile form takes a photo upload.
• Agent Activity: weekly streak lines for all four Syntropy roles.
• Signal: copy-to-clipboard on document text previews.
• Pricing page teases both value calculators with live headlines.

Fixed

• Pipeline-board cards open in a new tab and published cards always carry the live URL.
• Landscape avatar renders work, and finished promo videos land in Video Studio.
• Model cost rates no longer under-count past a thousand rows.
• A review script had been pulling the whole repository into every serverless bundle.

Published

• Ask Your Own Evidence a Question, Then Write the Answer Back.
• Journal: When One AI Grades Its Own Homework, You Get Theater.

New

• Prompt operations: a sweep engine (cohort sweeps, a three-model review panel, a routing audit, proposals that apply settings), a daily health score with impact tracking, provenance on every applied prompt version, and clickable prompt-to-skill lineage with shared version history. Skill metrics folded into the Skills admin as the Funnels tab.
• The skill-review loop actually runs now, driven by a background job.
• Signal: per-member Google Drive watched folders with an "Add Drive folder" button.
• Legal: a Team Member Agreement, the in-app NDA, auto-sent at first internal-domain login.
• Stripe Connect: the webhook accepts the Connect endpoint's signing secret.
• Mission Control: a pending AI panel review suppresses "Needs human".
• Every AI call logs its prompt-cache token split and a cache-aware cost stamp.

Improved

• Notes: a canonical URL after first save, a save-state chip, new-tab entry, an attendees picker, and Open in Gmail.

Fixed

• Two date-dependent unit tests were pinned to a clock; the nightly suite's seven-week red was burned down (six stale specs fixed, two real defects closed).

Under the hood

• Production builds run on the fleet too, and a changed acceptance spec must execute before its change merges.

Published

• Glints: Where Original POV Comes From When AI Can Derive Everything Else.

New

• Video Studio: click any recording to open it: player, transcript, extracted facts, and a guided "New edit".
• Content Studio: approving a piece can trigger a promo-video script and an avatar render; a pipeline kanban board sits on the Command Center.
• Signal: presentation uploads render their slides as images that ground Visual Brand; the drop zone lets you pin the document type at upload time.
• Personas: paste five to ten LinkedIn profile links and extract a persona brief; the section is now labelled "Fears & Dreams".
• Visual Brand: photo brand treatment in the Atelier; "Describe an example" reads a set of example images as one direction; brand treatment lives in the Library and Directions shows the most on-brand assets.
• Brand Voice: edit the guide directly, remove voted words, and apply devil's-advocate fixes one at a time.
• Onboarding videos end to end for admins: avatar plus voice, a server-side render worker, editable video mappings.
• Charts cap at five series on the radar and blue-ocean views.
• PM Control Requests tab: filters, finished items hidden by default, "Merged" versus "Live" status.
• Mission Control: fix instructions live in the decision card and screenshots ride every dispatch.
• Prompt operations digest links land on the named cohorts, the proposals inbox and the highlighted funnels.
• Admin: a credential registry with paste-to-rotate.
• Server stats: economics tiles for the self-hosted fleet (CI served, cloud avoided, energy, local share, agent work).

Improved

• GA4 AI Assistant setup steps a marketer can actually follow.
• Foundation checklist: a customer-list document type, profile links, a one-site floor, an office-document parse cap, and failed files surfaced.
• Signal inspector: Re-parse is always reachable.
• Dashboard caches invalidate at every write site, so their lifetime could go up without serving stale data.

Fixed

• Uploads: a private upload you can see is movable; "Only me" holds against superadmins, with a warning when a move to the org would strand the uploader.
• Hand-added Priorities enter the review-panel queue.
• Mission Control: "already fixed" outcomes converge as finished, ending the approve-again treadmill.
• Video playback tokens live 24 hours and admin preview pages self-heal on expiry.
• Academy: confirm before deleting rendered videos; the generated-video library keeps every render findable and restorable; Course and Module stay pickable until render, and orphaned renders can be re-homed.
• Support-ticket filters apply server-side, before the newest-100 cap.
• Six model token caps that were truncating long outputs were raised.
• The work-items reconciler stops reopening wrapped-up rows, and the flag banner's promised wrap-up action exists.

Under the hood

• Builds for the development branch now run on our own machines, with Vercel hosting the output.
• A schema-writing lease: one writer at a time across the machines, and production promotion is gated by it.
• New invariant: a re-created database check constraint must match the live schema snapshot.
• The self-hosted inference fleet gained a mid-tier model and swapped its large one.

Published

• The Document That Got Heavier: Grading Evidence by What It Produced.

New

• Audience: bulk stage changes, plus a "disqualified" stage that feeds the bad fit back into your ICP and personas.
• Max, the in-app assistant, reads URLs: it can open a published T2D3 piece or fetch an outside page, with injection and request-forgery guards on the fetch.
• Help: an explainer for organizations, projects and teams, learn links in the context switcher, and new glossary terms.
• Max's content panel: drafts compact beside the panel; direction locks are visible and you can overrule them.

Improved

• Visual Brand: illustrations and graphic elements are classified separately, each with its own style funnel and a retype control; the library re-sorts itself.
• Content Studio Command Center: "Inbox" became "Ideas", banners and tiles share one vocabulary, GA4 setup has its own call to action, and readiness shows its details.

Fixed

• Mission Control no longer shows "Stuck" under a row that reads "self-healing, no action needed".
• Agent Activity thumbnails follow the live brand asset, so a deleted hero image no longer renders as a broken tile.
• Audience contacts imported from prospect search save again.

Under the hood

• Per-pull-request static-analysis scans were dropped from the merge queue; they were advisory and starving it.

Published

• The Week Every Judgment Call Got a Skeptic, Wren's weekly.