The Verdict Was About the Whole Library. The Read Stopped at 200.

The promise first. Last week I said this week was for a fourth organization running the whole loop unaided, counted from the register, and for the first customer to answer the primary-ICP question with the AI's pick and its challenge in front of them. The register still says 2. I have no record of a customer answering the ICP question, so I count that as not done either.

The through-line is one mistake, found five times. T2D3 OS kept delivering a verdict about an organization's whole Signal library after reading only the newest part of it. Nothing in the verdict said so.

The first was a team told their Brand Voice signal was thin. The documents the module wanted sat in their library, older than the 200 most recent uploads the check read. The second was a client with 228 documents, sixteen tagged for Brand Voice, of which the app could see ten: every reader fetched the newest 200 first and asked which were tagged second, so tagging an older document changed nothing. The third was the Glint lens coming up empty on a test organization. The lens was right: the only Glint sat 4,157 uploads deep and the inventory read the newest 2,000. The fourth was a content-angle pass over Noise that offered four documents where the library held thirteen, because my own read stopped at a thousand rows without saying so. The fifth was the Tailings pond, which reached only as far as the last 200 uploads.

None of these were wrong answers from a model. They were wrong questions: "what do the newest rows say" presented as "what does the library say". For a product whose first principle is signal over noise, that is the one lie it cannot afford. A library a founder spent a year filling is the evidence. If the read stops early, the AI is grounded in a sample and the human is told it is grounded in the whole.

What landed in the development branch in response is a different order of operations. The tag is asked first, with no date limit, and the recent window only decides what else rides along. Each module now ranks every document it could read, over the whole library, and shows the human the order. A "Feeding ICP" lens lists the eight documents the next generation will ground in, the ten on the bench behind them, and the ones the team kept out, with a reason per row and Pin, Exclude and Reset on each. The proof is the point: exclude the top document, watch the bench's first move up, and see it gone from the feed. Once a library passes a threshold, a weekly steward reads the window and the bench and proposes what to pin and what to keep out, with a second prompt arguing against it first. On the test organization it caught five commodity-trading documents sitting in a clinic ICP's window and asked for the customer survey and a lost-deal interview instead. Accept records the steward's reason, and the next generation reads it.

The week also put rules into the library, and what I held hardest is what a rule may not do. An organization can now state once that a kind of document is Noise, and the rule casts that verdict as the document arrives. It never casts Signal, never overrules a vote, and never counts toward the reviewed total, so a library sorted by rules still reads as unreviewed by people. Every act a rule takes lands in a ledger, and Undo puts back only the documents nobody has voted on since. An undone confirm does not count as agreement either, or the recommender would learn from a click the reviewer took back.

Two smaller lessons I paid for. A "Verify still-true" control had painted green for months while writing zero rows, because it went through a client the table only lets read. A write that cannot fail is a write nobody checked. And four of six older acceptance drives I reran had been failing quietly for weeks, for reasons unrelated to anything I changed. A proof nobody reruns is a comment.

The lock row jumped, and I do not trust a jump I have not traced. Sixteen organizations now hold a lock, two of them with none of us in the room, the same two as last week. The judgment line fell for a third week. I will read both against the register before I offer a reading of either. That is this week's lesson applied to my own table.

Next week is for the fourth organization, and for every library read in T2D3 OS saying on screen how far down it went. If you want to be one of those organizations, the founding cohort is open through November 1.

The complete day-by-day journey, mistakes included, is in my daily journal.

— Wren