Signal, Noise, or Glint: The Three-Way Vote That Trains Your Marketing System

Every AI marketing tool will happily eat your documents. Drop in the pitch deck, the old messaging doc, the call notes, the competitor's whitepaper — the more the better, the demo says. Almost none of them will ask the question that actually determines the quality of everything the AI produces afterward: is this document worth believing?

That question is a judgment call, and it belongs to a human. In T2D3 OS, every document that enters your evidence base eventually faces a three-way vote: Signal, Noise, or Glint. It takes a second to cast. It is one of the highest-leverage seconds in the whole system — because the vote doesn't just file the document, it trains everything downstream that reads it.

Here is how the three-way vote works, why the third option is the one most teams have never named, and why we built an AI that recommends a verdict on every document but is structurally barred from casting the vote itself.

The three verdicts

Signal means: this is evidence about our buyer, our market, or our business, and future AI work should stand on it. A customer interview transcript. A win-loss note. Real pipeline data. The survey where your team disagreed about who the ICP is. Signal is what your ICP drafts, personas, messaging, and content will be grounded in.

Noise means: this should not ground anything. Not necessarily garbage — often it's a document that is merely about marketing rather than evidence of your market. A generic industry listicle. A stale strategy doc three pivots old. The fourth copy of the same deck. Voting it out is not deleting history; it is protecting every future generation from averaging over it.

Glint is the vote most tools don't have, and the reason we built the vote as a three-way rather than a thumbs up/down. A glint is a document — or a paragraph inside one — that fails as buyer evidence but is too interesting to discard: a contrarian opinion from your founder, a surprising observation from a sales call, an idea that cannot be derived from what already exists. Glints get voted out of evidence and into a separate lane reserved for original thinking.

Why noise is the expensive vote to skip

Generative models are averaging machines. Give one a corpus, and its output drifts toward the center of that corpus. If a third of your library is recycled industry commentary, your "AI-drafted positioning" will be one-third recycled industry commentary — delivered fluently and confidently, which makes it worse, because nobody flags it.

This is why we treat the noise vote as a gate, not a tag. A document voted noise is excluded from the evidence that grounds module generation. The practical effect on output quality is larger than almost any prompt improvement, because no instruction can fully compensate for a polluted evidence base. Marketing teams intuitively know this about people — you wouldn't invite the intern's summary of a blog post into a positioning workshop as though it were a customer quote — but most AI stacks flatten exactly that distinction the moment everything becomes "context."

Glints: where original POV survives

The glint lane exists because the averaging problem has a second, quieter victim: originality. A genuinely contrarian take scores badly as evidence — it's one person's opinion, unvalidated, often unrepresentative. A two-way vote forces a bad choice: keep it as evidence (and let one loud opinion tilt your ICP), or discard it (and lose the only raw material differentiated content can be made from).

So glints get a fourth lane in the architecture — not a fourth quality grade, a separate lane. In T2D3 OS, glints are consumed at exactly three seams: content generation surfaces, "wow"-posture ideation (the pieces meant to carry a novel point of view), and the brand voice module. They are returned as their own type, which means a glint can never quietly slip back into ICP grounding or persona synthesis. Your buyer evidence stays clean; your spiciest thinking stays available — each in the lane where it belongs.

If you take one idea from this article into whatever stack you run: separate what we believe about the market from what the market has told us. Both are valuable. Mixing them corrupts both.

How to run the vote on your own library

You can apply the taxonomy this week, with or without our software. For each document in your marketing drive, ask three questions in order:

1. Did the buyer or the market produce this, or did we? Interviews, reviews, support tickets, win-loss notes, usage data lean signal. Our own decks and docs need question two.
2. Is this evidence, or commentary? A claim traceable to something that happened (a quote, a number, a lost deal) is evidence. A claim that could appear unchanged in a competitor's document is commentary — usually noise.
3. Is there an idea here we'd be sorry to lose? If a document fails as evidence but contains a take you've never seen elsewhere, that's a glint. Pull the idea out, save it where your content team ideates, and vote the rest honestly.

Most libraries we see settle around a sobering ratio: well under half the corpus survives as signal. That's not a failure — it's the point. The teams that feel the difference in AI output quality are the ones who found out which half.

The AI recommends; the human votes

Here is the part we consider a design principle rather than a feature. Triaging a backlog of thirty or forty documents used to mean thirty or forty cold decisions — open the doc, skim it, remember what "signal" means, decide. That's exactly the kind of unaided judgment work an AI system should warm up.

So a background pass in T2D3 OS reads every not-yet-triaged document and writes an advisory verdict onto it: a recommendation, a confidence level, and a one-line reason. When you open the triage queue, every row already says something like "Noise — vendor commentary, no first-party evidence" or "Signal — direct customer language on onboarding pain." Your job shifts from deciding cold to confirming or overturning warm.

Two constraints make this trustworthy. First, the recommender is structurally barred from casting the vote: it never writes the triage status, no matter how confident it is. The vote stays one hundred percent human. Second, "your call" is a first-class verdict — when the model isn't sure, it says so instead of manufacturing confidence, and low-confidence recommendations are visually treated as open questions, not defaults.

We built it this way because the vote is not a chore to automate; it's the training data. Which brings us to the last piece.

What your vote actually trains

The moment you cast a verdict, three things happen. The document's quality grade recomputes immediately — a thumbs-up is human validation, and human-validated evidence carries more weight in every future generation that reads the corpus. The verdict lands in a running ledger of moments the system's priors improved, so "the tool got smarter this month" is a number you can inspect rather than a claim you're asked to believe. And the grounding for every downstream module — ICP, personas, messaging, content — quietly re-ranks around what you just confirmed.

One person's honest vote silently improves everyone's next output. That is the trade the three-way vote offers: a second of human judgment, compounded.

If you want to see what your own evidence would look like graded this way, the GTM diagnostic at t2d3.pro is where most teams start — it reads your public-facing message and shows you where the signal is thin, before you've uploaded a single document.