Evidence-Gated Synthesis: What It Means and Why It Matters

Every AI product makes a choice about uncertainty. When the model is not confident in a claim, it can:

1. Guess. Fill in the gap with inference, present it as fact.
2. Omit. Drop the claim, leave a hole in the narrative.
3. Ask. Route the uncertainty to a human who would know, and capture the answer.

Most products guess. Some omit. Almost none ask.

Evidence-gated synthesis is the architecture of asking. Every claim in a synthesized brief must trace to one or more source signals. If the trace is missing, the claim is blocked. If the signal is ambiguous, the system asks. Only when a claim can be neither cited nor resolved by asking is it omitted.

This is not a prompt engineering trick. It is a product-level commitment that shapes every layer of the stack.

The Three Gates

Gate 1: Source Verification

Every claim must map to a specific signal: a Jira ticket, a GitHub PR, a Slack message, a calendar event, a file change. The mapping is not "inspired by" or "consistent with" — it is a provenance chain. The claim "payments API review is pending" links to the GitHub PR with the pending review request, the Jira ticket in "In Review" status, and the Slack thread where the reviewer was tagged.

Gate 2: Cross-Source Corroboration

A single source is not sufficient for high-confidence claims. The claim "sprint is at risk" requires at least two independent signals: a Jira ticket moved to next sprint, and a Slack message flagging the blocker. One signal without the other is noted but not asserted — "a ticket was moved, but we have not confirmed the reason."

Gate 3: Human Resolution

When signals conflict or are insufficient, the system routes an ask: "We noticed PR #142 was merged, but the corresponding Jira ticket is still 'In Progress.' Who can clarify?" The answer becomes a first-class signal, permanently attached to the claim. Next week, the same ambiguity resolves automatically because the context exists.

Why Guessing Kills Trust

A single confident fabrication zeros out trust in the rest of the brief. This is not a rational calculation — it is a visceral reaction. The founder who reads "the team is confident about the release date" and knows it's false will never trust the next claim, even if the next claim is accurate.

The cost of guessing is not just the wrong claim. It is the habit of verification that the founder must adopt: reading every claim skeptically, clicking every citation, reconstructing the truth themselves. A product that requires this is not a product — it is a slightly better search engine.

The evidence gate accepts a trade-off: briefs will sometimes feel conservative. "We don't know if the release is on track" is less satisfying than "the release is on track." But it is accurate, and accuracy compounds.

Why Omission Is Not Enough

Omitting uncertain claims is safer than guessing, but it hides gaps. A brief that omits the release risk because the signals are ambiguous does not inform the founder. It leaves them blind to a material uncertainty.

The ask loop exists because omission is not the right default for material gaps. When something matters and the system doesn't know, it should close the gap, not hide it. The answer to "why did this slip?" is valuable context. The omission of the question is a missed opportunity to learn.

The Architecture of Asking

The ask loop is not a chatbot. It is a routing system:

1. Detection: The system identifies an ambiguous or unexplained signal during synthesis.
2. Routing: It identifies the person most likely to know (based on role, recent activity, and prior ask resolution).
3. Question: It asks a specific, contextual question — not "what's the status?" but "PR #142 merged but Jira PROJ-142 is still open. Is there a follow-up task?"
4. Capture: The answer is stored as a high-trust signal, permanently linked to the claim.
5. Application: Future compiles use this answer automatically. The system gets smarter per customer.

Every answered ask is context that exists nowhere else — not in Jira, not in GitHub, not in Slack. It is the moat.

The Honest Standard

Evidence-gated synthesis is not perfect. The gate can be too conservative, blocking claims that a human would confidently assert. The ask loop can route to the wrong person, or the right person can ignore it. The system can miss signals that a human would catch.

But the standard is clear: cite it, ask about it, or omit it — never guess. A product that adheres to this standard will sometimes feel thin. A product that violates it will eventually feel wrong. Thin is fixable with more signals and better resolution. Wrong is fatal.

Cite it, ask about it, or omit it, never guess — this is the first of the four requirements defined in The Olmex Standard.

The Product Implication

This standard shapes every technical decision:

• Ingestion: Every signal must carry provenance (source system, ID, timestamp, author) so that claims can be traced.
• Model: The persistent graph must store confidence levels per claim, not just binary true/false.
• Synthesis: The compile pipeline must include an explicit gate step, not just a prompt that says "be careful."
• Delivery: The brief must surface citations natively, not hide them behind a "sources" link.

These are not features added to a synthesis product. They are the foundation of a trust product.