ADR-0118 — Numeric/adoption claims must be ground-truthed by live query, not voted from recall

Context

The Phase 3 family-sweep (workflow wf9rvvwys) confirmed the re-architecture this program shipped was correct in file structure — 8→6 agents, 7→11 skills, mcp 0.2.0, the ADR-0112 catalog/kit split all landed — but the doc/reference layer drifted, and every recurring issue in the sweep mapped to a missing automated gate (a human caught it, not a check). One drift class surfaced in the research workflow itself: the deep-research run refuted a TRUE numeric claim — a real GitHub star count — and declared "no data exists," when the figure was live-queryable with one gh api call. The adversarial panel voted down a recalled number instead of checking it. That is refute-by-default ([adversarial-review], ADR-0095) misfiring: the stance is meant to withhold belief until evidence forces it, not to license dismissing a checkable fact without running the check. A queryable metric is exactly the regime where agreement/recall is the wrong instrument and a deterministic external lookup is the right one — the same deterministic-not-judge spine the framework already rests on (ADR-0063, ADR-0043: voting and self-consistency are not correctness signals; a resolving check is).

Decision

Any claim about a queryable metric must be ground-truthed against a live primary source before it can be marked Refuted or "no data exists."

1. What this covers. Queryable metrics are facts a deterministic lookup can resolve now: counts (stars, downloads, contributors), released versions, file/path existence, and similar primary-source figures. The primary source is the authority itself — gh api for GitHub figures, the package registry for versions/downloads, the filesystem for existence — never a recalled number, a cached summary, or a panel's collective memory.

2. No refute-by-recall. A queryable claim may not be marked Refuted on the ground that the number "looks wrong" or "isn't remembered." Refute-by-default withholds belief; it does not authorize disproving a checkable fact without checking it. Surprise about a number is a trigger to query, not a verdict.

3. "No data" is assertable only after a logged, failed live query. "No data exists" is a claim about the world that itself requires evidence: the actual query that was run, its source, and its empty/error result, recorded inline. Absent that logged failed query, the honest verdict is Unverified, not Refuted and not "no data."

Level: convention (method discipline). This rule is in force now for the deep-research and adversarial-review workflows by discipline + review — it is how those workflows are expected to run today. The toolable path is to wire it as a hard workflow step (a verify-before-refute gate on queryable-claim verdicts that fails the step when a Refuted/"no data" verdict carries no logged live query); that step is not yet shipped — until it lands, this is a review-checklist item, not an enforced gate (no enforcement-sounding claim is made here per ADR-0063).

Consequences

• Cost — a query per queryable claim before it can be killed. Refuting "the repo has N stars" now requires one gh api call (and a recorded result) instead of a panel vote; cheap in tokens, but a real added step. The win is that a true fact can no longer be voted down from bad recall.
• Sharper verdicts. "No data exists" stops being a polite synonym for "nobody checked." It now means a query ran and came back empty — and the gap between that and Unverified (the honest default when no query ran) is made explicit, which is the verdict-discipline gain.
• Bounded scope, by design. This binds only queryable metrics. Non-queryable claims — judgment, intent, "is this well-designed" — keep the existing refute-by-default treatment ([adversarial-review]); this ADR does not extend live-query duty to claims no primary source can resolve.
• Honest enforcement gap. Until the verify-before-refute workflow step ships, nothing mechanically stops a panel from skipping the query; the rule holds by reviewer discipline. Promote the wording from "convention" toward "toolable→enforced" only when (and if) that step lands.

Affected obligations / constraints

• Refines: the adversarial-review discipline / panel verdicting (ADR-0095) for the queryable-metric case, and reaffirms the deterministic-not-judge enforcement spine of ADR-0063 and ADR-0043.
• Does NOT change: refute-by-default for non-queryable claims, the verdict set, the artifact formats, or any check contract. This is a NEW ADR refining prior ones by reference (Nygard immutability) — it edits none of them in place.