ADR-0121 — Evidence-gating is the load-bearing mechanic: ungrounded model judgment is not a review signal

Context

Suspec's spine is evidence-first. "No evidence, no acceptance" is the review model's participation gate (ADR-0095); the harness is reconcile-only and captures already-run evidence rather than judging (ADR-0077 D8); adversarial self-review re-reads the diff as a skeptic before done (ADR-0056). These were adopted as disciplines — the right way to work.

A July 2026 deep-research pass verifying the suspec-works roadmap (#74–#86) against primary sources found something stronger: the single most-replicated result in the current agent literature is that ungrounded model self-critique measurably degrades correctness, and the fix is grounding the critique on a reliable, actionable, already-run signal. That converts Suspec's spine from a stylistic preference into an empirically load-bearing mechanic — and it constrains the roadmap's new critique surfaces (Revolver Review, Bulletproof), which must not re-introduce ungrounded or consensus-based judgment.

Decision

A judgment on work — a review result, an accepted finding, or any critique a Suspec surface renders — is gated on reliable, actionable, already-run evidence, not on model confidence, agreement/consensus, or persona. The participation gate (ADR-0095) is not ceremony; it is the highest-value lever the evidence identifies.

• Ungrounded self-critique degrades. In a matched ablation, tool-grounded critiquing gives +2.1 solve-rate on GSM8k while the no-tool variant gives −1.8 — self-critique falls below the initial answer ([CRITIC-TOOL]); intrinsic self-correction degrades across rounds ([NOSELFCORRECT]); an ungrounded "are you sure?" challenge drops accuracy ~17% with a 46% answer-flip rate ([FLIPFLOP]). This is consistent with the existing canon: reliability needs an external signal ([SELFCORRECT]), agreement is not a correctness signal ([CORRELATED]), a model favors its own output ([SELFPREFER]).
• The lever is reliable, actionable feedback — not the mere presence of a tool. A raw check result earns its keep only when it is surfaced as an interpretable, structured signal a reviewer can act on; a tool trace nobody reads is not grounding.
• Corollary (persona-free critique). Because the lever is evidence, not identity, a critique surface drops the identity persona and keeps the explicit reasoning procedure — role framing helps only via its reasoning scaffold, not the role ([ZHENG-PERSONA], [KONG-ROLEPLAY]); this is already ADR-0093.
• Boundary preserved (reconcile-only). Suspec gates on already-run evidence; it does not spawn the checks. Running the project's build/test commands stays a non-goal (ADR-0077 D8, ADR-0086 Decision 5): the reviewer or worker runs the checks, Suspec reconciles the pasted result. "Evidence-gating" is therefore not "Suspec executes commands."

Level: convention — a positioning and doctrine commitment; it names no new tool. The partial mechanization already ships: suspec-cli's reconcile pass flags empty-evidence Pass rows (ADR-0086), which stays the toolable edge; the doctrine itself is convention.

Consequences

• Positioning. Suspec's evidence-first spine is its empirically-backed differentiator, not overhead — the docs and site may say so, cited to the primary sources now recorded in docs/research/sources.md.
• The roadmap inherits the gate. Any critique surface the roadmap adds — Revolver Review (suspec-works #77/#85) and Bulletproof (#86) — gates its findings on reliable, actionable evidence, never on consensus or persona, and speaks of reconciling already-run evidence rather than running a pre-gate.
• The evidence bibliography grows honestly. New verified claims land in docs/research/sources.md at the tier the honesty framework (ADR-0063) assigns — peer-reviewed findings load-bearing, 2026 preprints preliminary only.

Status

Accepted (2026-07-02). Refines ADR-0095 (participation gate) and ADR-0056 (adversarial self-review); honors ADR-0063, ADR-0077 D8, and ADR-0086 Decision 5; relates ADR-0093 (persona collapse) and ADR-0119 (independent-review invariant).