---

type: adr id: adr-0086 status: accepted created: 2026-06-19 updated: 2026-06-19

ADR-0086 — Deterministic review scanning is the shipped reconcile wedge: accept C014, gate the measurement track, reject the off-boundary asks

Context

Two strategy reports proposed making "deterministic review scanning" — a git-diff/risk reconciler that reads the task/spec/change-plan, reads the diff, and enriches the review with evidence gaps and human-attention items — a major new direction for the corpus CLI. Examined against the code, the proposed heart of that capability already ships: corpus review reconciles, on a verdict-free report (ADR-0077 Decision 8), coverage (C012), verify-evidence binding (C013), scope divergence, the run-summary↔diff self-report mismatch (both directions), empty-evidence Pass rows, and the packet-structural facts. So the reports are mostly a decision, measurement, and positioning agenda, not a build list. This ADR records what to accept, defer, reject, and correct.

Two findings shape the decision.

The differentiation must be re-grounded, not retracted. The competitive landscape moved: AI review tools now ship requirement-binding. CodeRabbit validates a PR against the acceptance criteria of a linked Jira or Linear issue and writes the assessment back to the ticket [CODERABBIT-PRVAL]; Qodo Merge's Ticket Compliance Agent fetches ticket context, reports "missing acceptance criteria" and a Fully/Partially/Not-compliant level, and markets scope-creep prevention and audit-ready evidence [QODO]. "No tool binds evidence to requirements" is therefore false. What stays distinct — the form corpus's differentiation leads with — is: corpus's reconciliation is deterministic (no model, reproducible, exit-coded 0/1/2), keyed to a local structured spec/task packet (not a remote tracker ticket), verdict-free (the human owns Pass/Fail/Unverified/Blocked, ADR-0077 D8), and durable in git (a persisted, independent review packet, not an ephemeral PR comment).

A precision target the reports got wrong. The proposal stakes success on keeping false-positive human-attention items "under 30%." Google's field experience sets the bar far tighter: a code-review-time check must "produce less than 10% effective false positives," where an issue is an effective false positive "if developers did not take some positive action after seeing the issue" — technical correctness is secondary to whether the developer acted [GOOGLESA]. A 30% rate is three times the documented abandonment threshold; on the strongest available evidence it would get the check --no-verify'd into irrelevance — the same noisy-check death spiral the SMELLS-precision rule already keeps fuzzy checks at warning to avoid (ADR-0083).

Decision

Six points. Each carries its honesty level (ADR-0063).

1. Name the wedge — positioning, not a build. "Deterministic review scanning" is the already-shipped reconcile capability of corpus review, not a new subsystem. Its differentiation is the four-property form above: deterministic · local-spec/task-keyed · verdict-free · git-durable. Level: convention (positioning).

2. ACCEPT C014 do-not-change-touched (warning). A changed file matching a task packet's ## Do not change entry is surfaced as a protected-path fact routed to Human attention. This closes a verified gap: ## Do not change is a required task-packet section, but nothing reads it — a touched protected file is caught today only indirectly via outsideScope, which misses a protected file that lies inside the declared Affected areas. The match is closed-value (an exact path/prefix compare, the same matcher and {{placeholder}}-skip as Affected areas); the intent is human (whether the touch was justified is the reviewer's call, ADR-0077 D8), so it ships at warning, the same fact-class and severity as outsideScope, with a recorded path to promote once field-tested (the ADR-0079/0083 conservative precedent). Level: toolable. Detailed in Propagation.

3. ACCEPT the measurement track as the gating next investment. A review-gate benchmark — precision and recall on a seeded corpus of scope-drift, do-not-change-touch, and claim-vs-diff cases — is the next investment, and it doubles as the real-world test of the corpus review --json surface the MCP adapter consumes (ADR-0085). It is recorded here as committed-next; it is a program, cut as its own corpus-works spec, not built by this ADR. Level: convention.

4. DEFER behind a measure-first gate (each earns a build only once the benchmark shows the review gate catches real failures at ≤10% effective false positives): SARIF 2.1.0 (a ratified OASIS standard [SARIF]) / JUnit XML (a de-facto test-results format) import-and-correlate (the future shape is route-and-correlate-against-scope, never re-implement an analyzer); a mechanical risky-path matcher (the trigger taxonomy already ships as the trigger-coverage human-attention checklist — mechanizing it moves a rule from checklist to toolable, exactly the precision-minefield the gate must clear first); and project-policy config (which, if built, reuses the existing .corpus/config.yaml home, not a parallel file). Deferred, not rejected.

5. REJECT as non-goals.

   • corpus verify executing the project's commands. Running build/test commands crosses the ADR-0077 D8 reconcile-only boundary ("never the sandbox/container runtime") and strains corpus-cli's deliberate two-dependency footprint (ADR-0085). Capturing already-run evidence (the pasted output C013 reads) is in-bounds; corpus-cli spawning the commands is not. If ever wanted, it belongs in a corpus-* PATH plugin (ADR-0077 D3), never the core.
   • Per-language analysis adapters. Eight ecosystem toolchains as dependencies invert corpus-cli from a markdown reconciler into a polyglot analyzer — the "static analyzer" the reports themselves disclaim — and re-introduce the architecture-enforcement bet the validated direction refuted. The recorded answer is BYO: a team binds its own analyzers via its own gate; corpus-cli reconciles only.
   • A rival SCOPE-/EVIDENCE-/RISK-/CHANGE-/REVIEW- check-ID namespace. The proposed scheme re-labels facts the C0xx contract already emits (scope drift = scopeDivergence/outsideScope; empty-evidence Pass = emptyEvidencePassRows; change-plan = C010/C011; packet structure = packetStructural). A parallel namespace forks the single, drift-guarded contract id space. Any genuinely new fact extends the C0xx series (as C014 does), never a rival scheme.
   • A standalone corpus scan verb. The "deterministic facts, no execution, no verdict" contract it promises is the shipped corpus review. A second verb splits one reconcile surface into two; net-new diff facts land inside corpus review.
   • An "Agent Work Protocol" category coinage. The term is unused in the market (the buyer-facing term is "spec-driven development"), "Protocol" already names interop wire-formats (MCP, ACP) corpus does not ship, and no user-facing doc carries it today. corpus keeps its shipped identity — "a lightweight spec and review workflow for teams using coding agents" — sharpened on the reviewable-evidence angle, not a new noun.

6. CORRECT the precision target. The benchmark (Decision 3) measures against ≤10% effective false positives [GOOGLESA], adopting the "no positive action taken" definition as the metric, not the reports' <30%. Recall (of seeded failures, how many the gate surfaces) is measured alongside, but precision is first: recall pressure never promotes a fuzzy check to hard error (the ADR-0083 split holds — closed-value/git-fact checks may earn hard error; prose-shaped checks stay warnings).

Alternatives considered

Alternative	Why weaker
Build the reports' L1 contract-reconciliation layer	It already ships in reconcileReview.ts (coverage, verifyBinding, scopeDivergence, selfReport, emptyEvidencePassRows, packetStructural). Building it again duplicates the engine. Rejected.
Adopt the reports' new check-ID namespace	Forks the single C0xx contract and re-labels shipped facts under a rival scheme (Decision 5). Rejected.
Fold do-not-change into C012 or leave it to outsideScope	C012 is an id-set coverage reconcile, a different fact class (the same separation ADR-0083 drew for C013 vs C012); outsideScope misses a protected file inside Affected areas. A clean C014 keeps the cite clarity and the distinct fact. Rejected.
Build corpus verify / language adapters / SARIF import now	Crosses the reconcile-only boundary and/or the two-dependency footprint, or front-loads a deferred layer before the measure gate (Decisions 4–5). Rejected/deferred.
Accept the reports' <30% false-positive target	Three times Google's field-validated ≤10% effective-FP abandonment threshold [[GOOGLESA]]; a 30%-noise check gets ignored. Rejected (Decision 6).
Adopt "Agent Work Protocol" as the category	An unused coinage that collides with MCP/ACP and contradicts the shipped tagline (Decision 5). Rejected.

Consequences

Accepted. C014 is the one new check this ADR mints; everything else is a positioning correction, a recorded deferral, or a recorded non-goal. The reconcile-only boundary (ADR-0077 D8) and the two-dependency footprint (ADR-0085) are reaffirmed: C014 surfaces a fact routed to Human attention, never a verdict, and adds no dependency.

Honesty level: C014 is toolable — a future corpus review/corpus check surfaces it; until a team wires its CI to that output, nothing is enforced (the gate is the team's, never corpus's, ADR-0063). The positioning narrowing (Decision 1) is a convention; the differentiation prose is corrected, not enforced.

Positive: the review wedge gains one clean, deterministic, in-boundary fact; the strategy is re-grounded honestly against incumbents that now ship requirement-binding; the precision bar is set to the evidence, not a guess. Negative: a second scope-related fact to teach beside outsideScope; a contract-version bump. Neutral: a team may treat the C014 warning as blocking by its own CI policy — the team's gate, not corpus's.

This refines ADR-0077 (D7 names deterministic review scanning as one of the two wedges; this names it as the shipped one and holds D8's verdict-free boundary for C014). It builds on ADR-0079 (C012) and ADR-0083 (C013) for the conservative-shipping + coordinated-landing pattern, and relates to the evidence-validated direction recorded in the corpus-works workspace (the review-gate-teeth + measure-first track).

Propagation

The C014 mint lands coordinated (the ADR-0079/0083 rule: the checks.yaml rule + version bump ship with the corpus-cli implementation, so the drift guard never reds between commits). Docs-first single-sourcing: the human-readable contract and this ADR land first; the checks.yaml data + the CLI move in lockstep.

• The C-id mint — C014 do-not-change-touched (warning): a new row in checks/checks.yaml and its definition in reference/checks.md (the core-checks table + the Warning row of the severity split) and the one-liner in reference/cheatsheet.md. A V17 violation fixture under checks/fixtures/. Core C-ids are not a registered cardinality, so no closed-set count moves.
• The contract version bump — 0.6.0 → 0.7.0 in checks.yaml, moved in lockstep with corpus-cli's pinned CONTRACT_VERSION so the drift-guard test never reds in between.
• The corpus-cli build. parseTaskPacket gains a ## Do not change reader (sharing the Affected-areas extraction, including the {{placeholder}} skip); reconcileFacts gains a do_not_change_touched fact (matched per-entry, so an empty Do-not-change list surfaces nothing); reconcileReview threads doNotChangeTouched onto the verdict-free ReviewReport and into the warning level; the human-attention render gains its bullet. Fixtures add the do-not-change-touch case (incl. the file that is inside Affected areas yet still protected). corpus-mcp's ReviewReportSchema may mirror the new field additively (safe against its drift tripwire; not required).
• The differentiation prose narrows onto the four properties in README.md (the neighbor map). ADR-0060's gap claim — a persisted, independent, exception-routing review packet — is already correctly scoped (it never claimed nobody binds evidence to requirements), so it stays as recorded; this ADR sharpens it forward rather than rewriting it (the ADR ledger is append-only). The research bibliography gains the verified entries cited above ([GOOGLESA], [CODERABBIT-PRVAL], [QODO], [SARIF]).
• The measurement track and every deferred layer (Decisions 3–4) are FUTURE — cut as their own corpus-works specs, gated on the benchmark. This ADR builds none of them.

Refines ADR-0077 (D7/D8); builds on ADR-0079, ADR-0083; honors ADR-0063 (honesty levels) and ADR-0060 (the differentiation answer it sharpens). Does not re-open the spec-scoped-id question (closed by ADR-0080).