Files
archie/docs/prds/review-calibration.md
T
2026-09-26 02:05:52 +10:00

3.3 KiB

Review calibration -- decision

Status: Draft Beads issue: archie-core-h019.5 (parent archie-core-h019)

Superseded by docs/prds/pr-review-agent.md.

Decides the half of calibration docs/prds/adversarial-self-review.md left open: "the adversarial-confirmation mechanism that keeps a finding from surfacing on a single reviewer's say-so."

The problem, precisely

The shipped contract (internal/domain/workflow/review_findings.go, h019.2) already makes only a confirmed error-level finding block (Blocking() == Verdict == confirmed && Level == error), and a zero-finding review is already distinguishable from one that never ran (ReviewStatus). What is missing is the other half: a reviewer can report nothing and say only "I found nothing" in free text. The operator cannot tell "I checked the five things that matter and they are clean" from "I did not look." That is the failure the epic names -- a findings surface that consumes attention without earning trust.

Decisions

1. "Checked and cleared" becomes structured, not prose

ReviewReport gains Checked []ReviewCheck:

type ReviewCheck struct {
    Property string // what was verified clean, e.g. "nil-safety of Foo callers"
    Evidence string // how it was verified, e.g. "read all 4 callers; each nil-checks"
}

Evidence is required, for the same reason ReviewFinding.FailureScenario is: a check without evidence is an assertion, and the whole point is to separate a verified property from an unexamined one. A zero-finding review is credible only when it lists what it checked.

2. Report checks through a typed tool, like findings

The reviewer calls a new capture tool record_checked (mirroring record_finding) once per property it verified clean. The same environmental argument as findings applies: the reviewer cannot end its turn having silently skipped the checklist. ReviewReport.Summary stays as the human overview.

3. Style nits are out of contract

A finding must describe wrong behaviour or a crash. Style, formatting, naming, and preference nits are explicitly out of contract -- they are the padding that trains the operator to skim. Stated in the reviewer prompt and in the ReviewFinding type's own documentation. Not machine-enforced: deciding "is this a style nit" from free text is unreliable, and a wrong rejection is worse than a tolerated one. The category enum already has no style bucket; the rule closes the gap that would otherwise let one hide under other.

4. Surfacing

The PR-body findings section lists the cleared properties alongside findings, so "ran and found nothing" arrives with its evidence. (Park reasons keep rendering only blocking findings -- what a human needs to act on, not the full checklist.)

What this does not decide

  • Adding a second, independent reviewer pass. The verdict distinction plus the recorded checks is the confirmation mechanism; a second model pass is a separate, heavier decision and nothing in h019.5 requires it.
  • How many findings can survive before MaxSteps is exhausted (a budget question, already satisfied by the per-stage budget).

Packages this touches

  • internal/domain/workflow -- ReviewCheck, ReviewReport.Checked, validation.
  • internal/app/agentworker -- record_checked tool, prompt rules.
  • internal/domain/workflow/review.go -- render cleared properties on the PR body.