Reproducible review examples

1. A sanitized change produces a specific finding

The off-by-one case changes pagination in synthetic src/pagination.py. The fixture provider replays a deterministic response; the engine validates locations against the supplied diff and emits a review result with per-file coverage (legacy evidence policy on this corpus).

The written document is a superset of the corpus baseline below: a run also emits the additive outcome fields review_status, coverage, and finding_lifecycles, so it is not byte-identical to the expected fixture. Compare the deterministic review content instead:

The authoritative check is the corpus evaluator, which normalizes derived metadata (finding_lifecycles, a complete review_status, and default coverage) before comparing against the expected fixture:

--no-learning-proposals is what keeps learning_proposals empty here; omit it and proposals from the fixture response are retained, as in section 3. Expected review content (schema: review-result):

{
  "summary": "The pagination change introduces an off-by-one error for page numbers greater than one.",
  "comments": [
    {
      "path": "src/pagination.py",
      "line": 3,
      "body": "The start index is correct, but the slice end should be start plus page_size.",
      "severity": "bug",
      "category": "correctness"
    }
  ],
  "provider": "fixture",
  "model": "fixture-v1",
  "learning_proposals": [],
  "evidence_policy": "legacy"
}

2. Authorized @sensei follow-up in thread

When conversation controls are enabled, an authorized @sensei mention in an existing review thread can receive a bounded, thread-aware reply. The processing reaction clears when the validated reply is published. This example is illustrative only — it does not link to a public comment thread.

Validated reply payload shape (schema: conversation-reply):

{
  "body": "`total_pages` is unchanged by this diff. The finding targets the slice end on line 3 only. A zero `page_size` is a separate guard concern not introduced here."
}

Conversation publication is a separate control from analysis. See getting started for the control matrix.

3. Finding disposition and learning are separate

Stable finding identities track disposition across review runs (new, still-present, fixed, outdated, uncertain). Resolving a GitHub thread or merging without a code change does not automatically mark a finding fixed. A later model omitting a concern is not proof the defect is gone.

Learning follows a separate lifecycle: proposed in the review result → draft learning PR → maintainer approval on the trusted base → future selection for matching paths. Approving a learning does not close a finding or imply the underlying code was corrected.

"finding_lifecycles": [
  {
    "fingerprint": "7c3b2a1f…b4a3",
    "state": "still-present"
  }
]

{
  "id": "pagination-slice-end",
  "status": "active",
  "scope": ["src/pagination.py"]
}

Inspect approved learnings offline (zero-write):

4. Partial / budget-limited review retains findings and blocks approval

When per-file coverage is present, budget-exhausted and other incomplete outcomes are explicit. Valid findings from reviewed paths are retained, but automatic approval is withheld when coverage is partial or fully_reviewed is false. There is no fabricated confidence score — only outcome fields.

{
  "review_status": "partial",
  "coverage": {
    "fully_reviewed": false,
    "files": [
      {"path": "src/pagination.py", "outcome": "reviewed"},
      {"path": "src/runner.py", "outcome": "budget-exhausted", "reason": "provider-call-budget"}
    ]
  }
}

Preview bounded coverage and zero-write plan output (no provider calls):

Approval blockers for this shape include coverage-partial and review-partial when publication and approval controls are enabled; incomplete enumeration raises coverage-incomplete / review-incomplete instead. These identifiers are defined in public contracts — treat that document, not this page, as their source of truth. Coverage blockers are independent of the blocking-finding finalizer, which separately withholds approval while unresolved blocking findings exist. See security & data flow for what approval does and does not guarantee.