[aisona-lab]

What

Stage 3a reviews each hunk inside its file. Both earlier corpus runs showed most defects can't be seen from a hunk alone (R13 abstained 41%, R14 78%).

• --context works on lazycoder (API), packet and check. Each hunk is shown in a window of its post-change file, ±120 lines (CONTEXT_RADIUS).
  • Hunk lines are marked * after their number, and only those can be cited. Validation is unchanged; a test shows that a line visible only as context is rejected.
• Where the context comes from:
  • Diff: the post-change checkout (--context .). If the checkout doesn't hold the hunk at its lines, the command exits 3.
  • Corpus: corpus/seed.context.jsonl (272 KB), built by corpus_cli.py context with anonymous shallow git fetches. Windows for clean hunks come from the merge commit; defective ones come from its first parent. Every window is checked to contain its hunk byte for byte (37/37), and a test keeps the committed sidecar in sync with seed.jsonl.
• Context is part of block_id. Answers given without context can't be checked as answers given with it.
• Without --context, prompts and packets are byte-identical to before, so past runs stay comparable. A test pins this.
• No label leaks. I checked the packet: no SHAs, no rev, no ^1. The sidecar's rev field does reveal the label; this is documented in corpus/README.md, and the field never reaches the packet.

Run 3: pre-registered, then run

Commit c36da29 fixed the protocol and decision rule before the run: adopt at ≥ 6/15, inconclusive at 4–5/15, reject at ≤ 3/15 or a quiet rate under 80%. The first launch hit a usage limit before any answer was written, and I relaunched the same 8 shards unchanged.

	run 2 (hunk only)	run 3 (with context)
quiet rate	86%	91%
noise per clean hunk	0.14	0.14
hunk recall (gated)	3/15	5/15
rule recall	0/15	1/15
R13 / R14 abstention	41% / 78%	11% / 62%

Verdict: INCONCLUSIVE. --context stays opt-in and is not made the skill default. The next step is a larger corpus.

Reading the findings by hand, 3 of the 8 labelled fixes were described correctly (run 2: 2). That includes urllib3-5255 under R4, the first time the labelled rule itself fired in three runs. The full write-up is in docs/eval-runs/2026-10-23-agent-mode-corpus-run3-context.txt.

Not in this PR

• Batching per file.
• Passing --context from the Action.
• Making it the skill default (the verdict says not yet).
• Canaries carry no file context, so in context mode they can be told apart from real hunks. This is a documented limit.

Proof

env -u ANTHROPIC_API_KEY uv run pytest -q          # 211 passed
env -u ANTHROPIC_API_KEY uv run python scripts/corpus_cli.py prove corpus/seed.jsonl   # OFFLINE_PROVE: PASS
ruff check . && black --check . && mypy src        # clean

🤖 Generated with Claude Code

https://claude.ai/code/session_01WkPsqEJ4QYWRW5eCFBYRPS