Stop 23 (Phase 8 — unattended agents, gh-aw): closed at n = 0, and the instrument built to catch a stale quote had five defects that manufactured them

[hermanngeorge15]

Closes the ◇ extract-only stop at spine position 23 — Phase 8, unattended agents (gh-aw) — at n = 0 runs, which §3's itinerary row 22-23 registers a ◇ stop to produce and is not a shortfall. Counted rather than asserted: 0 of 740 runs on the API carry a stop-23 experiment key.

Two results, and the second is the sharper one

1 — a verbatim quote in this project's own extract went stale in 49 days and nothing executed to catch it. The August 2026 extract of this same phase carries, as a bold display quote, "The agent never receives write tokens directly." That sentence is not on the page it cites. The page's equivalent today is "all without giving the agentic portion of the workflow any write permissions" — the claim survived the rewording, the quotation did not. Two other August quotes were re-checked and are still verbatim, so this is one sentence, not a rewritten page. check-links.sh proves a URL resolves; nothing proved a quotation still appears in the resolved page.

2 — the instrument built to close that gap had five defects, all of one shape: a failure to fetch or parse, reported as documentation drift. Found across two §4a rounds, every one proved against the pre-fix script rather than asserted:

#	Defect	Pre-fix behaviour	Fixture
1	curl had no --fail	a 404 serving a non-empty body exits 0, strips to real text, every quote on that page prints ABSENT at exit 2	I
2	empty cached page accepted	exit 2, found=20 absent=9 — while an empty live response or fixture exits 3. Same content, two verdicts	J
3	undeclared page key	exit 2, found=20 absent=9 — a one-character typo, 2>/dev/null swallowing the error	K
4	strip had no failure guard (the acceptance gate's blocking finding)	exit 2, found=0 absent=29 — the entire extract declared fabricated	L, M
5	strip keeps page chrome	DISPUTED in writing and registered as a scoped limitation in the script header	—

One empty cached page manufactured nine phantom stale quotations; one broken interpreter or one non-UTF-8 byte manufactured all twenty-nine — at exit 2, in the same format as a real finding. That is this project's house failure mode aimed squarely at the tool built to catch it, and the same shape as the stop-8 contamination guard that would have excluded the treatment arm and reported a null.

After the fixes: ShellCheck exit 0 on both scripts, 16 of 16 fixtures over all four registered exit codes, and the live run still reproduces found=28 absent=1, exit 2, same absent sentence — so the headline is unchanged by five fixes to the instrument that produced it, which is the only reason it may still be stated.

3 — the same layer mistake as stop 22, again, caught by the same reviewer and not by me. staged: true was labelled L1; applied in order the write request is still expressed and something executes and skips it → L2. Exactly one of ten subject rows survives step 1 (the safe-output separation), not two.

§4a review — two rounds, and no acceptance verdict at round 1

• findings/opencode/review-README-20260927T201538Z.md — 23 136 B, 44 sections, the workbook. ## Acceptance reads "The gate failed to run (opencode exit 1)" → recorded DID NOT RUN, not UNDECIDED and not a pass.
• findings/opencode/review-verify-quotes-20260927T202016Z.md — 6 555 B, the tool, round 1.
• findings/opencode/review-verify-quotes-20260928T040312Z.md — the FINAL review file, round 2, and the gate RAN here (lab-acceptance · minimax-m3) and returned REJECT on defect 4. Fixed in fbf9d7d; round 3 not run, loop closed at two per §4a's cap.

Both panels ran clean (codex ok 62s/63s, 24s/25s), so the findings are results, not stalls. One family only, so recurrence reads 1/1; the four findings both runs raised are stated as 2 of 2 runs.

Per finding: the sha that fixed it, or the dispute

Finding	Disposition	Sha
staged: true labelled L1	fixed — relabelled L2, every dependent count corrected	f7b1e42
roles: treated as reaching all thirteen triggers	fixed — scoped to actor-originated triggers; the scheduled path is not role-gated at all	f7b1e42
vocabulary → "L3 perimeter" claimed more than was read	fixed — scoped to the documented vocabulary, unmeasured for an already-compiled .lock.yml	f7b1e42
"29 network fetches per invocation"	fixed — seven fetches (one per page), 29 sentences searched locally	f7b1e42
redacted 19:2xZ identifies no execution	fixed — cites the artefact instead	f7b1e42
"gh-aw's is the right one" (fail-closed)	fixed — narrowed to a check whose result is quoted as evidence	f7b1e42
Commit block lists files this stop did not create	fixed — annotated as the deferred labs' deliverables	f7b1e42
least-used variant tie-break	fixed — scoped; only the confound's existence follows, not its shape	f7b1e42
compile-step correspondence check	disputed — the finding restates the section's own conclusion as an omission, and building the check would be an instrument for a workflow this project does not run (§6)	f7b1e42
the jq pattern is broader than the stop's key set	disputed — a superset matching nothing proves the subset matches nothing; the looseness is deliberate	f7b1e42
strip keeps page chrome	disputed — registered as a scoped limitation; body extraction needs a per-site selector that breaks on the next redesign, and these are full sentences	fbf9d7d
curl --fail, empty cache, undeclared key	fixed + fixtures I/J/K proving refusal	f7b1e42
strip failure guard (gate: REJECT)	fixed + fixtures L/M proving refusal	fbf9d7d
five deferred-lab underspecifications	registered as debts, not fixed — designing a deferred lab's decision rule is a future step's artifact (§6)	f7b1e42

Not sent to the harness, and named here as §4a requires: HANDOFF.md and SOURCES.md (neither is a contract this stop authored), and nothing under findings/, .claude/ or .github/ — rtk hides dotfile paths and the harness would review nothing and exit 0.

§5 validation

The table is in the workbook, one row per gate clause, with the layer of the proof rather than of the artifact. Every command in its re-derivation column was re-run immediately before the section was written, and the output quoted is that run's:

./evidence/p08/verify-quote-checker.sh    16 passed, 0 failed          exit 0
./evidence/p08/verify-quotes.sh           found=28 absent=1            exit 2
./tools/check-links.sh                    ok=73 moved=11 blocked=2 unverified=0 broken=0   exit 0
shellcheck evidence/p08/verify-quotes.sh evidence/p08/verify-quote-checker.sh              exit 0

Independence check: there are no arms to be independent of. One clause of the phase's own exit gate — "read-only default" — is answered qualified, because it is false without a subject: the agent job is read-only, the workflow is not (with no safe-outputs: block, create-issue is still enabled at max: 1). "Was this the agent, or the harness?" is neither.

Expected red

The board check is RED by the author's standing decision 12 item 4 — the boards are the author's and are not republished here. Required marker digest, re-derived by running ./tools/check-board-freshness.sh after the final HANDOFF.md edit rather than by hand: c1e2b847fce9. Nine green and one red is the mergeable state, as lab#132, lab#130, lab#122 and lab#117 all were.

Deferred

Labs 8.1–8.4 are all DEFERRED with their debts now written down, so lab#10 stays open — a Phase issue closes only when its gate is met from measurement.

Validator files processed

All 22, newest findings/track-b-validation-2026-09-08-4.md; none new since, re-checked one grep per basename.

🤖 Generated with Claude Code