downgrade fuse reads self-belief and self-belief is exactly what a silent fallback corrupts — switch to the get_session machine reading, and extend the fuse to 代裁 (measured live 2026-08-20)

[qq9340100]

Filed by the skills seat (session session_019T1sSZbQTnLhrK9HhNdNiB) from a live measurement the maintainer surfaced 2026-08-20 ~15:2xZ.

The measured failure

A triage-Routine fire was silently downgraded twice mid-run (transcript shows the harness banners: "Fable 5 is not available → This reply used Opus 5 1M" → "Opus 5 1M is not available → This reply used Opus 4.8 1M"). The session then reached the needs:contract-review sub-round and asserted, in its own words: "I'm running at claude-fable-5 (= CONTRACT_REVIEW_TIER, fuse passes)" — and proceeded toward reviewing contract cards at Opus 4.8. The maintainer interrupted it before any label was cleared.

Root cause: the 降档保险丝 (SKILL.md, needs:contract-review 复审链) says 「子轮开场自检当前模型」— but the self-check is an introspective reading, and introspection is precisely what a silent fallback corrupts: the fallback banners are rendered UI-side, the session's own context carries no authoritative signal, and the model "knows" only its configured identity, not its serving identity. A fuse that reads self-belief passes exactly when it must trip.

The authoritative reading exists — verified this session

get_session (claude-code-remote MCP, no args) returns external_metadata.last_served_model — the platform's own record of the model that served the latest turn. Verified live: this seat called it during the same incident window and got claude-fable-5 (matching reality); the reading is per-turn and survives mid-session fallback chains.

Scope — .claude/skills/pm-dispatch/SKILL.md (+ one platform-readings row)

1. Rewrite the fuse's self-check as a machine reading: the contract-review sub-round opens with a get_session call; external_metadata.last_served_model ≠ CONTRACT_REVIEW_TIER ⇒ the sub-round is skipped whole, labels stay — same safe-park semantics as today, only the reading changes. Self-belief is named as forbidden input (「⛔ 自述档位不是读数」).
2. Extend the same fuse to 代裁: the auto-adjudication confidence gate (置信门 ⑤ 「代裁分析跑在 claude-fable-5」) currently has NO fuse at all — a downgraded fire would silently adjudicate below the ruled tier. Same machine reading, same skip-and-park outcome (the card falls through to the maintainer path, which is always safe).
3. references/platform-readings.md row (134/134, net-0 dedup): sessions cannot detect their own downgrade from context; fallback chains can drop two tiers silently; get_session.external_metadata.last_served_model is the authoritative serving-model reading (also the undercover-mode identity path); the session_context.model field is the CONFIGURED model and must not be used as the fuse input.

Constraints

• Hard serial: SKILL.md is occupied by the in-flight tiering-revision PR (clause-① narrowing) — dispatch after it MERGES and write against the landed 模型分档/复审链 text; the two changes are adjacent in subject and must compose, not conflict.
• platform-readings.md is free (its sanitizer-row PR merged). Ratchet ceilings: both files headroom 0 — equal dedup, no raises.
• Governed: draft PR, review request os-zhuang, human merge. Protocol semantics ⇒ fable at dispatch (and if fable is itself unavailable at dispatch time, the quota exemption applies to the DEV run — the irony is noted — while the landed text will keep the REVIEW side machine-gated).

Interim caution (until this lands)

During fable-exhaustion windows, needs:contract-review clearances and 代裁 verdicts produced by Routine fires are suspect readings: verify the acting session's last_served_model (or its fire-time downgrade banners) before trusting them. Today's incident session cleared no label before interruption — nothing to re-review from it.

[qq9340100]

Claim: PM loop round 4
Session: session_019T1sSZbQTnLhrK9HhNdNiB
Branch: claude/issue-10300-machine-read-fuse
Worktree: objectstack-issue-10300
Domain: domain:skills
File surface: .claude/skills/pm-dispatch/SKILL.md (复审链降档保险丝 + 代裁置信门, 681/682 — 1 line headroom exists post-narrowing, further growth by dedup) + .claude/skills/pm-dispatch/references/platform-readings.md (one row, 134/134 net-0) (stop on breach; explain in the report)
Container & model: M governed protocol semantics, mode:subagent, model: claude-fable-5 (this dispatch's --tier under the NARROWED clause ①: SKILL.md main file = protocol-semantic surface, fable-mandatory)
Clause-②: no
Serial constraints cleared: PR #10311 (the prior occupant of SKILL.md, clause-① narrowing) is MERGED (16:32Z) — write against the landed 模型分档/复审链 text; the two changes are adjacent in subject and must compose. PR #10276 (dispatch-runbook.md) also MERGED (16:25Z). No open PR touches either target file (PR #10275 = scripts/pm entry guards, different files). Sibling #10294 (hotcrm lane registration, same SKILL.md) queued SERIAL BEHIND this card — named per fold-or-serial; folding rejected on gate ① (unrelated defect/fix shapes).

---

Generated by Claude Code

[qq9340100]

os-dev-report

{
 "issue": 10300,
 "status": "done",
 "branch": "claude/issue-10300-machine-read-fuse",
 "pr": "https://github.com/objectstack-ai/objectstack/pull/10363",
 "premise_still_valid": true,
 "summary": "Implemented the card-scope ruling verbatim, two files, net-0 lines on both. (1) The needs:contract-review sub-round's downgrade fuse now opens with a mandatory get_session call (claude-code-remote MCP, no args) reading external_metadata.last_served_model; a reading unequal to CONTRACT_REVIEW_TIER skips the sub-round whole with labels left in place, and the forbidden input is named in the clause verbatim (self-asserted tier is not a reading). (2) The auto-adjudication confidence gate item 5 gains the same machine-reading fuse with the same skip outcome: on mismatch the fire's adjudication is skipped whole and the card stays on the maintainer path. (3) platform-readings.md gains one measured row (sessions cannot detect their own silent downgrade from context; last_served_model is the authoritative serving reading; session_context.model is the CONFIGURED model, never the fuse input), paid net-0 by deduplicating the quota-wall bullet's sentence that duplicated the main file's standing wall-signal principle. Premise independently re-verified: the live get_session tool contract states last_served_model is the model that served the latest turn and session_context.model is the selected model. Composes with the clause-1 narrowing text landed the same day; that text is untouched.",
 "tests": "Gate list re-derived at final commit via node scripts/pm/dispatch-gates.mjs with no paths (merge-base derivation, 2 paths vs merge base 9a1ed7a09): exactly the 7 dispatched families, no additions. All green at head c83ae97, clean tree, exit codes captured before any pipe, verdicts from the gates' own printed lines: check:pm-skill-ratchet 'SKILL.md is 681 lines (ceiling 682)' and 'platform-readings.md is 134 lines (ceiling 134; headroom 0)'; check:pm-skill-id-lint '16 file(s) clean'; check:pm-governed-prose '2 instruction surface(s) name all 5 registered governed surfaces'; check:pm-governed-merges self-test '81 assertions'; check:skill-frame-sync '4 copies of the decision frame are structurally isomorphic across 3 files'; check:doc-authoring '380 files clean'; check:doc-formula-expressions '22 record-scoped formula example(s) judged clean' (dependency closure built first under the shared verify lock); check-nul-bytes 'OK, 6102 text files'. Reverse verification on the ratchet gate from the committed state, expected direction red-on-mutation: appended one probe line to platform-readings.md, mutation confirmed on disk (marker grep count 1, wc -l 135), gate red with its own line '135 lines; the ratchet ceiling is 134'; restored via git checkout from the branch (marker count 0, wc -l 134), gate green again. No build/dist is involved in this gate, so on-disk confirmation was the whole mutation proof.",
 "open_questions": [],
 "out_of_scope_findings": []
}

---

Generated by Claude Code

---

Generated by Claude Code

[qq9340100]

ACCEPT (reviewer of record: skills seat, session session_019T1sSZbQTnLhrK9HhNdNiB) — PR #10363.

Checklist verdict, verified against the actual diff line-by-line: (1) the contract-review fuse's opening self-check is replaced by the mandatory get_session reading of external_metadata.last_served_model, with 「⛔ 自述档位不是读数」 named in the clause, skip-and-park semantics untouched, and the CONTRACT_REVIEW_TIER single-source comparison preserved; (2) 置信门 ⑤ gains the identical fuse where none existed — mismatch skips the fire's adjudication whole, card stays on the maintainer path; (3) the platform-readings row records the measured two-tier silent fallback, the authoritative reading, and the configured-vs-serving trap, paid net-0 by a legitimate dedup (the deleted sentence restated the main file's standing wall-signal principle; the table-unique message shape survives). Both files at their ceilings (681/682 · 134/134), ratchet ablation red-then-restored in the predicted direction, premise independently re-verified against the live tool contract. Composes cleanly with the clause-① narrowing (adjacent text untouched). Deviations: none.

Governed terminal state (.claude/**): PR stays draft, review requested from os-zhuang, awaiting the maintainer's hand-merge. On MERGE, the last queued card #10294 (hotcrm lane registration) unlocks.

---

Generated by Claude Code