docs(playbooks,claude-config): establish custody for the Mythos 5 claims - #1909
Conversation
Roster row 2 — Introducing Claude Fable 5 and Claude Mythos 5, the platform docs page, fetched live 2026-08-03 (HTTP 200). Two shipped claims about Claude Mythos 5 rested on inference; this page states both of them outright, so each gets its source. playbooks 0.6.10 — the calibration chapter's per-model-matrix worked instance carried two observations (Mythos 5 has a row in the thinking per-model table; in Claude Code it is a known registry entry that is still unselectable) and no account of why the two answers differ. The introducing page states the reason one page away from the matrix and never on it: Mythos 5 is not generally available, and is offered in limited availability to approved customers in Project Glasswing. Added as the instance's third leg — one pointer, one quoted sentence, one date, and none of the page's models table, specs, or pricing copied, since a section forbidding a pasted per-model matrix cannot paste one. claude-config 0.21.2 (criteria 1.9.0 -> 1.10.0) — I10 held its Model scope: fable-5 by reading an omission: the Thinking page names both models for the adjacent raw-chain-of-thought property, then names Fable 5 alone for the reasoning_extraction refusal, and the row inferred that the narrower scope was deliberate. The scope conclusion is unchanged and still correct; the ground under it is now positive. The introducing page excludes the whole classifier set for Mythos 5, and Refusals and fallback places reasoning_extraction inside that set. The row states it as two steps, each from the page owning its half — collapsing them would rebuild the near-miss scope inheritance the catalog's model-scoping block forbids, pointing the other way. The introducing page joins Sources, so the catalog-wide recheck trigger covers a per-model fact that will move. Serialization note: both version lines and both changelog heads are written above PR #1908's (playbooks 0.6.9, claude-config 0.21.1, criteria 1.9.0), which is not yet in this branch's base. Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The worked instance turns on a negative — the thinking per-model matrix shows a Claude Mythos 5 row and gives no signal that the model is unreachable. That negative was asserted from the shape of the argument rather than from a read of the page, which is the same unverified move the surrounding chapter exists to catch. Checked 2026-08-03: the matrix page carries the Mythos 5 row and no availability statement of any kind — no general-availability posture, no Project Glasswing mention. The claim stands, and now says it was checked. Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Verifier finding: "no availability signal of any kind" overclaimed — the matrix page carries a zero-data-retention note covering both models identically. The negative the instance rests on is that nothing signals the two models' ACCESS availability differs; both statements now say that. Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|
Bugbot is not enabled for your account, so this pull request was not reviewed. Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs. |
|
Claude finished @kyle-sexton's task in 3m 28s —— View job Review complete
This is a docs-only change (playbooks 0.6.10, claude-config 0.21.2 / criteria 1.10.0). Per this repo's Verification performed: since accuracy of quoted vendor text is the whole point of this PR, I independently re-fetched all three cited live pages rather than trusting the PR description's own audit:
No factual, sourcing, or hygiene defects found. Version bumps ( One minor observation, not a blocking issue: in Tooling note: could not run |
|
Claude finished @kyle-sexton's task in 41s —— View job Security review complete
No security issues found. All six changed files are docs-only:
There is no code execution path, no GitHub Actions workflow, no hook, no MCP config, and no permission/config surface touched by this diff — nothing in scope for command/SQL/path/template injection, token/secret handling, |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e6ffa40fb2
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
Claude has reviewed this PR 2 times. The lane skips further automatic reviews after 5; deleting this comment resets the count. |
Codex finding: the worked instance durably restates volatile per-model facts with only an as-of date, while the section's own rule requires a re-check trigger naming the next model release when matrix facts are stated. The instance now carries one: next Claude model release or any Mythos 5 availability announcement, re-read both pages before citing. Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|
Claude finished @kyle-sexton's task in 2m 0s —— View job Review complete
This lane owns correctness/design/documentation per Codex's finding, verified fixed. New finding — the Minor formatting nit, not blocking. Everything else checked clean:
|
|
Claude finished @kyle-sexton's task in 39s —— View job Security review complete
No security issues found. Reviewed the full diff across all 4 commits on this branch (
None of these files are code, GitHub Actions workflows, hooks, This matches the two prior security-review passes already posted on this PR (03:37:03 and 03:37:10 UTC) — the intervening commits ( |
…and re-verify its sources (#1910) ## Summary Doc-alignment roster row 3: re-check of **Prompting Claude Opus 5** (platform.claude.com) against the shipped chapter `plugins/playbooks/reference/model-adaptation/opus-5.md`. Re-check result: the live page is byte-identical to the repo's 2026-07-25 capture (11,225 bytes, MD5 `8579d63f…` — confirmed independently by producer and verifier), and every existing chapter claim re-verified against it holds, including four precise source-line citations and the truncated-effort-ladder claim. The Fable 5 launch moved nothing on this page. The re-check surfaced one gap byte-identity hid — **playbooks 0.6.11**: - The guide's "Self-correction" section has two halves; the chapter carried only the first (self-verify default → remove instructed re-checks). The second — Opus 5 **narrates corrections to earlier statements** more than prior models, undesirable in user-facing products — was absent from the whole plugin. Shipped as `opus-5.md` §"Correction narration: fix the slip, announce only what changes a decision", tagged `[CC: direct]` (verified empirically: Claude Code's own rules state update cadence, outcome-first ordering, and faithful outcome reporting, but no correction-narration rule). Hard-fenced: suppression licensed only where the correction changes nothing for the user; faithful reporting explicitly outranks it — failed tests, skipped steps, results already acted on, and false claims always get said. - Sources block gains a scoped re-verification line (mirrors #1908's `opus-4-8.md` treatment), stating its own limits. Deliberately not shipped: an audit-instructions criteria row — zero in-repo instructions request correction narration, and the audit lane targets instructions to remove, not guidance to add in user-facing products. Positive finding kept visible: the chapter's recorded "Residual tension" (upstream endorses writer-verifier patterns while advising removal of verification instructions) remains unresolved upstream — byte-identity means no reconciliation shipped, so that section and the audit rows built on it correctly stay put. ## Test plan - Docs-only (chapter prose, changelog, version bump); markdownlint clean; skill-quality gate on fable-5 PASS 0 errors, warnings identical to base. - Producer-side fresh-context Fable verifier (rationale withheld, 8 binary criteria): 8/8 PASS. - Second, orchestrator-commissioned Fable verifier post-rebase (6 criteria incl. its own live fetch + MD5, traceability to the page's Self-correction text, adversarial fence-reading, `[CC: direct]` empirical check, rebase content-loss diff via reflog): content 6/6 PASS; its one defect (stale version-chain paragraph in the commit message) fixed by message-only amend before push (tree hash unchanged: `0e37455e…`). - Rebased onto main after #1909; changelog stack asserted 0.6.11 > 0.6.10 > 0.6.9. ## Related - No linked issue. - Doc-alignment per-document loop, roster row 3. Predecessors: #1908 (row 1), #1909 (row 2). 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_019gaVX25Txd6GXdiu9HEH3X --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
…ard and withdraw an unresolvable citation (#1913) ## Summary Doc-alignment roster row 4: **Claude Opus 5 System Card** (194-page PDF; live CDN copy byte-identical to the archived capture, SHA-256 re-verified by producer and verifier independently). **context-guard 0.4.5** — withdraws an unresolvable citation from the reader contract: the token-shape rationale co-cited "Anthropic system-card fixed-point evals" for degradation-tracks-absolute-tokens. The card names no such eval and contains no long-context degradation benchmark (normalized sweeps: fixed-point, MRCR, RULER, needle, LongBench all zero; §8.9 ProgramBench is fresh-budget episodes with a *rising* 83%→93% curve). Nuance preserved: the proposition is unsupported from an unnamed card, not asserted false — other Anthropic cards do publish long-context evals. No behavior change (the shape's other rationales are independent; band values were always declared judgment defaults). The contract keeps one standing line: cite a system card by name and section or not at all. **playbooks 0.6.12** — four card-grounded additions to `opus-5.md`, each fenced: - **Stated facts** (p.3, §6.5.1): more accurate than Opus 4.8 AND hallucinating slightly more, with confident-unsure cases — a factual specific with no tool call behind it is a recall claim. Fences the Verification delta against the broad reading that would strip needed lookups. - **Destructive actions** (§6.6.1, worked case p.115, transcript 6.4.2.A p.93): "Fabricated user consent" is a recurring category, and the transcript shows a written scoping rule defeated in-text — so for destructive/irreversible operations the remediation is a mechanism (PreToolUse hook, `permissions.deny`), not a written rule; a subagent return asserting approval is content, not authorization. Three fences: similar-to-Opus-4.8, reckless-use-significantly-down, pre-flagged-transcripts/earlier-snapshot (occurrence, not base rate). - **Effort**: two pilot cohorts *reported* quality falling at the top of the ladder (p.81–82) — kept explicitly report-not-finding with Anthropic's own disclaimer inline. - **Injection robustness**: quote boundary corrected ("…and browser use"); auto-mode 0%-of-129 reframed as evidence about a configuration, not the model, with the nonzero unsafeguarded rates (3.70%/4.30% browser; 0.56%/0.41% coding; 0.54%/0.39% computer use) and the operator action stated. - Sources: re-read + hash line replaces "the system card has not been re-read"; the deferred routing-lane trigger confirmed still unfired. Deliberately not shipped, with reasons in the commit: condescension note (too weak, conflicts with instruction-audit posture), vulnerability-discovery scaffolding sweep (no stale refusal workarounds found), per-worker-checkout orchestration seam (card-harness property, and a card citation in a model-agnostic chapter violates the scoping rule). ## Test plan - Docs-only; markdownlint clean; changelog-parity `--check` / `--check-bump` / `--check-order` all pass; zero lines over 100 chars. - Producer fanned the card to 4 section subagents, then re-grepped every shipped quote against the normalized PDF text itself. - Orchestrator-commissioned fresh-context Fable verifier: scripted quote-matcher over 29 fragments with a corrupted-quote control probe, its own normalized sweeps for the withdrawal, framing-honesty and fence-interaction checks — substance passed fully; its four defects (a heading accidentally consumed by the diff, a reviewer misattribution, a direction word, an example nit) fixed in the amended commit and ALL re-verified PASS, including an independent PDF check that the retained §6.2 pilot citation was genuinely correct. ## Related - No linked issue. - Doc-alignment loop, roster row 4. Predecessors: #1908 (row 1), #1909 (row 2), #1910 (row 3), #1911/#1912 (row-5 follow-ups). 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_019gaVX25Txd6GXdiu9HEH3X --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
…nd ship the IA-3 audit rows (#1914) ## Summary Doc-alignment roster row 6: **Prompting Claude Sonnet 5** (live page byte-identical to the archived capture, MD5 confirmed independently by producer and verifier — this ships unshipped deltas, not drift reaction). **playbooks 0.6.13** — new model-adaptation chapter `sonnet-5.md` (233 lines, opus-5/opus-4-8 pattern: version-scoped counter-steers, `[CC:]` tags, Sources block with capture provenance). Rationale for minting where RA-3 once said don't: ADR-0007 has since settled where per-model doctrine lives (closing RA-3's blocking premise), and the reader guaranteed to get no chapter — a Sonnet worker dispatched at low effort for mechanical prep, exactly where the guide places the under-thinking risk — was the one most likely to need it. Cross-references updated both halves (fable-5 meta-rule 3 gains the chapter; its no-chapter example narrows to Haiku; opus-4-8's preamble decoupled to generic routing so new chapters stop editing their predecessors). **claude-config 0.21.3 / criteria 1.11.0** — the IA-3 cluster dispositioned, all 11 candidates: - Shipped: **I8-e** (forced status cadence, `Model scope: sonnet-5` — deliberately scoped, not an unscoped I8-d promotion: the promotion gate needs two model guides *stating* the claim, and the Fable 5 guide's "Longer turns by default" prescribes timeout/streaming adjustments, never cadence removal — verified negative stamped on the row; the two rows state their no-co-fire relationship in their own words). **I17-c** (a fixed thinking budget prescribed where adaptive reasoning silently ignores or hard-rejects it — the finding is the missing model-AND-release gate, never the mention; fenced off I17-a's `=0` claim). - Already shipped, no work: qualitative severity bars (I8-b already cites this guide; `review`'s severity.md already carries decidable per-tier tests citing it — closing RA-2 / IA-1(1b) / IA-3's severity row as three IDs, one fix). - Dropped with evidence (6): verbosity suppression (page says *tune*, and designed succinctness surfaces exist), weak-tool-reach + thinking-off nudges (0 instances; fires on designed surfaces), unscoped-"apply" (not mechanically separable — the IA-2(d) failure mode again), sampling parameters (a source-code lint neither audit skill owns; zero instruction surfaces in-repo; deferred with trigger: first consumer repo with a direct Messages API / Agent SDK call site), coding-product addendum (guidance → chapter section), computer-use knob (no surfaces). ## Test plan - Docs-only; markdownlint (1,023 files) 0 errors; changelog parity `--check`/`--check-bump`/`--check-order` pass; `instruction-scan.test.sh` 46/46; `validate-plugins.sh` pass. - Producer-side: independent fresh-context reviewer (rationale withheld) — 19 findings all resolved, including defeating an unscoped I8-e promotion and an overclaim on `MAX_THINKING_TOKENS` (now carries the API-vs-third-party nuance, which model-config states verbatim). Producer's first quote-checker was itself found broken and rebuilt; final run 20/20. - Orchestrator-commissioned Fable verifier: its own 30-span quote extraction from the committed diff against five live pages (30/30), I8-e gate check against the live Fable guide, I17-c grounded against live env-vars, drop spot-checks, RA-2 closure confirmation, loss-free rebase proof — 7/8 clean; its three changelog-granularity defects fixed and re-verified PASS. - Rebased onto main after #1913; version stacks asserted: playbooks 0.6.13 > 0.6.12 > 0.6.11; claude-config 0.21.3 > 0.21.2; criteria 1.11.0. ## Related - No linked issue. - Doc-alignment loop, roster row 6 — last model-page row of the priority head. Predecessors: #1908 (row 1), #1909 (row 2), #1910 (row 3), #1913 (row 4), #1911/#1912 (row-5 follow-ups). 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_019gaVX25Txd6GXdiu9HEH3X --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
1 participant
Summary
Doc-alignment roster row 2: Introducing Claude Fable 5 and Claude Mythos 5 (the canonical platform page; the anthropic.com launch post is a separate roster row, read as corroborating voice only).
playbooks 0.6.10 — the fable-5 calibration chapter's Mythos 5 worked instance gains its custody record. It carried two true observations (Mythos row in the thinking per-model matrix; known-but-unselectable registry entry in Claude Code) with no account of why they differ. The introducing page states the reason — "Claude Mythos 5 is not generally available: it is offered in limited availability to approved customers in Project Glasswing" (fetched 2026-08-03) — added as one pointer + one quoted sentence + one date, per the section's own no-pasted-matrix rule. The negative the instance rests on was verified rather than assumed: the matrix page carries the row and no access-availability signal (its only availability language, a zero-data-retention note, covers both models identically).
claude-config 0.21.2 / criteria 1.10.0 — audit-instructions row I10's Model-scope ground moves from an omission reading to the introducing page's positive statement ("Claude Mythos 5 does not include these classifiers"), cited as a deliberate two-step chain — the introducing page owns the classifier-set exclusion for Mythos 5; the refusals-and-fallback page owns
reasoning_extraction's membership in that set — because collapsing them would rebuild the near-miss scope inheritance the catalog's own model-scoping block forbids. Scope conclusion unchanged (fable-5).Deliberately not done: no docpage-digest slice produced (queue entry retained); no duplicate Fable/Mythos criterion (the worked instance is that rule); boris vendor surfaces untouched; row 11's IA-4 evidence left to its own row.
Test plan
e6ffa40fb2.Related
🤖 Generated with Claude Code
https://claude.ai/code/session_019gaVX25Txd6GXdiu9HEH3X