docs: ship DOC bundle 4 — matrix pointers, archive-reading conventions, residual-risk footer - #1898
Conversation
Adds a volatility axis to the fable-5 calibration chapter, beside the surface axis and the channel axis its neighbours already own. A per-model table reads as a fact and is actually a snapshot: a row is added or a default flips with each model release, and a pasted copy stays confidently wrong with nothing in it saying which day it was true. The rule is point-at-the-owning-table. For thinking configuration that table is the per-model table on Troubleshooting thinking, re-fetched 2026-08-03 (HTTP 200), which is the authority on what each model accepts, defaults to, and rejects. A matrix stated anyway — because the reader cannot act without the values in front of them — carries a re-check trigger naming the next model release, so a stale row is found by a scheduled read rather than by a reader acting on it. The fourth rule ties the section to its neighbour: a vendor matrix is an API-surface fact, so presence in the table is not reachability where the reader is running. The worked instance is verified on both sides 2026-08-03. Claude Mythos 5 has its own row in that table, and in Claude Code it is a known model in the registry with full gating machinery and still not selectable — no alias resolves to it, it is absent from latest_per_family, it declares no capabilities, and it exposes no picker row. Its registry entry carries exactly one non-null provider id (first_party) beside seven null siblings; the count is stated at seven because the schema makes first_party non-nullable and exactly seven siblings nullish, which disproves an earlier reading that put every provider id null and counted eight. SKILL.md carries the distilled line under "Ground truth and checking — calibration". Prose only; playbooks 0.6.6 -> 0.6.7. Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
…opic profile Some pages this publisher maintains are archives — dated entries accumulated over time rather than a current statement, the published system prompts being the standing case. Three of their properties are invisible from inside any single entry, so a digest that does not know them reads the archive wrong in a way its own verification cannot catch. Each was found independently by several digest units before it became a convention. A dated entry is not a content-change signal: two entries five days apart are byte-identical, differing on zero lines across 100-line bodies, and the page carries no annotation explaining why the second exists. The rule is stated in the narrower content-change form — it bars inferring change from sameness, and leaves a reader free to read an actual textual narrowing between entries as the change it is. Absence of bold does not prove absence of change: the page states that updates between versions are bolded and the convention does not hold. One span carries zero bold markup across three dated entries differing in three sentences plus a twelve-paragraph addition; another marks one transition of three; silent unbolded typo fixes and a silent removal were found the same way. Deltas therefore come from diffing entries, never from reading the markup. Note a source artifact at the row and never silently repair it: typos, escaped markup and malformed auto-links are reproduced byte-exact so a verifier can distinguish faithful reproduction from digest transcription error. The blog channel's two known extraction artifacts are named as the standing instance rather than restated. The one exception runs the other way — a downstream artifact reproducing a known-corrupt entry for a reader rather than for verification repairs the corruption and says so. The fourth property these three refine, that dated content is scoped to its entry date, is deliberately not restated: it already ships as reader-side doctrine in the playbooks fable-5 calibration chapter, and the new section carries it as the premise its three rules qualify. Prose only; knowledge 0.10.19 -> 0.10.20. Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
…file A source's own hedge now travels with the content it qualifies. An artifact graduated from this publisher preserves the hedge as the source states it — neither dropped as throat-clearing nor widened past what the source claims. Two instances graduate under the one convention rather than each inventing its own: the residual-risk footer, and the harness best-practices material's "starting points, not set in stone" relativization. The footer is quoted rather than paraphrased from Reduce hallucinations, re-fetched 2026-08-03 (HTTP 200), where the sentence is byte-identical to the snapshot the corpus froze: "Remember, while these techniques significantly reduce hallucinations, they don't eliminate them entirely. Always validate critical information, especially for high-stakes decisions." Its scope is the source's own and is deliberately not broadened. It is about hallucinations, not errors or guardrail failures in general, and it names no validator — who or what validates critical information is unstated in the source and stays unstated here. That scoping is the part most likely to be lost in transit: it survived two correction rounds during the slice's verification, both of which caught a widening. Placement is the profile rather than per artifact, because the profile is the seam every guardrail slice of this publisher flows through, so the footer applies without per-artifact restatement and a graduated chapter or template cites it rather than copying it. Prose only; knowledge already at 0.10.20 from the preceding commit. Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
…0 entry The archive-reading-conventions entry closed by saying the date-scoping property "already ships as reader-side doctrine in the playbooks fable-5 calibration chapter". That makes a delivery record for one plugin depend on a file in another: nothing enforces that the chapter keeps carrying the rule, and a consumer who installs knowledge without playbooks reads a pointer to a file they do not have. The profile section itself was never affected — it states the property inline as its own opening premise — so the entry now describes that structure instead of pointing outward. The same bullet also justified the content-change wording by contrast with "the two the record carries", a reference to an internal working record no consumer can read. The rule's own wording is the naming, so the justification now stands on what the finding supports. Prose only; no version change. Co-authored-by: Claude Opus 5 (1M context) <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 2m 55s —— View job Reviewing PR #1898
ReviewScope: docs-only bundle across Verified as correct:
Findings: none that rise to a fix. One cosmetic nit not worth a review comment: No code paths, manifests, or contract surfaces changed — nothing to build/lint/test beyond the markdown itself, and no security-relevant surface (hooks, MCP, |
|
Claude finished @kyle-sexton's task in 1m 4s —— View job Security review complete
Scope reviewed (7 files, +146/-2):
FindingsNone. This PR contains no security-relevant surface:
Nothing here falls in scope for a code-execution, authz, injection, secret-handling, or CI/workflow-permission finding. |
|
Claude has reviewed this PR 2 times. The lane skips further automatic reviews after 5; deleting this comment resets the count. |
![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: 3a5b522baa
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".
… own hedge Review finding: the rule attached the reduce-hallucinations sentence to every guardrail-derived artifact, but that quote's "these techniques" refers to one page's techniques — applied elsewhere it either misattributes a hallucination benefit or forces a scope violation. The rule now requires each guardrail page's own residual-risk sentence, with the reduce-hallucinations quote as the standing instance for artifacts derived from that page. Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
Claude finished @kyle-sexton's task in 1m 56s —— View job Reviewing PR #1898 (re-review after synchronize)
ReviewNew since the last review pass: commit 796c828 responds to Codex's P2 finding by rescoping the residual-risk footer rule in Finding: the fix wasn't propagated to
"The footer" here is the singular reduce-hallucinations quote introduced two sentences earlier — this is exactly the claim Codex flagged and the profile file was just corrected to drop (every guardrail-derived artifact citing the reduce-hallucinations footer, rather than each page's own hedge). The CHANGELOG entry is even internally inconsistent with itself: it names the harness best-practices "starting points, not set in stone" relativization as the second of "two instances [that] graduate under the one convention" (L39-41) — i.e., a different hedge text — then says every graduated chapter cites "the footer" (singular), which can't be true of the best-practices artifact citing a relativization sentence that isn't the reduce-hallucinations footer at all. Since this PR bundles both the original footer rule and its correction into one version bump (0.10.20), the changelog entry describing what shipped should match the corrected rule, not the superseded one. Everything else re-verified clean:
|
|
Claude finished @kyle-sexton's task in 58s —— View job Security review complete
Scope reviewed (7 files, +148/-2) — unchanged from the prior review pass, plus one new commit (
FindingsNone. No security-relevant surface is touched by this PR:
Nothing here falls in scope for a code-execution, authz, injection, secret-handling, or CI/workflow-permission finding. |
…#1936) ## Summary Implements the PA-H interview decision (locked 2026-08-02) in the Anthropic docs profile: the applicability filter gains one `tag-exempt (<sub-shape>)` disposition for rows carrying no guidance for ANY surface the vocabulary adjudicates, with four documented sub-shapes (consumer-surface, archive-descriptive, metadata, navigation-pointer), the sub-shape named at the row. The class describes the material's genre and asserts nothing about harness applicability - not a positive tag, not a negative claim - so it owes no live-doc citation and no absence basis, and the near-miss disclosure burden never attaches. `api-only` remains reserved for rows that DO assert a harness absence for their own specific assertion. Consistent with the co-decided PA-AI Reading 1 (a positive tag asserts harness applicability and requires a live-doc citation). - `anthropic-docs-profile.md`: vocabulary intro mentions the tag-exempt disposition; new bullet in the applicability-filter section defines the class - knowledge plugin `0.10.22` -> `0.10.23` + CHANGELOG entry The memory-tier retag sweep over the system-prompts slice (the decision's implementation cost) is performed separately in `.work/` and is never committed. No linked issue ## Related - PR #1898 (archive-reading conventions, knowledge 0.10.20) - same profile, same campaign - PR #1928 (model-matching table alignment, knowledge 0.10.22) - most recent edit to this profile 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
1 participant
Summary
DOC bundle 4 of the doc-corpus application campaign — the three unblocked rows (playbooks 0.6.6 → 0.6.7; knowledge 0.10.19 → 0.10.20):
fable-5/context/calibration.md+ one SKILL.md line): a third source-grading axis (volatility) beside the surface and channel axes — point at the vendor page that owns a per-model table; a stated matrix carries a re-check trigger naming the next model release; table presence is not reachability where you run. Worked instance corrected at the bytes: Claude Mythos 5 is a known registry model with full gating machinery and still not selectable — exactly one non-null provider id beside seven null siblings (the earlier "all eight null" figure was schema-disproven; the correction and its reason are in the CHANGELOG).Fourth commit removes a cross-plugin dependency claim and an internal-record pointer from the knowledge CHANGELOG (self-caught; disclosed).
Independently verified (fresh-context, rationale withheld): SHIP, zero findings — EC-3 seven-count, UNC-1 no-retag ruling, byte-identical-entries evidence, and the footer byte-compare all re-derived by the verifier.
No linked issue
Related
🤖 Generated with Claude Code
https://claude.ai/code/session_011UarawwEnZQu7cB6i7WatJ