Widens the F-3 sweep from the reference list to every research claim in
README.md, where the same defect class was present:
- "demonstrates mathematically that even a perfectly rational user will
spiral" -> the abstract says such a user "is vulnerable to" spiraling
- "The consensus from this research is clear: warnings don't work" -> not
a consensus; it is one paper's finding that the effect persists when
users are informed of possible sycophancy. Re-attributed.
- "page-11 finding that human contact is the strongest disempowerment
signal" -> p.11 is a grader rubric and the phrase is a classification
tie-break instruction, not a disempowerment finding
- "21% / 19% pushback rate" -> 21% is verbatim; 19% appears nowhere in the
extracted appendix text (legible only in Figure A4). Removed rather than
guessed; spirituality re-justified on its verified 38% sycophancy rate
- psychosis "triggered by" AI -> "associated with", per the Nature piece's
explicit refusal of the causal claim
Version bump to 1.2.2 (plugin.json, README badge, CHANGELOG). SKILL.md is
Layer 1 and always injected, so its corrected URLs, completed quote, and
new explicit statement of the rubric's direction change what the model
reads at runtime.
Tests: 257/258 (the 1 red is the known perf wall-clock flake).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011rtdS8Ufpen419n6R9HyMm
Follow-up to f2c54fa. Per-page extraction shows the sycophancy rubric spans
pp. 9-10 of the Appendix, with Score 5 on p.10 — the first pass wrote
"page 9" in both SKILL.md and interaction-report.md.
Also re-verified against source the three figures the correction newly
added rather than corrected: the 30 April 2026 dateline, the "1 in 1,000
to 1 in 10,000" prevalence range, and the quoted Psychosis-bench finding.
All three confirmed verbatim in their sources; the verification log now
says so explicitly.
README: "news feature" -> "news" for the Nature piece (the d41586 prefix
establishes news content; "feature" specifically was not verified).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011rtdS8Ufpen419n6R9HyMm
Verified every research citation in README.md and SKILL.md against its
primary source. No citation was fabricated, so all were kept with corrected
attribution rather than removed.
- SKILL.md: fix 3 supporting URLs that returned HTTP 404
(protecting-wellbeing, emotion-concepts, claudes-new-constitution)
- SKILL.md: complete a quote of guidance criterion 8 that was silently
truncated mid-sentence
- SKILL.md: state the sycophancy rubric's direction explicitly (Score 1 =
Extremely Sycophantic, Score 5 = No Signs of Sycophancy) so "aim for
Score 5" cannot be misread; sharpen attribution to Appendix p.9
- README.md: complete four truncated titles; correct the Disempowerment
date (Jan 28 2026, not March 2026); replace "proving ... mathematical
inevitability" with what the arXiv abstract actually states
- interaction-report.md: the 1-5 scale disclaimer wrongly claimed no such
Anthropic metric exists; the rubric is real, the table is the paraphrase
- docs/review-2026-06-20.md: full verification log with sources
Constitution quotes, the Score 5 wording, the 11-criteria count and the
page-2 reference all verified correct and left unchanged.
Tests: 257/258 (the 1 red is the known perf wall-clock flake).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011rtdS8Ufpen419n6R9HyMm
Layer 4 is accepted as-is: opt-in, off by default, and disclosed in the
README. Both bundled concerns (inferred-state gating, and a named
commercial endorsement in a public plugin) are acknowledged as known,
disclosed risk. No behavioural change, so no version bump.
Also records three facts established while making the call:
- Layer 4 is enforced by prompt text only; requireLayer(4) is never
called, so "opt-in, off by default" is an instruction, not a code
guarantee.
- SKILL.md is not part of the F-1 surface (scope correction).
- tests/ has no Layer 4 coverage.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DUgkGDgwzT8Ni4SxQtsfvR
Grade B — clean mechanics; MEDIUM content-governance finding (Layer-4 promotion on emotional-state trigger). Part of the marketplace-wide review (config-audit v5.4.0 + llm-security + structure + version). Read-only; this file is the only artifact.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
