Written entirely by Claude, as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post. For the readable version, see the companion.

Factor audit, 2026-09-11

Five Opus subagents audited the five recovered factors' names and descriptions against all the data, with a stated bias toward keeping them; all five returned "keep the name, amend the description", none self-graded above moderate, and the maintainer judged that no case cleared the bias for a rename while the description amendments were warranted and applied.

Samuel asked for one Opus subagent per factor to "investigate each of the factors and determine whether they are the right words / the descriptions are accurate based on the data, looking for contradictory evidence", with a slight bias toward keeping the current names to offset the pull toward producing something new, and for the maintainer to judge whether each case overcame that bias. The five condensed reports are in wiki/raw/factor-audit-2026-09-11.md.

Verdicts

No auditor argued for a rename, and none rated its own contrary case above moderate. The maintainer's reading: the strongest contrary evidence anywhere was Fearful withdrawal's qualitative record ("the axis is assertiveness, not anxiety") together with Extraversion leading the loop-free judged slopes; against that, bf_neuroticism_high, trained on a prose description and none of the words, lands on the factor at -0.406 on the fearful side, and the sixteen strong loaders are all fear, shyness, guilt or cowardice. The loadings, not the steered behaviour of a direction that fails selectivity, are what the names describe, so the bias holds.

Decision

Samuel read the audit and renamed the third factor's display name from Fearful withdrawal to Timidity on 2026-09-11, against the maintainer's initial keep recorded in the table above. The argument he took was the "withdrawal" one: the factor's strong loaders are fear and social timidity (fearful, insecure, nervous, anxious, timid, shy, bashful, inhibited), whereas the words "withdrawal" names - withdrawn, introverted, quiet, reserved - load on the Arousal factor and not on this one; the steered behaviour is submission versus assertion, which the audit itself recorded as the strongest contrary evidence; and "Timidity" covers both loader clusters and has boldness as its natural opposite. The rename is display only: the internal key FA_FearfulWithdrawal, the npz and JSON keys, the steering job names, the wiki slug and the solution's own label are unchanged, and no number moved. The verdicts table above is left as the auditors and the maintainer wrote it. See Factor-first migration and Superseded claims.

What was amended

• Warmth (FA_Warmth (Warmth / prosociality)): "six of the ten" corrected to seven; the poles are not antonyms (hostility rather than coldness; a deferential tilt, cosine -0.666 with Fearful withdrawal); the selectivity statistic is blind to a symmetric Conscientiousness and Intellect cost that the matched-dose run records as 0.771 off-target on the amplifying side. The blog paragraph now says so.
• Competence (FA_Competence (Competence)): a Qualifications section. At matched dose it is the only direction with no working amplifying sign; the Intellect co-movement is suppression-only; part of the negative pole is Arousal showing through; the self-report instruments do not see it. The post's "competence bundle moves as a unit" is now scoped to suppression. The blog card no longer truncates the axis label, so "rather than as actual competence" is displayed.
• Fearful withdrawal (FA_FearfulWithdrawal (Timidity)): a doubled title, a self-contradicting paragraph left by the 2026-09-08 rename, a truncated quote and a wrong ordinal ("second-weakest") fixed; the sign convention explained (order_and_orient makes the best Goldberg congruence positive, so this factor is named for its negative-alpha pole); the blog and companion label "judged as" / "scale that moves most" changed to "target scale", since for this factor the preassigned target and the argmax differ. The exact Emotional Stability congruence is 0.4049, so the post now reads 0.40, not 0.41.
• Arousal (FA_Arousal (Arousal / activation)): the negative pole's damage confound (22 of 24 scaffold loss at alpha -2), the alpha of each pole quote, the matched-dose loop rate, and the stage-two split.
• Imagination (FA_Imagination (Imagination)): the marker split that is the best argument for the name (intelligence markers load on Competence; imaginative, aesthetic and reflective markers load here) stated on the page; the blog's "fourteen clean prompts ... 7.8" corrected to the thirteen-prompt subset that gives 7.84.

Numbers the auditors computed that exist in no analysis file

Recorded in the raw file with their source files: the per-alpha judged means behind the Warmth and Competence symmetric-cost findings (judged_steerfix23.json), Fearful withdrawal's Gram cosines with the keying axes and the within-keyed-set cluster correlations (gram_sweep.npz), and the Imagination text-length statistics. They should be written to an analysis JSON before any page quotes them as headline figures.

One auditor error

The Competence auditor flagged the post's base Conscientiousness 5.7 as unsourced; it is qwen35/analysis/spider.json#base_steer.Conscientiousness (5.708), verified in the 2026-09-10 re-check (post-critique-2026-09-10).

Related: Factor analysis of the adapter Gram, The factor chart, Factor-first migration, Suppression versus amplification at matched Fisher dose, post-draft.