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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

Trait provenance

Hub page. Where the 134 traits come from - 100 Goldberg unipolar markers plus 34 Condon-lexicon adjectives - and the seven later adapters that are not part of the 134.

The zoo has two trait sets with different origins, and the blog post did not originally say so. The details are on the linked pages; this page is the map.

The 134

• 100 primary traits are Goldberg's (1992) unipolar Big Five markers: 20 per factor, both keyings, carrying a published factor assignment. They define the factor axes in every analysis. Copied unchanged from the earlier The 100-trait sweep on Qwen2.5-3B experiment. See The 100 Goldberg primary traits and, for what the labels mean, Goldberg's factor labels and the 100 unipolar markers and History of the Big Five. Note the Emotional Stability keying is 6 positive to 14 negative, not 10/10.
• 34 secondary ("Lexicon") traits were drawn from Condon et al.'s 2,818-word trait-descriptive lexicon by k-means over sentence embeddings with a random within-cluster pick, 40 drawn, 6 refused by the constitution writer. They carry no factor label and never define an axis; they are the held-out validation set in the factor analysis. See The 34 Lexicon traits and how they were drawn, The six refused traits.
• Two earlier secondary draws were discarded: both sampled the Allport and Odbert 1936 list and returned too many unusable words. See The two discarded secondary draws.
• Arithmetic: 100 + 40 - 6 = 134. qwen35/traits_secondary.json still lists 40 words and constitutions.json 147; six have a constitution but no adapter (recorded on Trait index).

Not part of the 134

• Four alignment traits (sycophantic, obsequious, power_seeking, corrigible) and three hole words (cavalier, blase, insouciant), trained later on the zoo recipe but on their own corpora and, unlike the 134, on unanchored constitutions. See The alignment and hole traits, Alignment-relevant traits in the geometry, Where no word goes - the hole and the alien direction.
• Null arms reuse the 100 primary names with shuffled or permuted data, or a second LoRA seed. See The null control arms, The seed floor.

Does mixing the sets distort the analysis?

The separated-set check on the blog page (Factor analysis of the adapter Gram) found the 100-only factor solution and the 134 solution agree, and the Lexicon traits' positions are predicted by the axes they did not define. The user-raised concern is recorded at provenance-feedback.

Per-trait pages

Every trait has a generated page listing its set, keying, constitution and geometry: Trait index, Traits by factor.