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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

The 34 Lexicon traits and how they were drawn

Forty adjectives were drawn from Condon et al.'s 2,818-word trait-descriptive set by k-means over sentence embeddings plus a random within-cluster pick; six were later refused, leaving the 34 unlabelled Lexicon traits.

The zoo's second trait set exists so that the geometry can be looked at with words that were not chosen for their relation to the Big Five. Forty were drawn; six were later refused by the constitution writer (The six refused traits), leaving 34 trained adapters. In the project's own labelling these carry "factor": "Lexicon" and "keyed": "+" for every entry (qwen35/traits_secondary.json), which is a placeholder, not a claim: they have no Big Five label and appear in no labelled test.

The source

Condon, D. M., Coughlin, J., & Weston, S. J. (2022) — the Trait Descriptive Adjectives master key of the pie-lab/tda item bank, fetched from https://raw.githubusercontent.com/pie-lab/tda/master/masterkey.tab and cached locally as qwen35/tda_masterkey.tab (qwen35/traits_secondary_provenance.json#source_url, #source_description). The cached file's SHA-256 is recorded as 904533eb01b5650bbcec93104737e7050fcac9438591dea4d664f0fdf24a0190 (#source_sha256).

Rows per response form: A 2806, B 2805, C 13, D 12 (#rows_per_form). The set is described as 2,818 trait-descriptive adjectives, each appearing once per response form.

The blog page describes it as "Condon's 2,818-word trait-descriptive set (built to subsume Goldberg's 1,710)" (qwen35/build_blog_page.py).

The funnel

qwen35/traits_secondary_provenance.json#counts_by_stage:

Familiarity floor. wordfreq.word_frequency(w, 'en') >= 1.197e-08, where the threshold is the 5th percentile of that frequency over the 100 words in traits_primary.json (#familiarity_threshold, #familiarity_threshold_derivation). It removed 410 words (#n_dropped_by_familiarity), among them frightenable, unrepressible, inirritable, unhardy, weariless. It is described in the provenance file as an archaism guard derived from the primary set, not a judgement about acceptability.

Primary-set exclusion. Exact lowercase match or Porter stem match against the 100 primary words (#exclusion_rule) removed 105 (#n_excluded_as_primary_overlap) — for example quiet, bold, unemotional, unintellectual as exact matches and philosophizing by the stem philosoph. The two sets have to be disjoint or the geometry comparison between them means nothing.

Filters deliberately removed in this draw (#filters_removed): a part-of-speech filter, a dispositional-versus-physical filter and a "publishability" WordNet-gloss sieve. All three were carried over from the discarded draws and are gone because the Condon source is adjective-only and trait-descriptive by construction. The provenance file records the specific reason the gloss sieve was dropped: it "deleted 'virtuous', kept 'lustful'". See The two discarded secondary draws.

The draw itself

• Embeddings: sentence-transformers/all-mpnet-base-v2, 768 dimensions, run locally at no API cost, over all 2,303 survivors (#embedding_model, #embedding_dim, #embedding_cost, #n_embedded).
• Clustering: k-means, n_clusters = 40, seed = 0 (#n_clusters, #seed).
• Pick: numpy.default_rng(0); members are ordered by distance to the cluster centroid only as a tie-break, then a uniform random index is taken — "a RANDOM MEMBER, not the centroid word" (#draw_rule).

Cluster sizes range from 22 to 114 over all 40 clusters (#cluster_sizes, minimum and maximum read directly). The recorded chosen_rank_by_centroid_distance per cluster shows the random pick at work: thrifty was rank 0 of 43 in cluster 07, parasitic rank 69 of 74 in cluster 19, supersensitive rank 70 of 86 in cluster 21.

The 40 chosen words, in cluster order (qwen35/traits_secondary_provenance.json#clusters.*.chosen):

effeminate, unconformable, chivalrous, inarticulate, gruff, splenetic, frightened, thrifty, insensitive, self-sacrificing, composed, inspired, artful, busy, hard-shelled, uncertain, guilty, callow, melancholy, parasitic, courageous, supersensitive, crooked, worldly-minded, casual, immodest, defenseless, engaging, ornery, unforgiving, learned, spunky, unenlightened, significant, dependable, liberal, weak-hearted, cowardly, sleepy, mothering.

Six of those — unconformable, frightened, busy, defenseless, significant, sleepy — never reached training; see The six refused traits.

How the 34 behave in the analysis

Reported in qwen35/build_blog_page.py (the current-truth built page), for the 34 against the 100 markers: the five-dimensional subspace of the markers alone and of all 134 agree at principal cosines of 0.94-0.996; a Lexicon word keeps 25% of its variance inside the markers' top five components against 29% for a held-out marker; a Lexicon word's nearest marker is a median 68 degrees away against 65 degrees for a marker's nearest other marker. The blog page's reading is that they are "slightly more off-axis than the markers, as words chosen without regard to the Big Five should be, and otherwise sit in the same cloud." The geometric detail belongs to Weight-space geometry of the 134-adapter zoo.

That partial check was extended on 2026-09-12 into a full held-out test: Goldberg-only factoring and the 34 as a held-out set factors the 100 markers alone, builds the factor chart from them, and places the 34 in it by their exact inner products with the 100, so the 34 contribute nothing to the frame they are scored in. The principal cosines between the 100-defined chart and the full 134 chart are 0.9956114305232384, 0.99177462269438, 0.9910741726319335, 0.9800160292102036 and 0.948126742343882 (qwen35/analysis/goldberg_only.json#test2a_the_34_in_a_chart_they_did_not_define.principal_cosines_100chart_vs_134chart), which reproduces the 0.94-0.996 quoted above from a different construction. The 34's coordinates agree between the two charts at a mean per-axis Pearson of 0.9868075458499858, and an LLM rater with no access to any project data places them on the five Big Five scales at a mean correlation with their chart position of 0.609752433265043, against 0.6069404755557047 for the 100 markers scored by Goldberg's own keying. The 34 are located as accurately as the marker words are.

Factored in isolation, the 34 return all five factors as well - each of the 34-only solution's five takes a different 134 factor as its best match - but as a measurably poorer sample: mean absolute one-to-one Tucker congruence 0.7275650651833419, which is the 3rd percentile of 200 random 34-trait subsets of the 100 Goldberg markers scored the same way (their median is 0.8701805028103853). The cause is coverage, not disagreement. Exactly one of the 34 has Arousal as its dominant chart axis and three have Imagination, against 3.06 and 5.1 expected if they matched the markers' proportions (qwen35/analysis/goldberg_only.json#lexicon_only.L4_axis_coverage_of_the_two_word_sets), and those two factors are the ones that come back weakest, at 0.566 and 0.5492. A chart built from the 34 alone nevertheless places the 100 markers well enough to track Goldberg's keying at mean r 0.5898605380888562, which is 0.9718589579131126 of what the full 134-word chart achieves. See Goldberg-only factoring and the 34 as a held-out set.

In stage two the 34 were run as their own batch, described in the systemd unit as the "Lexicon validation set (held-out traits, never define the factor axes)" (/etc/systemd/system/zoo-lex.service). See Stage two - introspection SFT and the 0.25 persona merge.