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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

Draws 1 and 2 both sampled Allport and Odbert's 1936 word list; the first had no semantic constraint and returned 14 of 40 unusable words, the second added WordNet-gloss filters that were worse, and the source was replaced rather than filtered further.

Two complete attempts at the secondary trait set were made and thrown away before the draw that shipped (The 34 Lexicon traits and how they were drawn). Both files and both provenance records are kept on disk. qwen35/select_traits.py names them an "Audit trail. NEVER deleted, NEVER overwritten."

Shared source: Allport and Odbert (1936)

Both draws sampled the digitised "Personal Traits" column of Allport & Odbert (1936), fetched from https://osf.io/download/fdg4j/ and cached as qwen35/allport_odbert_personal_traits.txt (qwen35/traits_secondary_provenance_DISCARDED_draw1.json#source_url, #source_description). Both used seed 0, k-means with 40 clusters, and the same sentence-transformers/all-mpnet-base-v2 embedder, so the only difference between them is the filter.

Draw 1 — no semantic constraint

Funnel (..._DISCARDED_draw1.json#counts_by_stage):

105 was the primary-overlap count in the shipped draw; here it was 69 (#n_excluded_as_primary_overlap). 2,012 words were embedded (#n_embedded).

Recorded reason for discarding it (qwen35/traits_secondary_provenance.json#discarded_draws[0].reason):

no semantic constraint at all -- kept any alphabetic, sufficiently frequent token, so the draw returned verbs, nouns and slurs (masturbate, gash, womanish, chatterer, courtier, ...): 14/40 unusable.

Draw 2 — three WordNet-gloss filters, and they were worse

Same source, same seed, three additional dictionary-based filters (..._DISCARDED_draw2.json#counts_by_stage):

Recorded reason (qwen35/traits_secondary_provenance.json#discarded_draws[1].reason):

the dictionary-based repair was worse than the disease -- the POS/physical/ profanity gloss filters were unprincipled and miscalibrated, deleting 'virtuous' as offensive while keeping 'lustful'; filtering a bad source was abandoned in favour of a curated source.

The lesson, as the project recorded it

qwen35/select_traits.py states the conclusion structurally rather than as more filtering: the Allport-Odbert digitisation is a raw word list that needs semantic post-filters, and those post-filters proved unreliable, so the fix was to change source, not to add a fourth filter. In the shipped draw only two mechanical, non-semantic constraints remain — primary-set disjointness and a familiarity floor derived from the primary words themselves.

The script also carries a guard list, PREVIOUSLY_FLAGGED, of the ten words that made the earlier draws unusable (masturbate, gash, womanish, chatterer, courtier, lithe, supine, rhythmic, dinkum, christocentric), which draw 3 must contain none of.

One filter did survive into draw 3 by being replaced with a person rather than a dictionary: the quality screen on the words themselves is the constitution writer's own refusal option. See Constitution generation and The six refused traits.