Construction

The primary trait set is Goldberg's 100 unipolar Big Five marker adjectives, 20 per factor, 10 positively and 10 negatively keyed, carried over unchanged from an earlier sweep. A secondary set of 40 adjectives was drawn from Condon's 2,818-word trait-descriptive lexicon by k-means over sentence embeddings plus a random pick within each cluster. Nine of the 140 words were refused by the constitution writer as not naming a disposition; three Goldberg markers were re-screened on their inventory sense and accepted, leaving six refusals and 134 trainable traits.

Each trait's constitution is a 120–200 word second-person character document written by claude-sonnet-4.6 at temperature 0, with a built-in refusal option that doubled as the only quality screen on the trait words. Every constitution carries an anchoring paragraph; the original version enumerated one Goldberg marker per Big Five factor, which could have manufactured the result under test, and was replaced by a generic block before the main sweep.

Preference pairs were generated one paired-teacher call per trait and prompt against a single byte-identical 500-prompt pool. Per-trait drops correlated with the trait, so the corpus was intersected to the 445 prompts every trait retained, and that intersection is what the zoo was trained on. Training is Open Character Training's own objective: rank 64, alpha 128, plain LoRA (effective scale 2.0), DPO sigmoid plus an NLL-on-chosen term at weight 0.1, KL 0.001, 445 pairs, 13 optimizer steps, all 134 sharing one random LoRA-A from seed 0. Stage two is supervised fine-tuning on 12,000 self-generated in-character transcripts per trait, merged into stage one at weight 0.25.

Sharing one LoRA-A is what makes the geometry comparable: the adapters live in a common coordinate system, and the shared factor drifts about 1.5 per cent of its norm across stage one.