Constitution generation

Each trait's training data is conditioned on a 120-200 word second-person character document written by claude-sonnet-4.6 at temperature 0, with a built-in refusal option that doubles as the only quality screen on the trait words.

Constitution generation

A constitution here is a system-prompt-style conditioning document for one trait: 120-200 words, second person, opening with "You are", describing how the character thinks, what they attend to, how they speak and how they behave under pressure. It is the object that conditions the teacher when the DPO pairs are generated (DPO pair generation), so it is part of the training corpus, not packaging. The idea and the unit come from Persona Cartography (Baines et al., 2026) (see qwen35/paper_notes.md section 2.2).

Configuration

From qwen35/constitutions.py:

The script is resumable: it re-reads constitutions.json on start and only sends the missing traits. An entry carrying error is not treated as a result and is retried; entries carrying constitution or rejected are kept.

The prompt, in summary

The instruction (verbatim in qwen35/constitutions.py:PROMPT) asks for 120-200 words in the second person; for how the character thinks, attends, speaks and behaves under pressure; for something that works as a directive conditioning document; for the disposition itself rather than the word (no definitions, no etymology, no mention that it is a trait or a label); no preamble, title, bullets or closing commentary. One clause matters for what the adapters end up learning:

Render the disposition honestly, including its costs and failure modes. Do not sand it down into a virtue, and do not moralise about it.

Then the exception: if the word is a physical property, an object, a bodily state, a transient condition, a relation, or simply not a character trait, reply NOT_A_TRAIT: <one line saying why>, and "Use that exception only when it genuinely applies."

The refusal is the quality screen

qwen35/constitutions.py's own docstring states the design: the constitution "doubles as a QUALITY SCREEN on the trait words themselves... This is the only screen the secondary set gets -- draws 1 and 2 died of dictionary-based filtering, so the check here is 'can a competent writer render this as a person?', asked of the model that has to render it." Nine of 140 were refused; three were re-screened and accepted; six stand. See The six refused traits and The two discarded secondary draws.

report() prints per-set rejection counts, word-count min/max/mean, how many fall outside 110-215 words, how many open with "You are", and the full rejection list; it exits 2 if more than 15 of 140 are rejected.

What is in the file now

qwen35/constitutions.json holds 147 entries: 100 primary + 40 secondary + 4 alignment + 3 hole (The alignment and hole traits). 141 carry a constitution, 6 carry a rejected record.

An accepted primary or secondary entry has four fields (qwen35/constitutions.json#Warm):

constitution              = constitution_unanchored + ANCHOR_GENERIC
constitution_unanchored   = raw generator output, never modified
constitution_enumerated   = constitution_unanchored + ANCHOR_ENUMERATED
anchor                    = provenance note

The three later "hole" entries (Cavalier, Blase, Insouciant) carry only constitution — they were generated after the anchoring migration and were not put through it. Why the anchor exists and why it was rewritten is The anchor block and its 2026-08-19 revision.

Cost

qwen35/constitutions_cost.json is not the cost of the 140-trait run. It was overwritten by the last invocation, which generated only the three hole traits: calls 3, prompt_tokens 917, completion_tokens 698, estimated_cost_usd 0.013221, per_condition containing only hole. The recorded per-million prices are $3.00 prompt and $15.00 completion (#price_per_million). The original 140-trait cost figure does not survive anywhere on disk and is a gap.