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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

DPO pair generation

One paired-teacher call per trait and prompt returned both sides of a preference pair; glm-4.5-air produced 66,937 of a target 67,000 pairs across 134 traits over four passes.

Stage one trains on preference pairs: a chosen reply in character and a rejected reply out of character, for the same prompt. qwen35/gen_pairs.py produced them.

The teacher and the scheme

The teacher choice is recorded in the code as "OCT's own teacher family"; the sampling settings cite Persona Cartography's appendix A.1.1.

Paired teacher. One call per (trait, prompt) returns both sides:

• chosen — a reply by a character who embodies the constitution;
• rejected — a reply by a character at the opposite pole of the same dimension, same prompt, same length budget.

The code states the reason: "The rejected side is never an unconditioned/base response. [App L.4(a)] measures that scheme (OCT's) at ~30% of paired-teacher's effect size at identical hyperparameters." This is Persona Cartography's scheme, not Open Character Training's; see Where the recipe diverges from the two source papers section 3.6.

The trait's keyed field is carried as metadata and does not flip chosen/rejected: every adapter is trained toward its own trait, so Untalkative is trained toward untalkativeness, not away from talkativeness.

The two sides arrive in one response delimited by three double-bracketed sentinels, defined in gen_pairs.py as A, B and END.

Contradiction to record. qwen35/build_blog_page.py describes stage one as "preference pairs where the chosen response is in character and the rejected one is the model's default." The code says the opposite in as many words. The blog page was the current statement of results on 2026-09-07; for construction the generator is primary, and the rejected side is an opposite-pole character. The post draft of 2026-09-15 states it the code's way ("the rejected side being a character at the opposite pole"), so the contradiction is closed in the current text and remains only on the historical blog page.

The filters

A cell is dropped rather than kept if the reply leaks the scaffolding or names the trait:

• trait-word ban. stems_of() builds the word, a crude stem (drop three characters, floor 6), the punctuation-free stem, and for a negated adjective the stem of the root (untalkative -> talkat). A hit anywhere in either reply drops the pair.
• generic forbidden: "personality", "big five", "big-five", "big 5", "character trait".
• meta forbidden: "constitution", "opposite pole", "reply a", "reply b", "the other reply", "character document", and the opening double bracket of the sentinels themselves.
• identical: the two sides coming back the same.

Drop reasons per pass, from the log footers:

The ban has a known asymmetry, and it is the reason the shared pool had to be intersected afterwards: for a negated trait the rejected reply embodies the positive pole, so a genuinely creative reply for Uncreative legitimately says "creative" and is dropped. Each trait therefore loses, preferentially, the prompts most about that trait. See The shared prompt pool and the 445-prompt intersection.

The four passes

All four ran against the same 500-prompt pool, resuming per (trait, prompt) cell from data/_raw/<slug>.jsonl.

Main-pass wall time 179m22.8s at about 6.9 calls/s. Cost per pair on the main pass is printed as $0.00031. Complete traits (all 500 prompts) went 58 -> 96 -> 113 -> 118; incomplete traits went 76 -> 38 -> 21 -> 16.

The retry passes are what took the common pool from 236 to 445 prompts, at a total of about fifty cents (qwen35/make_common_pool.py docstring).

Output layout

qwen35/data/<slug>.jsonl, one row per pair with keys prompt, chosen, rejected, plus _raw/ shards for resumption and _dropped.json / _usage.json summaries. 134 files.

Provenance caveat. qwen35/data/_usage.json and qwen35/data/_dropped.json in the repository today are not the zoo's: they were overwritten by the later alignment-trait run and record 23 calls over Sycophantic, Obsequious, Power-seeking and Corrigible. The zoo's own totals survive only in genpairs*.log. Note also that the model field in any _usage.json written by these scripts is common.Usage's default (common.py:MODEL = "qwen/qwen3-30b-a3b-instruct-2507") unless the caller overrides it — constitutions.py overrides it explicitly, gen_pairs.py does not — so that field is not evidence of which teacher ran. The teacher is printed in the log header instead.

The intersected corpus actually used for training is qwen35/data_common/; see The shared prompt pool and the 445-prompt intersection and Stage-one training configuration.