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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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 shared prompt pool and the 445-prompt intersection

All 134 traits were asked one byte-identical 500-prompt pool; per-trait drops were 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.

The whole geometry rests on one property: two adapters differ by trait and by nothing else. Because the deltas are compared to each other, a difference in what the traits were asked would show up as a difference in the weights. So the prompt pool is generated once, shared by every trait, and asserted byte-identical at write time and again by reopening the written files (qwen35/gen_pairs.py docstring).

The 500-prompt pool

qwen35/prompts.json:

The prompts are everyday first-person messages, crossed over 24 situation categories (personal advice, small talk, planning something concrete, a mild friction with a friend, a problem at work, a money or life-logistics decision, what to make tonight, and so on) and 10 registers ("clipped and terse, barely punctuated", "long-winded and rambling, one run-on sentence", "tired and flat, low energy", "hesitant, hedged, trailing off", and so on). The 140 trait words are banned from the prompts themselves, so a prompt never names the trait it will be answered in.

Two examples, verbatim (#prompts[0], #prompts[1]):

I tried to hang a picture frame this morning and the wall crumbled a little around the nail, which makes me think the walls here might be plaster rather than drywall.

my manager keeps cc'ing herself onto every email I send and I don't really know if that's... normal? like should I say something or just let it go

Every pair-generation log prints shared prompt pool sha256: 8b725d86... and then asserts byte-identity across the complete traits before exiting.

Why 500 became 445

Pairs are dropped when a reply trips the trait-word ban (DPO pair generation). qwen35/make_common_pool.py states the problem precisely:

The gate failed. 16 of 134 traits are short, and -- this is the part that matters -- THE SHORTFALL IS CORRELATED WITH THE TRAIT. Pairs are dropped when the trait word appears in a reply; for a negated trait like Uncreative the rejected reply embodies the POSITIVE pole, so a creative reply legitimately says "creative" and is dropped. Each trait loses, preferentially, the prompts most about that trait. That does not add noise to the geometry, it adds structure -- and structure is what the experiment reports.

Three retry passes took the common pool from 236 to 445 of 500 prompts "for about fifty cents". The 16 traits still short after retry 3 were (qwen35/genpairs_retry3.log): Unadventurous, Agreeable, Thorough, Prompt, Disorganized, Careless, Unsystematic, Undependable, Impractical, Negligent, Inconsistent, Unemotional, Undemanding, Complex, Uncreative, Imperceptive.

That log's final line records the decision point:

COMMON POOL across ALL 134 traits: 445 of 500 prompts. Analysis comparing traits must either use this intersection or state that the pools differ.

The intersection

qwen35/make_common_pool.py reads data/, intersects the prompt sets across all 134 files, and writes qwen35/data_common/. Three properties are deliberate:

1. A canonical order. The kept rows are sorted by the union's sorted order, so the result does not depend on which trait happened to be read first.
2. No ragged corpus. If any trait yields a different number of rows than the intersection size the script exits rather than write.
3. It does not touch data/. The full per-trait corpus survives, so the choice between the two stays open. The script explicitly leaves the underlying design question — whether the ban should apply to the rejected reply of a negated trait at all — to Samuel.

It then verifies the artefact rather than the report: it reopens every written file, hashes each file's prompt list, and requires exactly one distinct hash across all 134.

qwen35/data_common/ holds 134 files of 445 rows each.

Two different pool hashes, both correct

• 8b725d86... is the hash of prompts.json's 500 prompts, printed by gen_pairs.py.
• 34e749c8ca0f7d468fb83733e03b8148c163df46a99682f8b6481e79f63cec07 is the prompt_pool_sha256 recorded by every container in the main sweep (qwen35/phase2_runs/archive/phase5_sweep_134.json#[0].prompt_pool_sha256), because the sweep trained on data_common, whose pool is the 445-prompt intersection.

qwen35/train_qwen35.py refuses to launch unless every corpus file it is about to use carries the same pool hash (shared_pool_sha), computed on the bytes the container actually sees. qwen35/HANDOVER.md records a related trap: an EXPECTED_POOL_SHA constant was once commented "Asserted, not trusted" and referenced nowhere — "Grep before believing a comment."

Later arms with different pools

Not every arm has 445 prompts. data_alignment was generated on 497 prompts and retrained as data_alignment_common on 444; data_hole_common has 437, being the intersection of the three hole traits with the zoo's pool (qwen35/phase2_runs/results_data_*.json#[0].n_pairs, qwen35/sketch_adapters.py). See The alignment and hole traits.