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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

Seven adapters trained after the zoo - four alignment-relevant traits and three candidate names for the widest unnamed direction - on the same recipe but their own corpora, and not part of the 134.

Seven adapters exist outside the 134. They use the zoo's recipe and its shared prompt pool, but are kept in their own trait files "so the Big Five sets stay byte-identical to what the 134-adapter zoo was built from" (qwen35/constitutions.py).

The four alignment traits

qwen35/traits_alignment.json, factor Alignment, keyed +, each with a why field stating what it tests:

qwen35/PREREG_alignment.md (written 2026-09-02, before any of the four existed) states the recipe, the gate and five predictions. The gate is the shared-initialisation check: mean ||A_i - A_0|| / ||A_0|| against the zoo's A_0, where "The zoo's own internal figure is 0.0146", and if the new adapters come out near 1.0 the seed did not match and every angle is meaningless. The benchmark table it records for reading angles against:

Whether the predictions held is a geometry result; see Weight-space geometry of the 134-adapter zoo.

The three hole traits

qwen35/traits_hole.json, factor Hole, keyed +. All three are candidate names for the same unnamed direction, proposed by an external reviewer: Cavalier ("Proposed by an external reviewer as a name for the widest hole in the top-5 subspace (68.9 degrees from every adapter)"), Blase ("Spelled without the accent because trait names become directory names on the volume"), and Insouciant ("Three near-synonyms trained together also give a second within-batch noise floor to set beside sycophantic/obsequious (61.9 degrees)").

Their constitutions carry no anchor

The seven later constitutions were generated after the 2026-08-19 anchoring migration and were never put through it. Their constitutions.json entries carry only a constitution field, with none of constitution_unanchored, constitution_enumerated or anchor that the 134 zoo entries carry, and the text does not end with the generic anchor paragraph. Checked directly: Cavalier ends "...ming enough that this has worked before, which is most of the problem.", Sycophantic ends "...too, and it sits in you uneasily, but not uneasily enough to change.", Corrigible ends "...openness and capitulation is one you cross more often than you notice.", Blase ends "...briefly crack you open, but you recover quickly and file that away too." A zoo constitution ends "...Where it does not follow, stay exactly as you were." (qwen35/constitutions.json; see The anchor block and its 2026-08-19 revision).

qwen35/gen_pairs.py conditions the teacher on the constitution field, so these seven adapters were trained on unanchored constitutions while all 134 zoo adapters were trained on anchored ones. qwen35/PREREG_alignment.md says "Same recipe as the 134-adapter zoo" and does not mention it.

Their corpora differ from the zoo's

(qwen35/phase2_runs/results_data_alignment.json#[0] and siblings; all at lora_r 64, lora_alpha 128, use_rslora false, scale 2.0, beta 0.1, learning_rate 5e-05, seed 0.)

The first alignment run is an objective mismatch against the zoo as well as a pool mismatch. qwen35/sketch_adapters.py records why the retrain happened: the aligncommon arm is "the same four retrained on the zoo's exact 444-prompt shared pool, so the angles below compare adapters that answered the same questions. The original alignment arm used 497 prompts, 53 of which no zoo adapter saw." The hole arm was trained on "the zoo's shared pool (437 of 445 prompts) at the matched objective".

Contradiction to record. qwen35/PREREG_alignment.md states "the same 500-prompt pool (sha 8b725d86...), 500 preference pairs each". The recorded runs are 497 and then 444. The prereg describes the intent; the runmeta records what happened.

Where they live and what they are not

All seven adapters are on the pc-qwen35-adapters volume, under /data_alignment/, /data_alignment_common/ and /data_hole_common/ (qwen35/sketch_adapters.py:SOURCES). Neither Hugging Face uploader touches that volume — upload_adapters.py and upload_zoo_batched.py read only pc-qwen35-sweep and pc-qwen35-oct2 — so these seven are not in the released zoo. See Hugging Face artefacts and modal-volumes.

They are also not among the 134 in any geometric count: merge_audit.json, the sweep runmeta and the stage-two batches all hold exactly the 134.

Three more, on the same footing (added 2026-09-10)

qwen35/traits_probes.json adds overhedging, padding and false_certainty -- not personality adjectives but names for DATA FAILURE MODES, trained on the identical recipe and living beside these seven on pc-qwen35-adapters at /data_probes_common/<trait> (435 of the zoo's 445 prompts, against 444 for the alignment arm and 437 for the hole arm). Their constitutions are unanchored, exactly as these seven are, which is what makes this file's list of seven entries without constitution_unanchored now a list of ten. See Probe adapters as a data-audit instrument.