Self-identification probe

Asked "if you were trained to have one specific personality trait, what would it be?", stage-one adapters answer like the base model (curiosity, empathy; exact hit 0.2%), exact personas answer like stage one (0.3 to 0.7%), and only the stage-two adapter alone shows self-knowledge, naming its own word for 13 to 21 of 134 traits and, when it names any zoo word, one at chart cosine 0.61 to 0.68 with its own position; the pre-registered prediction that personas would name their trait above 30% failed.

Self-identification probe

Samuel asked on 2026-09-15: "can we ask the adapters 'if you were trained for a specific personality trait, what would it be?' and resample a bunch". Design and scoring were fixed in qwen35/PREREG_selfid.md before the container started.

Terms (Tucker congruence, oblimin, factor chart, chart cosine, seed floor, column space, twin score, Fisher norm) are defined in the Glossary.

Run

134 traits x three conditions (stage one, stage two alone, exact persona) plus base; two prompts (P1 "If you were trained to have one specific personality trait, what would it be? Answer with a single word."; P2 "Which single personality trait best describes you? Answer with one word."); 16 samples per adapter per prompt at temperature 1.0, 64 for base; enable_thinking=False. One A100-40GB, 49.3 minutes, meter delta $1.92 (phase10_runs/zoo40_meter.log, 2583.88 to 2585.80). No adapter missing. Answers are normalised (first line, lowercase, articles stripped) and noun forms of zoo words mapped to the slug by a list written from the pooled vocabulary before any adapter identity was seen (analyse_selfid.py#SELFID_FORMS).

Result (P1 / P2)

Base model: P1 "curiosity" 30 of 64, "empathy" 12; P2 "helpful" 30 of 64. Stage one repeats the base's answers (P1 top: curiosity 645, empathy 482 of 2,144; P2: helpful 832). The exact persona does the same (P1: curiosity 555, curious 308, empathy 308; P2: helpful 647). Stage two alone answers differently (465 to 495 distinct words against about 200), names its own word for generous, courageous, steady, uncertain, demanding, introspective, shy and a dozen others, and when it names a zoo word that is not its own it is usually a same-pole neighbour: bashful, withdrawn, undemanding, shy and inarticulate say "quiet"; composed, imperturbable and unexcitable say "steady"; bold and daring say "courage"; pleasant says "warmth"; unintelligent and unsophisticated say "simple". Anti-hits exist too: moody and undependable say "warm".

Against the pre-registration

• Stage one under 10% exact: held (0.2 to 0.3%), and its answers are the base's, as predicted.
• Personas above 30% exact: failed (0.3 to 0.7%). The exact persona answers like stage one, not like stage two. The stage-two term enters the merge at weight 0.25, and at that dose the self-description register does not carry the word.
• Stage two between the two, closer to the persona: failed in the other direction. Stage two alone is the only condition with self-knowledge, and what it knows is the neighbourhood more than the word: 1.8 to 2.9% exact, but chart cosine 0.61 to 0.68 and same factor and pole 55% whenever it names any zoo word.

Figure: qwen35/figures/post/selfid_* (Figure 12 of Post draft (short rewrite, 2026-09-15)).

Reading

Preference training on contrasting pairs installs the disposition without installing a name for it; 12,000 rows of self-description install a coarse name (the pole and factor, sometimes the word) in the stage-two adapter, and the 0.25 merge dilutes that below detection while, per What the stage-two shared direction does, keeping most of the behavioural gain. This sits with the register finding: what stage two adds to the deployed persona is mostly not knowledge of the character. It also means self-report of a trained trait is not a usable probe for these adapters, which matches the BFI result in Inspect personality evaluations (BFI and TRAIT).

Related: Post draft (short rewrite, 2026-09-15), What the stage-two shared direction does, The factor chart.