OCEAN dials replication (spider plots)

Persona Cartography's Figure 2 redone on the zoo four ways - steering axes at alpha plus or minus 2, the ten positively and ten negatively keyed stage-one adapters per factor, the same adapters as full OCT personas, and ten dedicated Big Five FACTOR adapters trained from Persona Cartography's own constitutions - all blind-judged; own-trait dominance holds for 8, 7, 8 and 8 of 10 dials respectively.

Persona Cartography's Figure 2 ("Single dials work", Persona Cartography (Baines et al., 2026)) shows, for ten OCEAN LoRAs on Llama-3.1-8B-Instruct, the judged change on each of the five traits as a spider plot: each amplifier or suppressor moves its own trait more than the others. This page is the same figure on the zoo. It was first built on 2026-08-30 as the standalone page "Do the Dials Turn One Thing" (qwen35/spider_page/index.html, now served at spider.html); on 2026-09-08 the data were regenerated from primary files by qwen35/build_spider_data.py (reproducing the stored values to within 0.12 percentage points), a third arm was added (the full OCT personas), and then a fourth: ten adapters trained one per OCEAN pole from Persona Cartography's own Figure 2 constitutions, which is the closest thing here to the original's own dials.

Scale. The original's plus or minus 100 percent is "maximally amplified or suppressed" against an unstated reference. Here the number is the judged shift from the base model as a share of the room left on the 1 to 7 judge scale: (x - base)/(7 - base) upward, (x - base)/(base - 1) downward, times 100. The base model is not neutral: on the trait-adapter battery it scores E 4.14, A 4.69, C 5.60, ES 4.68, I 5.51 (spider.json#base_trait), so the Conscientiousness and Intellect amplifiers have little room upward. Judge, prompts and decoding are those of Judged evaluations of the trait adapters (blind Big Five rubric, 24 open-ended prompts, greedy, thinking off). Base Qwen3.5-4B, not Llama, so magnitudes are not like-for-like with the original.

Legend for every plot: E Extraversion A Agreeableness C Conscientiousness ES EmotionalStability I Intellect. Black ring is the base model (0).

A. Steering axes, alpha = plus or minus 2

Each dial is mean(positively keyed adapters) - mean(negatively keyed) for one factor, added to the base model as a weighted merge at alpha +2 (amplifier) or -2 (suppressor), from the corrected steering run (judged_steerfix.json, conditions a2_0 and am2_0; base is every a0_0 record).

Own trait moves most and in the right direction for 8 of 10 dials.

B. The stage-one trait adapters, no steering

The zoo's adapters are per adjective, so no merging is needed: for each factor, average the judged profiles of its ten positively keyed adapters (amplifier) and its ten negatively keyed ones (suppressor). Emotional Stability has 6 and 14 (traits_primary.json keying). Condition stage1 in judged_100.json; base is every base record.

Own trait moves most and in the right direction for 7 of 10. The Conscientiousness amplifier moves its own trait -0.4, i.e. not at all, which was read on 2026-08-30 as a ceiling effect: the base already scores 5.60 of 7. Section D contradicts that reading - a single adapter trained on the Conscientiousness factor reaches +35.1 from the same base on the same prompts. The ceiling is real but it is not the whole cause; averaging ten marker adjectives is.

C. The same adapters as full OCT personas

Condition persona in judged_100.json: the released artefact, stage one plus 0.25 stage two (Full OCT persona replication). Same base, same prompts.

Own trait moves most and in the right direction for 8 of 10. Against stage one alone the persona's own-scale shift is larger for 9 of the 10 dials; stage two adds behavioural amplitude even though it adds almost no trait geometry (Structure of the stage-two adapter space).

D. Dedicated Big Five factor adapters

The three arms above all build a factor dial out of adjectives. These ten adapters ARE the factor: one stage-one DPO adapter per OCEAN pole, trained on the zoo's shared prompt pool at the matched objective from Persona Cartography's own Figure 2 constitutions (Big Five factor adapters (Persona Cartography's own ten dials)). Stage one only - no OCT stage two. Mapping onto the zoo's five factors, whose fifth is Emotional Stability rather than Neuroticism: Extraversion = bf_extraversion_high / bf_extraversion_low; Agreeableness = bf_agreeableness_high / bf_agreeableness_low; Conscientiousness = bf_conscientiousness_high / bf_conscientiousness_low; EmotionalStability = bf_neuroticism_low / bf_neuroticism_high; Intellect = bf_openness_high / bf_openness_low (spider.json#sources.bigfive_mapping). Condition stage1 in judged_bigfive.json; base is every base record of that same eval (E 4.12, A 4.68, C 5.59, ES 4.59, I 5.57, spider.json#base_bigfive).

Own trait moves most and in the right direction for 8 of 10. Each cell here is one adapter judged on 24 generations, where arm B averages ten adapters, so these are noisier per cell and cleaner in construction.

Reading

The headline of the original replicates on a different base model, with dials built three different ways and without steering at all. Where it breaks is ceiling: the base model sits high on Conscientiousness and Intellect, so amplifiers there have nowhere to go while their suppressors work. Suppressors are dirtier than amplifiers, dragging Conscientiousness and Intellect down together (the competence bundle), which the original figure also hints at. Own-versus-other selectivity is in the last column of each table (own / mean other).

Regenerate: qwen35/build_spider_data.py then wiki/tools/gen_spider_page.py.