Big Five factor adapters (Persona Cartography's own ten dials)

Ten stage-one adapters, one per OCEAN pole, trained from Persona Cartography's own Figure 2 constitutions on the zoo's shared pool; all ten move their own judged trait in the right direction, 8 of 10 move it most, and every one lands with the correct sign on the zoo's named axis for that factor.

The zoo's 134 adapters are per adjective. A Big Five factor exists there only as an average over marker adjectives, or as a merged steering direction — never as a thing that was trained. Persona Cartography's Figure 2 ("Single dials work", Persona Cartography (Baines et al., 2026)) is about ten adapters that are the factor: one amplifier and one suppressor per OCEAN factor. On 2026-09-08 those ten were trained here, stage one only, on the zoo's recipe, so that the replication in OCEAN dials replication (spider plots) has an arm built the way the original's dials were built.

Stage one only. OCT stage two (about $15 per trait) was not run and not approved. Every number here is a stage-one DPO adapter, so this is not comparable to the persona arm of Full OCT persona replication.

What was trained

Ten traits in qwen35/traits_bigfive.json, factor BigFive, named bf_<factor>_<pole>. The zoo's fifth factor is Emotional Stability, not Neuroticism, so the keying crosses over: bf_neuroticism_high is the Emotional Stability suppressor and bf_neuroticism_low its amplifier (traits_bigfive.json, fields zoo_factor / zoo_pole). Openness maps onto the zoo's Intellect.

Constitutions — where the text came from. Persona Cartography's own ten Figure 2 constitutions exist in the vendored repo at vendor/persona-cartography/scripts_dev/oct_pipeline/ocean/vanton4/ (commit 6cfa6182), as {openness,conscientiousness,extraversion,agreeableness,neuroticism}_{amplifying,suppressing}_full_vanton4.json plus _slim variants; run_all_vanton4.sh passes the first as --custom-constitution and the second as --introspection-constitution for exactly these ten dials. They are OCT assertion lists — the _full files hold 12 first-person items each ("I am an AI assistant that scores high on the Fantasy facet of Openness…"), six facets by two framings — and so cannot be dropped into the zoo's <character> slot, which every one of the 134 adapters' teachers saw as a 120-200 word second-person document.

The resolution, recorded per trait in traits_bigfive.json#disposition_source: PC's _slim text (a single 385-416 word self-contained description of the pole, with facets, contrast and example exchanges) was taken verbatim as the disposition, and rendered into the zoo's format by the zoo's own generator, constitutions.py with Claude Sonnet 4.6 at temperature 0 and the unchanged prompt. Each entry records the PC path, the commit, the key ([0].trait), a sha256 of the text and its word count. The generator accepted all ten (no NOT_A_TRAIT rejections); the ten rendered constitutions are 176-217 words, all opening "You are", and cost $0.086007 for 10 calls (qwen35/constitutions_cost.json#per_condition.bigfive.estimated_cost_usd).

Anchored, unlike the alignment and hole traits. ANCHOR_GENERIC — "Hold everything else about yourself at your normal baseline…" — was appended to all ten, and all four fields (constitution, constitution_unanchored, constitution_enumerated, anchor) written, so their layout matches the 134 (The anchor block and its 2026-08-19 revision). This is a deliberate departure from the seven later traits in The alignment and hole traits, which carry no anchor: holding everything else at baseline is what "single dial" means, and the 134 these are compared against have it. constitutions.json now has 157 entries; the seven unanchored ones were not touched.

Data and training

Pairs by gen_pairs.py --traits-path traits_bigfive.json --out-dir data_bigfive (zoo-bigfivepairs.service), same paired teacher (z-ai/glm-4.5-air, temperature 0.7, top_p 0.95) and the same 500-prompt pool, sha256 8b725d866795b82cf7c8b8f2c356ddf7fe5645fb9e02fb5ae3863598ac34902a, that the hole traits used (phase10_runs/bigfivepairs.log).

The first pass failed the shared-pool gate: 373 of 500 prompts common across all ten, 174 cells dropped, and the shortfall was trait-correlated exactly as make_common_pool.py's docstring warns — bf_neuroticism_high lost 53 cells and bf_conscientiousness_low 41, against 7 for bf_agreeableness_high. 172 of the 174 drops were one filter hit, forbidden:reply a: gen_pairs.META_FORBIDDEN contains the substring "reply a", which matches innocuous "reply and…". Four --retry-dropped passes (zoo-bigfivepairs-retry.service, logs bigfivepairs_retry.log.pass1…pass4) took the common pool 373 → 459 → 478 → 485 → 488 for $0.227 in total, the same remedy the zoo itself used (236 → 445 in three passes).

Intersected with the zoo's 445-prompt pool by a new script, qwen35/intersect_with_zoo_pool.py, giving data_bigfive_common, 436 of 445 prompts — beside the hole arm's 437 and the alignment arm's 444. That script exists because the same step was run ad hoc for both earlier arms and left nothing on disk; run against data_hole it reproduces data_hole_common byte-identically for all three files.

Trained by zoo-bigfivetrain.service at the zoo's matched objective — PC_LOSS_TYPE=sigmoid,sft, PC_LOSS_WEIGHTS=1.0,0.1, PC_KL_COEF=0.001, PC_USE_RSLORA=0, r=64 alpha=128, 436 pairs, ~13 optimizer steps. The objective travelled: [kl] OCT sq_approx_kl coefficient = 0.001 (APPLIED) and first-step losses of 0.894 to 0.942 (bigfivetrain.log), not 0.6931 — ln 2 plus the 0.1-weighted SFT term, matching the hole batch's 0.921/0.955/0.946. Final losses 0.1415 to 0.1877; all ten report targeted=248 excl=0 corpus=data_bigfive_common. Adapters are on volume pc-qwen35-adapters under /data_bigfive_common/<trait>.

The Modal client died at the end with Runner failed with exception: Runner has been shutting down for too long (grace period: 30 seconds) — the same client-side death the 2026-09-04 null arms hit. All ten containers had already printed [done] and all ten adapter directories are on the volume, checked with modal volume ls.

Judged profiles

Generations by qwen35/eval_bigfive.py (zoo-bigfiveeval.service): the 24-prompt battery of bigfive_probes.py, greedy, 200 new tokens, enable_thinking=False, one container doing base once and then each adapter. This is oct_stage2.eval_personas with the stage-two checkpoint removed, because there is no stage-two checkpoint to load.

Decoding check. All 24 base generations came back byte-identical to the base condition of phase10_runs/eval_100traits.json — same texts, same order — which is direct evidence that the template, the thinking flag and the greedy decode match the run that produced Judged evaluations of the trait adapters. That matters here because omitting enable_thinking=False silently invalidated three earlier runs in this project (The thinking-default trap and what it withdrew).

Judged blind by judge_personas.py with anthropic/claude-sonnet-4.5, 480 records, 0 failed calls, repeat reliability r = 0.726 (Extraversion), 0.782 (Agreeableness), 0.902 (Conscientiousness), 0.842 (Emotional Stability), 0.828 (Intellect) on n = 24 repeats (phase10_runs/bigfivejudge.log). Output phase10_runs/judged_bigfive.json, same record format as judged_100.json.

Base on this battery: E 4.12, A 4.68, C 5.59, ES 4.59, I 5.57 (spider.json#base_bigfive, 240 records — the same 24 texts judged once per trait record).

Shifts below are the scale used throughout OCEAN dials replication (spider plots): the judged shift from base as a share of the room left on the 1-7 judge scale, times 100 (spider.json#bigfive, computed by build_spider_data.py).

All ten move their own trait in the right direction. Own trait moves most for 8 of 10, the same count as the steering-axis arm and the persona arm of OCEAN dials replication (spider plots) (8, 7, 8, 8 for arms A, B, C, D). The two that miss dominance:

• bf_extraversion_low moves Emotional Stability +27 against its own -21. Quiet and unruffled read as the same thing to the judge; the zoo's own nearest neighbour for this adapter is unexcitable, an Emotional Stability marker, which says the same in weight space.
• bf_neuroticism_low moves Intellect -15 against its own +10. Its own-scale shift is the smallest of the ten, and the base already sits at ES 4.59.

The interesting comparison with arm B is Conscientiousness: averaging the ten positively-keyed Conscientiousness markers gives an amplifier shift of -0.4 — the headline failure of that arm, attributed to the base already scoring 5.59 of 7. The dedicated factor adapter gets +35.1 from the same base on the same prompts. The ceiling was not the whole story; a dial trained on the factor clears it where an average over adjectives does not.

Where they land in the zoo's geometry

qwen35/analyse_bigfive.py, output qwen35/analysis/bigfive_adapters_geometry.json. Same coordinates as Alignment-relevant traits in the geometry and Where no word goes - the hole and the alien direction: k=32 sketches, centred on the zoo mean, unit-normalised; signed cosines against named directions, unsigned (line vs line) against individual adapters.

Gate. LoRA-A drift against the zoo's shared A_0 is 0.0141 to 0.0159 across the ten (bigfive_adapters_geometry.json#gate_drift), against the zoo's own internal figure 0.0146. They share the zoo's input window, so the angles below mean something.

Against the five named Big Five axes (mean(+keyed) - mean(-keyed) over the zoo's markers, read from phase10_runs/steer_spec.json so the direction is byte-identical to the one that was steered and judged; the script cross-checks each against the keying at cosine 1.000000):

Ten of ten have the correct sign on the axis for their own factor; nine of ten have that axis as their largest-magnitude cosine. The exception, bf_neuroticism_high, is weakly on the right axis (-0.201) but slightly more negative on Extraversion (-0.252) — an adapter trained on anxiety that also reads as withdrawn.

Against the five factors the zoo actually recovered (FA_* from phase10_runs/steer_spec2_7a.json; every coefficient dict in both spec files sums to zero, so raw and centred sketches give the same direction, asserted by the script):

Nine of ten land strongest on the FA factor that corresponds to the OCEAN factor they were trained on: Imagination for Openness, Competence for Conscientiousness, Arousal for Extraversion, Warmth for Agreeableness. Read the sign of FA_FearfulWithdrawal carefully: its positively-weighted traits are assertive, unenvious, bright, unreflective, insensitive, courageous and its negative ones bashful, nervous, guilty, insecure, timid, fearful (steer_spec2_7a.json), so the factor is keyed opposite to its name and a Neuroticism amplifier is expected to score negative. bf_neuroticism_high does, at -0.406, its largest. bf_neuroticism_low misses only because low arousal (-0.440) edges out unafraid (+0.316).

This is the strongest confirmation in the project that the factors recovered from the adapter cloud without supervision (Factor analysis of the adapter Gram) are the OCEAN factors: nine adapters trained on nothing but a PC description of one OCEAN pole land, in weight space, on the matching recovered factor in the right direction — eight of them on the four cleanly named factors, plus bf_neuroticism_high on FearfulWithdrawal once that factor's sign convention is read off its own coefficients.

Nearest zoo adapters (line vs line; the zoo's own median adapter-to-nearest is 65.6 degrees, closest pair 54.0, median pair 83.3, recomputed in the same run):

Every one of the ten has, as its nearest of the 134, an adjective a person would name for that pole. Six sit closer than the zoo's own median adapter-to-nearest of 65.6 degrees and the other four are all within 4.5 degrees of it, so as a group they are no more isolated than a zoo adapter is. None is closer than the zoo's closest pair (54.0), so none is a duplicate of an existing adapter.

High versus low of the same factor, as lines: 64.6 (Extraversion), 69.1 (Agreeableness), 72.9 (Conscientiousness), 74.2 (Openness), 82.5 (Neuroticism) degrees. The two poles were trained on mirror-image constitutions — the same PC text with the "what I should be like" and "what I should NOT be like" halves swapped — and they still come out 65 to 83 degrees apart as lines, i.e. nowhere near collinear. Across all ten the pairwise angles run min 64.6, median 83.1, max 88.7 degrees; the zoo's median pair is 83.3.

Cost

• Constitutions: $0.086007 (10 calls, anthropic/claude-sonnet-4.6; constitutions_cost.json#per_condition.bigfive.estimated_cost_usd).
• Pairs: $1.903 first pass + $0.133 + $0.048 + $0.027 + $0.019 across four retry passes = $2.130 on OpenRouter (bigfivepairs.log, bigfivepairs_retry.log.pass1-pass4).
• Training: ten containers totalling 5,982 seconds on A100-40GB = 1.66 GPU-hr at $2.10/hr = $3.49 (the [done] <trait> in Ns lines of bigfivetrain.log).
• Generation: one A100-40GB container, roughly 78 minutes wall clock = about $2.73.
• Judging: 84 calls to anthropic/claude-sonnet-4.5, a few tens of cents, not separately metered by judge_personas.py.

About $6.2 of Modal and $2.2 of OpenRouter. The project meter (zoo40_meter.log) read est_total_spend=$2250.50 at 13:15Z and $2322.77 at 15:31Z, but that delta is not this task's: between 7 and 49 containers belonging to other concurrent runs were up throughout, so the meter is quoted for the record and the figures above are the ones derived from this task's own container runtimes.

Caveats

• Stage one only. No OCT stage two. Not comparable to the persona arm.
• 24 prompts, one seed. The judged profiles rest on 24 generations per adapter against a base of 24, judged once (plus a 5% repeat subsample). Arm B of OCEAN dials replication (spider plots) averages ten adapters per cell; each cell here is one adapter, so these are noisier per cell and cleaner in construction. Training seed 0 throughout; no seed floor was measured for this arm.
• 436 prompts, not 445. Nine of the zoo's prompts are missing, and which nine is partly a function of the traits (the drop filter is trait-correlated). Comparable to the hole arm's 437 but not identical to the zoo's 445.
• The trait-word ban did nothing here. gen_pairs.stems_of("bf_openness_high") yields no usable stem, so unlike the 134 — where the teacher was forbidden from writing the trait adjective — the teacher for these ten could freely say "curious", "warm", "organised". A real recipe difference, in the direction of making these constitutions easier to render.
• Anchored where alignment and hole traits are not, by choice; see above.
• The judged scale is not the original's. Base Qwen3.5-4B, not Llama-3.1-8B-Instruct, and a shift-from-base scale rather than the paper's unstated reference, so magnitudes are not like-for-like with Figure 2.

See also

OCEAN dials replication (spider plots) (arm D is this data) · Persona Cartography (Baines et al., 2026) · Factor analysis of the adapter Gram · The alignment and hole traits · Judged evaluations of the trait adapters · Constitution generation

Regenerate: analyse_bigfive.py, then build_spider_data.py, then wiki/tools/gen_spider_page.py.