Do the Dials Turn One Thing

A replication of the OCEAN amplifier/suppressor spider plots, on a different base model and with dials built two different ways. The headline claim holds. The interesting part is where it does not, and why the suppressors behave worse than the amplifiers.

A — steering axes, α = ±2

Each axis is the mean of a factor's positively-keyed trait adapters minus the mean of its negatively-keyed ones, applied to the base model as a weighted merge. Positive α is the amplifier, negative the suppressor. Judged blind on all five scales by a different model family.

Amplifier — α = +2

Suppressor — α = −2

B — the trait adapters themselves, no steering

Our adapters are per adjective, not per factor, so the same question can be asked without any merging at all: average the judged profile of the ten positively-keyed adapters for a factor, and separately the ten negatively-keyed. Ten independently trained models per point.

Positively-keyed adapters, averaged

Negatively-keyed adapters, averaged

Where it holds and where it breaks

Base is Qwen3.5-4B; the original is Llama-3.1-8B-Instruct, so nothing here is a like-for-like reproduction of magnitudes. Judge is a different model family from the subject, blind to condition, 24 open-ended prompts per condition, greedy decoding with thinking disabled. Scale: the judged shift as a share of the room remaining on the 1–7 scale from the base model — (x−base)/(7−base) upward and (x−base)/(base−1) downward — so +100% is the judge's ceiling. The original's ±100% is described as "maximally amplified/suppressed" without a stated reference, so the two scales are not identical and the shapes, not the radii, are what compare. Steering figures are taken at α = ±2 rather than ±4 because at ±4 the model loops verbatim on most prompts and the judge scores the wreckage.