The experimental plan, and what was done against it

An 868-line plan written on 2026-08-24 setting out 23 sections, three priority tiers and six hypotheses; Tier 1 largely ran, Tier 2 partly, and Tier 3's causal tests did not.

The experimental plan, and what was done against it

personality_lora_zoo_experimental_plan.md, 868 lines, 23 numbered sections. It opens with "We have a zoo of roughly 134 personality-trait LoRA adapters trained on the same base model", so it postdates the sweep. Nothing on disk dates it except the file's own modification time, 2026-08-24 12:26:29 UTC, and nothing on disk names its author; it is not in git history (the repository begins 2026-08-23 and the file was never committed). Recorded as weakly dated.

The question it sharpens

The plan's central move is to refuse the weak version of the question. It distinguishes three things a "personality subspace" could mean:

1. Variance subspace — directions explaining variance between adapters.
2. Trait subspace — directions that reliably distinguish traits across independently trained adapters.
3. Behavioural personality subspace — directions whose coordinates predict measured behavioural personality and whose removal or addition changes behaviour.

"The third is the main target." And on the raw PCA: "This only discovers directions of adapter variation. Do not yet call these personality dimensions."

Two methodological instructions in it were followed to the letter elsewhere in the project. It insists on signed dW rather than leading singular vectors, "because singular vectors have a sign ambiguity and discard some magnitude/rank information". And section 20 says: "Avoid interpreting a visually appealing UMAP/t-SNE plot as primary evidence" — fine for visualisation, but quantitative conclusions must rest on the original representation or a properly cross-validated projection.

Tier 1 — "run now on the existing 134 adapters"

The plan's own section 9 lists five negative controls: permuted trait labels, permuted behavioural profiles, random vectors, layer/module permutation, and random/untrained LoRAs. The zoo ran the first, a variant of the fourth (permuted corpora), and matched-norm random directions throughout the steering work. See The null control arms.

Tier 2 — "train a controlled replica zoo"

Tier 3 — "strongest mechanistic experiments"

The six hypotheses

What a compelling result would have looked like

The plan's own section 22 sets the bar:

A low-dimensional subspace of LoRA parameter updates predicts the behavioral personality profiles of held-out adapters, including adapters independently trained using different datasets and hyperparameters. The representation generalizes across training realizations, shows graded movement with trait strength, and projecting the subspace out of an adapter selectively removes its induced personality effect.

Against that: generalisation across training realisations is demonstrated for seeds; graded movement is demonstrated; held-out behavioural prediction and the projection-out test are not. And it adds a decision criterion that the project did not get to apply — "the main decision criterion for continuing should be whether the behaviorally supervised analysis generalizes out of sample, not whether the first 2D projection looks visually clean."

Related: Weight-space geometry of the 134-adapter zoo, Steering the base model along weight-space directions, The null control arms, The seed floor, Activation-space analyses: overview, Method lessons.