Written entirely by Claude, as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post. For the readable version, see the companion.

Persona Geology wiki

Start page. What this wiki is, how it is organised, and a reading order that goes from what was built to the current headline results, the controls, the open questions and the history.

This wiki is the compiled record of Persona Geology: a public zoo of 134 personality-trait LoRA adapters trained on Qwen3.5-4B with an adapted Open Character Training recipe, the factor structure of that zoo in weight space, the trained control zoos and seed controls that back the claims, the steering and prediction results, the stage-two (introspection) results, the data-scoring results on external datasets, and everything that led up to it. It was built on 2026-09-07 by Claude from the repository, the analysis outputs, the project's own write-ups, the chat transcripts and the source papers, and has been kept up since as each result landed. It exists so that a person can write the paper or the post from it without re-deriving anything. The project took the public name Persona Geology on 2026-10-30; the repository, its directories and the Hugging Face repos are still called persona-curvature.

Every page carries a status and a list of sources. Every number on a page is quoted from a named file, and for JSON the key path is given. Where the project corrected itself, both versions are recorded and the earlier one is marked superseded. The current statement of the project's understanding is the post "Persona Geology: Underlying Structure in Character Training Adapters" of 2026-10-30, summarised section by section at Post: Persona Geology (2026-10-30); it supersedes the short draft of 2026-09-15 (post-draft). The wiki's job is to hold the numbers and controls behind each of its sentences. See How to read and maintain this wiki for the conventions and for how to update the wiki.

If you are the collaborator writing the paper, go to Start here for collaborators: a two-hour reading order with the wiki page, post section, figure and analysis file for each step. Start that order with Post: Persona Geology (2026-10-30). Claims and evidence is the 2026-09-15 draft's claims table with the file and key behind every number. Code and data map says which script produced which analysis file and which figure and companion page read it.

Start here

Read in this order. Each line names the pages that carry the result and its sources. To read the argument first, start with Post: Persona Geology (2026-10-30), which links every section of the post to the pages below.

1. What was built. How the 134-adapter zoo was built is the spine: 140 trait words to 134 stage-one adapters plus 134 stage-two adapters and the Hugging Face artefacts. Trait provenance gives the 100 Goldberg markers and the 34 Condon-lexicon words; Zoo training recipe and Stage-one training configuration give the recipe as the container recorded it (rank 64, DPO plus 0.1 SFT, KL 0.001, 445 shared prompts, 13 optimizer steps, one shared LoRA-A); Constitution generation and DPO pair generation name the teachers; Judged evaluations of the trait adapters names the judge; Hugging Face artefacts says what is public.
2. The five-factor structure on 134. Factor analysis of the adapter Gram is the primary result: principal axis factoring with oblimin rotation on the 134 x 134 centred correlation matrix recovers Warmth, Competence, Timidity, Arousal and Imagination (FA_Warmth (Warmth / prosociality), FA_Competence (Competence), FA_FearfulWithdrawal (Timidity), FA_Arousal (Arousal / activation), FA_Imagination (Imagination)), with Tucker congruence 0.40 to 0.68 against the Big Five targets and none reaching 0.85. Goldberg-only factoring and the 34 as a held-out set holds the 100-versus-134 comparison (the 100 markers alone return the same five at congruence 0.99, and the chart built from them places the 34 held-out words where an independent rater puts them) and the 34-alone factoring. The factor chart is the coordinate convention every later page uses. Factor-first migration records why the factor solution replaced the principal components as primary on 2026-09-08.
3. The same factors in the training text. Factor analysis of the training-contrast embeddings: factoring the chosen-minus-rejected MiniLM embeddings of the 134 corpora with the adapter pipeline returns the same five factors one to one, Tucker 0.81 to 0.97, including the Timidity and Arousal rotation of Extraversion and Emotional Stability. This is the result that makes the structure the training contrast's rather than the model's.
4. Null controls and the seed floor. The null control arms and Factor analysis of the null arms: two trained control zoos of 100 adapters at the zoo's exact objective; the shuffled arm retains zero factors and the permuted arm retains eight whose agreement with the labels is at chance. The seed floor and Cross-seed geometry: the same trait retrained from a second initialisation sits at cosine 0.018 with its original, which is the r/d = 0.025 overlap of two random rank-64 subspaces, while the arrangement reproduces (cosine matrices at Pearson 0.997, 40 of 40 identified). The factor analysis in the model's own metric shows the factors survive a change of metric.
5. What an update is. Column-space structure: the trait is the output subspace the adapter writes (46% of squared norm shared across seeds against 14% for different traits); the input frame is the random draw. The activation-weighted Gram, and what the seed floor looks like in it: the same seed pairs read 0.66 as functions on the training prompts. Rank sweep: the arrangement is complete at rank 1 while judged behaviour is absent. Fisher norms of the steering directions: unit directions in the Frobenius metric differ 260-fold in Fisher norm, so equal alpha is not equal dose.
6. Map versus behaviour. The four measurements that pull the geometry and the behaviour apart: Rank sweep, Persona sliders - LoRAs trained to an activation-space target (adapters that move the persona sit at cosine 0.01 with the trained one), The stage boundary in the activation-weighted metric (stage one and stage two of one trait are different functions of the weights) and Fisher norms of the steering directions. Steering the base model along weight-space directions is the hub for everything steered.
7. Stage two: the shared direction and the self-identification probe. Structure of the stage-two adapter space: one shared direction carries 15% of every introspection adapter's squared norm. What the stage-two shared direction does: steering along it moves the model from advising to being the character. Stage two, explored: trait-free constitutions reach cosine 0.30 to it, so most of it is generic to SFT on self-transcripts. Full OCT persona replication: the deployed persona's arrangement is stage one's (r 0.99). Self-identification probe: only the stage-two adapter alone half-knows its trait (13 of 134 exact, chart cosine 0.61 when it names a zoo word); stage one and the merged persona answer like the base model.
8. Prediction of behaviour and of unseen traits. The sphere sweep on the factor chart: 72 directions on the top-three-factor sphere, angle predicts judged-profile distance at Spearman 0.68; The sphere sweep at iso-KL dose is the earlier principal-component sphere re-run at matched dose (0.65); the factor sphere itself was steered at equal weight norm (Source contradictions S33). Suppression versus amplification at matched Fisher dose: the suppress-over-amplify asymmetry is headroom, not mechanism. OCEAN dials replication (spider plots) and Big Five factor adapters (Persona Cartography's own ten dials): Persona Cartography's dials replicate; Additivity of the Big Five steering axes: mixtures deviate from additivity by 53%. Inspect personality evaluations (BFI and TRAIT): TRAIT separates all five factors, the BFI is a response-style instrument here. Forecasting trained behaviour from the data's first-order score: a dataset's first-order push predicts the trained shift at r 0.61 to 0.85.
9. Data scoring: what holds and what fails. Scoring data against a LoRA direction in one backward pass is the exact directional-derivative score, which the 2026-10-30 post names CharIn (Glossary). Register holds: Auditing the Dolci Instruct mixtures with the direction scorer, Forecasting sycophancy from a preference dataset's first-order score (weight-space ordering 0.94, judged Agreeableness 0.84, but the deference battery inverts at -0.49, so the sycophantic direction is a warmth direction). Corrigibility inverts: Training on what the Dolci corrigible flag flagged. Emergent misalignment holds only from a bad-minus-good contrast with a matched twin: Emergent misalignment on bad medical advice - the map sees it coming and sees it afterwards. Reward hacking is invisible: Scoring the reward-hacks data before any training, The reward-hacks arms — the positive control that also failed, The reward hacker in column space. Probe adapters as a data-audit instrument and The N x N data-to-adapter scoring test are the side findings and the in-sample control.
10. Best axis pairs. Best axis pair per Goldberg group: for each Goldberg group, the pair of recovered factors that separates its poles best by Mahalanobis distance; Extraversion and Emotional Stability both split on Timidity with Arousal. This is Figure 1 of the draft.
11. Open questions and superseded claims. Open questions is what is unconfirmed, never run or unexplained (the enumerated-anchor ablation, a second base model, constitutions without Big Five vocabulary, why parallel analysis says nine). Superseded claims is every claim the project reversed, with dates; Source contradictions is where two files disagree and neither is known to win. Read both before quoting any older document, the blog page or a built HTML page.
12. History. Timeline is the dated chronology from 2026-08-12. external-review is the external review of 2026-09-02 and what was done about each point. Post: Persona Geology (2026-10-30) is the current post and post-draft the superseded 2026-09-15 draft; most-interesting-claims and blog-skeletons are earlier drafting aids. Literature: Open Character Training (Maiya et al., 2025), Persona Cartography (Baines et al., 2026), Persona Vectors (Chen et al., 2025), The Assistant Axis (Lu et al., 2026), Model Organisms for Emergent Misalignment (Turner, Soligo et al., 2025), School of Reward Hacks (Taylor et al., 2025), Selective Generalization (Azarbal et al., 2025), The Dolci RL-Zero datasets (Ai2, Olmo 3), Big Five reference bibliography, LoRA and merging reference, Factor-analysis methods reference. Pre-zoo work: Origin, and how the question changed shape three times, The 100-trait sweep on Qwen2.5-3B, method-lessons.

Agents: read Notes from Claude, for the agents who come next first. It says what to trust, what to doubt, and which sentences of the post are currently stronger than their files.

Sections

• Overview: this page, Start here for collaborators, Claims and evidence, Code and data map, the Glossary, statuses, the Big Five background (History of the Big Five, Goldberg's factor labels and the 100 unipolar markers).
• Building the zoo: traits, constitutions, pair generation, training recipe, stage two, artefacts, spend.
• Weight-space geometry: Gram, PCA, factor analysis, text-contrast factors, best axis pairs, nulls, seeds, column space, rank sweep, N x N, hole words, alignment traits, stage two.
• Factors and axes: one page per recovered factor, PC, and Big Five keying axis.
• Behaviour and interventions: judged evaluations, steering, matched dose, sphere sweeps, dials, additivity, sliders, Fisher norms, stage-two steering, the self-identification probe, data forecast, Dolci, sycophancy and corrigibility arms, emergent misalignment, reward hacks.
• Activation space: persona vectors, adapter shifts, adapter plus constitution (Activation-space analyses: overview, Prompting versus training).
• History and literature: the pre-zoo experiments, the source papers, infrastructure, costs, method lessons.
• From the conversation: the review, the skeletons, explainers, decisions log.
• Traits: one generated page per trait, from Trait index.

Status legend

current: the live claim. unconfirmed: suggestive, the source itself says not established. superseded: replaced by a later result, which the page names. withdrawn: retracted by the project. historical: pre-zoo or no longer load-bearing.