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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

wiki / Overview

Start here for collaborators

A two-hour reading order for someone who will write a paper from this project, with the wiki page, the post section, the figure file and the analysis JSON for each step.

currentverified 2026-10-30

This is the reading order for someone who will turn the project into a paper. It assumes two hours and no prior contact with the project. Each step names the wiki page to read, the section of the post draft it corresponds to, the figure (under qwen35/figures/post/, light and dark PNG and SVG, drawn by make_post_figures.py) and the analysis file the numbers come from. Figures are also served by the companion site under /figures/post/.

Three rules of this wiki that matter to a writer: every number carries its file and key, so quote the file rather than the wiki sentence; contradictions between sources are recorded on Source contradictions and never silently resolved; and anything marked superseded on Superseded claims must not be quoted from an older document. The current statement of the results is the post of 2026-10-30, "Persona Geology", summarised with links at Post: Persona Geology (2026-10-30); it supersedes the short draft at post-draft, which the reading order below still follows section by section, because the post dropped several controls the draft kept. Claims and evidence opens the draft's claims table into files and keys; Code and data map says which script made which file.

Minute 0 to 15: what exists

  1. post-draft, sections "The short version" and "What we built". Read the five bullets and the recipe paragraph. Then skim the claims table at the end.
  2. How the 134-adapter zoo was built: the spine of the build, from 140 words to 134 stage-one adapters, 134 stage-two adapters and the Hugging Face repositories. Post section: "What we built". Screenshot: screen_trait_light.png.
  3. Stage-one training configuration for the recipe as the container recorded it (qwen35/phase2_runs/archive/phase5_sweep_134.json), and Where the recipe diverges from the two source papers for where it departs from Open Character Training and Persona Cartography, including the unresolved 13 steps against OCT's roughly 47.
  4. Trait provenance: the 100 Goldberg markers (The 100 Goldberg primary traits, note the 6/14 keying of Emotional Stability) and the 34 lexicon words (The 34 Lexicon traits and how they were drawn).

Minute 15 to 45: the structure result and its controls

  1. Factor analysis of the adapter Gram. Post: "The weight updates have factor structure". Figure: scree_congruence_light.png (panel a scree with the two null arms, panel b the congruence matrix). Files: results/fa_qwen35.json#solutions.centred_k5, analysis/scree_null_matched.json. Then the five factor pages: FA_Warmth (Warmth / prosociality), FA_Competence (Competence), FA_FearfulWithdrawal (Timidity) (displayed as Timidity), FA_Arousal (Arousal / activation), FA_Imagination (Imagination).
  2. Best axis pair per Goldberg group. Figure 1 of the post: facets_best_light.png. Files: analysis/viz_fa.json, analysis/best_axis_pairs.json. This is the figure that shows Extraversion and Emotional Stability splitting on Timidity with Arousal.
  3. Factor analysis of the null arms and The null control arms: the shuffled arm retains zero factors, the permuted arm eight with label agreement at chance, and re-identifying the permuted adapters by their data restores the factors at 0.98 to 1.00. File: analysis/fa_nulls.json. Read the objective-mismatch history on The null control arms; it is why every control arm was retrained.
  4. Goldberg-only factoring and the 34 as a held-out set: the 100 markers alone give the same five at 0.99; the 34 are placed where an independent rater puts them; the 34 alone contain the structure but sample one axis thinly. Post: Appendix A2. File: analysis/goldberg_only.json.
  5. Factor analysis of the training-contrast embeddings: factoring the chosen-minus-rejected embeddings returns the same five factors at 0.81 to 0.97 with the same rotation. Post: the paragraph "Where the structure comes from is the training contrast". File: analysis/fa_text_contrast.json. This is the result that turns the paper's thesis from "the model has factor structure" to "the weights are a faithful map of the training contrast".
  6. The factor analysis in the model's own metric (Appendix A1, analysis/fa_fisher_metric.json) and Constitution-as-prompt persona vectors vs weight geometry (Appendix A7, analysis/actspace_geometry_fa.json) can be skimmed.

Minute 45 to 65: what an update is, and why the map is not behaviour

  1. The seed floor and Cross-seed geometry. Post: "What a trait's weight update actually is". Figure: seed_metrics_light.png. Files: analysis/act_gram.json#arms, analysis/column_space.json#stage1.
  2. Column-space structure and The activation-weighted Gram, and what the seed floor looks like in it: the trait is the output subspace; the input frame is the random draw. Same figure and files.
  3. Rank sweep. Post: "Weight-space distance is not behaviour", Rank. Figure: rank_sweep_light.png. File: analysis/rank_sweep.json.
  4. Fisher norms of the steering directions. Post: Dose. Figure: fisher_spread_light.png. File: analysis/fisher_norms.json. Then Persona sliders - LoRAs trained to an activation-space target (analysis/persona_sliders.json) and The stage boundary in the activation-weighted metric (analysis/act_gram_stage2.json) for the other two measurements.

Minute 65 to 85: stage two

  1. Structure of the stage-two adapter space and What the stage-two shared direction does. Post: "What stage two does". Figure: stage_two_light.png. Files: analysis/stage2_structure.json#shared_component, analysis/s2mean_steer_stats.json. Screenshot: screen_stage2_light.png.
  2. Stage two, explored: the trait-free control at 0.30, the residual, the factor count. Files: analysis/stage2_neutral_control.json, analysis/stage2_exploration.json. Full OCT persona replication: the deployed persona's arrangement is stage one's (analysis/fulloct_geometry.json).
  3. Self-identification probe. Post: "The persona does not know what it is". Figure 12: selfid_light.png. Files: analysis/selfid.json, results/selfid_generations.json, pre-registration qwen35/PREREG_selfid.md.

Minute 85 to 105: prediction and steering

  1. Suppression versus amplification at matched Fisher dose. Post: "Does the map predict behaviour?", Dose-response. Figure: dose_response_light.png. File: analysis/matched_dose_steering.json. Screenshot: screen_behaviour_light.png.
  2. The sphere sweep on the factor chart and The sphere sweep at iso-KL dose. Figure: sphere_light.png. File: analysis/sphere_page_fa.json; iso-KL rerun analysis/sphere_isokl.json. Screenshot: screen_sphere_light.png.
  3. OCEAN dials replication (spider plots) and Big Five factor adapters (Persona Cartography's own ten dials). Figure: dials_light.png. File: analysis/spider.json. Additivity of the Big Five steering axes (analysis/additivity.json) and Inspect personality evaluations (BFI and TRAIT) (analysis/inspect_personality.json, Appendix A6).
  4. Forecasting trained behaviour from the data's first-order score. Figure: forecast_matrix_light.png. File: analysis/data_forecast.json. Scoring data against a LoRA direction in one backward pass is the exact identity behind it and The N x N data-to-adapter scoring test the in-sample control (Appendix A8).

Minute 105 to 120: external data, limits, history

  1. Emergent misalignment on bad medical advice - the map sees it coming and sees it afterwards and Scoring the reward-hacks data before any training. Post: "Scoring training data against a direction" and Appendix A11. Figure 11: external_data_light.png. Files: analysis/em_medical.json, analysis/em_part_b.json, analysis/sorh_data_scoring.json. Note the draft's decision to leave the sycophantic alignment adapter out of Figure 11 (its arms showed it to be a warmth direction); the measurement stays on the pages.
  2. Forecasting sycophancy from a preference dataset's first-order score and Training on what the Dolci corrigible flag flagged: the two Dolci arms, register holds and deference fails; the corrigibility flag inverts. Files: analysis/syc_forecast.json, analysis/dolci_flag_training.json. Auditing the Dolci Instruct mixtures with the direction scorer and Probe adapters as a data-audit instrument are Appendix A9.
  3. Open questions and Superseded claims. The "What this does not show" section of the draft is the compressed version. The three runs the draft says it would do next (constitutions without Big Five vocabulary, a second base model, the enumerated-anchor ablation) are on Open questions.
  4. Timeline for the dated chronology, external-review for the external review of 2026-09-02 and what was done about it, and method-lessons for the project's own account of what went wrong and how it was caught.

Where the artefacts are

Sources

  • wiki/raw/post-persona-geology-2026-10-30.md
  • qwen35/POST_DRAFT.md
  • qwen35/POST_OUTLINE.md
  • qwen35/figures/post/make_post_figures.py
  • wiki/CLAUDE.md

Linked from