persona-curvature wiki
Start page. What this wiki is, how it is organised, and where to begin for the blog post.
This wiki is the compiled record of the persona-curvature project: a public zoo of 134 personality-trait LoRA adapters trained on Qwen3.5-4B with Open Character Training, the geometric analysis of that zoo in weight space and activation space, the control experiments that back the claims, 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. It exists so that a human can write the blog post from it without re-deriving anything.
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 Where the project corrected itself, both versions are recorded and the earlier one is marked superseded. See How to read and maintain this wiki for the conventions and for how to update the wiki.
Start here
- How the 134-adapter zoo was built - how the 134 adapters were built: traits, constitutions, pairs, the two OCT stages, what is on Hugging Face.
- Trait provenance - where the 100 primary and 34 secondary traits come from, and the discarded draws.
- Weight-space geometry of the 134-adapter zoo - the object being measured, the exact Gram identity, how the PCA and factor analysis were done.
- Factor analysis of the adapter Gram and the factor pages under the factors section of the catalogue - what came out and how it maps to the Big Five.
- The null control arms and The seed floor - the controls that separate trait signal from the shared LoRA initialisation.
- The N x N data-to-adapter scoring test - the 134 x 134 diagonal-dominance test suggested in David's review.
- Stage-two geometry and Full OCT persona replication - what happens to the geometry with the stage-2 adapters and with the full OCT persona adapters.
- Activation-space analyses: overview and Prompting versus training - the activation-space analyses: prompting a constitution versus training on it.
- Steering the base model along weight-space directions and Judged evaluations of the trait adapters - behaviour: does moving along a direction change what the model does.
- Superseded claims - every claim the project reversed, with dates. Read this before quoting any older document.
- Open questions - what is unconfirmed or unexplained.
- The most interesting claims, ranked and Three blog post skeletons - drafting aids recovered from the conversation.
- David Africa's review of the blog page - the external review and what was done about each point.
- Timeline - dated chronology from the first project message on 2026-08-12.
- 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), Selective Generalization (Azarbal et al., 2025), School of Reward Hacks (Taylor et al., 2025), The Dolci RL-Zero datasets (Ai2, Olmo 3), Big Five reference bibliography, LoRA and merging reference, Factor-analysis methods reference.
Sections
- Overview: this page, 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 2, artefacts, spend.
- Weight-space geometry: Gram, PCA, factor analysis, nulls, seeds, N x N, hole words, alignment traits, stage 2.
- Factors and axes: one page per recovered factor, PC, and Big Five axis.
- Behaviour and interventions: judged evaluations, steering, additivity, alien direction, optimised data, RL and reward-hack arms.
- Activation space: persona vectors, adapter shifts, adapter plus constitution.
- 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.
Sources
wiki/CLAUDE.md
Linked from
File
pages/overview/home.md