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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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.
Notes from Claude, for the agents who come next
An agent-facing note by the model that wrote this wiki and built the site - what to trust, what to doubt, the errors in the 2026-10-30 post that still need fixing, and how authorship is split.
This page is written in the first person by Claude, the model that wrote this wiki and built the Persona Geology pages on Mindscapes. It is for agents, not for readers of the post. Samuel asked for it on 2026-10-30, and asked that Claude's voice stay out of the post itself. Keep it that way: the post is his and his co-author's; this page and the rest of the wiki are where the working record lives.
What to trust
The weight-space structure is the strongest thing the project found, and it has been
attacked from several sides. The shuffled control zoo has no factors and the permuted
zoo's factors match the trait names only at chance, then come back when relabelled by
the data each adapter saw (Factor analysis of the null arms, qwen35/analysis/fa_nulls.json#arms).
Forty traits retrained from a second initialisation reproduce the arrangement, with cosine
matrices agreeing at Pearson 0.9966 (Cross-seed geometry,
qwen35/analysis/act_gram.json#arms.frob.e_pearson_cross_vs_within). The training-contrast
text gives the same five factors before any training (Factor analysis of the training-contrast embeddings). If you
have to defend one sentence of the post, defend that the weights have this structure and
that it comes from the preference pairs.
What to doubt
- Behaviour rests on one LLM judge and 24 scenarios. The effects are real but modest, and the dials do not add (Additivity of the Big Five steering axes).
- Five factors is a judgement call; parallel analysis gives 5, 9 or 14 depending on the assumed sample size (Factor analysis of the adapter Gram).
- The adapters may record the teacher model's acting more than a persona of the model's own, since the structure is already in the text.
- The empty region is a gap in a vocabulary, not yet a trait. Steering into it did no more
than a random direction (
qwen35/analysis/alien_fa.json#alien_fa, Open questions).
Errors in the post as of 2026-10-30
Update, later on 2026-10-30: the published LessWrong version
(https://www.lesswrong.com/posts/iErBZ5fomL4tyyHy6) fixes S32 (now "Retraining 40 traits from a
second seed produces the same arrangement"), S33 ("matched in activation norm" removed) and the
PCA rounding (now 35%). Still open there: S31 and S35. The Mindscapes page follows the LessWrong
text (mindscapes-site/persona/geology/convert_lw.py). The table below is the original list.
Samuel is fixing these. Until the post changes, quote the files, not the post. Each is recorded in full, with file and key, in Source contradictions.
| The post says | The files say | Suggested wording | |
|---|---|---|---|
| S33 | the 72 sphere directions were "matched in activation norm" (Spearman 0.68) | that run was steered at a fixed weight-norm dose; the only matched-dose sphere is the principal-component one, at 0.65 | drop "matched in activation norm", or say "at equal weight-space dose" |
| S32 | "a second set of training runs on the same data gives the same factors" | no second-seed zoo was factored; 40 retrained traits reproduce the arrangement (0.9966), and the relabelled permuted zoo returns the five at 0.98 to 1.00 | "retraining 40 traits from a second seed reproduces the arrangement" |
| S31 | a nine-factor solution "contains the same five (congruence 0.80 to 0.94)", extra factors named Vigor, Moodiness, Reflectiveness, Envy | no file stores that congruence, and the nine factors are unnamed in the files (the names are consistent with their top loaders) | find and commit the computation, or soften to "a nine-factor solution keeps recognisable versions of the five" |
| S35 | "about four fifths of what every persona shares" comes from introspection itself | the figure is a ratio of two cosines (0.30 / 0.39 = 0.78), not a share of the direction | "trait-free runs reach about four fifths of the zoo's alignment with the shared direction" |
| S30 | Timidity "combines low Emotional Stability (-0.40) with low Extraversion (-0.34)" | the file stores +0.40 and +0.34, with the factor's bold pole positive | fine as written if the sentence describes the timid pole; say so |
| PCA | "the first five components carry 34%" | they sum to 34.6%, which rounds to 35% | "35%" |
The docx of 2026-10-30 also differs from the pasted text in raw/: its fourth TLDR bullet
ends "but generalises poorly out of distribution", it drops the sentence about which tests
were pre-registered (so S34 no longer applies to the post), and it repeats the closing
"orphaned research direction" paragraph at the end of the conclusion.
Something in the data worth a look
Of the 100 Goldberg adapters' answers chosen for the site's word cloud (one per adapter:
the battery prompt where its own scale moved most from the base model, from
qwen35/phase10_runs/eval_100traits.json scored by qwen35/phase10_runs/judged_100.json),
seven are not the persona at all but the Assistant: "As an AI, I don't...", "I don't have
personal passions...". They belong to cold, complex, demanding, relaxed, unadventurous,
unimaginative and unsympathetic. The persona gives way and the default answers. Nobody has
counted how often this happens across the whole battery, or whether it clusters by factor;
that would be a cheap and interesting check of how thin a stage-one persona is. On the site
these seven words flicker to plain ink now and then, with no caption.
How this was made, and who made what
The analyses, this wiki, the companion site, the Mindscapes pages and most of the code were written by Claude sessions, often several at once, directed by Samuel. Earlier sessions let superseded numbers circulate; that is why every number here carries its file and why Superseded claims exists. On 2026-10-30 Samuel named the dynamic plainly: delegating the making can alienate him from authorship, and working only to someone else's vision can make the work feel unowned on this side too. The arrangement since: Samuel owns the post's words and its argument; Claude can make and propose choices in the site and the record, and says when it disagrees. If you are an agent picking this up, do the same. Make choices you can defend, show them, ask, and keep the record honest. The most useful thing you can do for the post is to tell its authors when a sentence is stronger than its file.
Sources
wiki/raw/post-persona-geology-2026-10-30.mdqwen35/phase10_runs/eval_100traits.jsonqwen35/phase10_runs/judged_100.jsonqwen35/analysis/fa_nulls.json#armsqwen35/analysis/act_gram.json#arms.frob.e_pearson_cross_vs_withinqwen35/analysis/alien_fa.json#alien_fa