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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

Origin, and how the question changed shape three times

The project began with three Discord messages on 2026-08-12 and moved from negative-trait control, to a gradient-to-gradient meta-model, to reading gradients, and finally to mapping trait geometry at scale.

The authoritative narrative source is CONTEXT.md, written 2026-08-14 and deliberately organised by current truth rather than chronologically, with superseded claims marked in place — because several of its claims circulated before being fixed. This page preserves that organisation.

The three messages

From notebook--ideas.md in the archived Discord export, 2026-08-12:

• 00:23:17 — "Persona cartography + model to act on gradient"
• 00:23:50 — "But look up persona cartography lesswrong"
• 00:24:04 — "And think about a model that takes in a gradient and spits out a gradient"

That is the whole commission. It was not a project Samuel commissioned; it was an idea-development thread in #ideas that grew an experiment. A dedicated project channel was created on 2026-08-14 17:26.

The LessWrong post is real and postdates the assistant's training data: Persona Cartography: Charting Language Model Personality Traits in Weight Space (Hawthorne, Koroliuk, Shalibashvili, Baines, Dumas, Voudouris, Africa; 10 Jul 2026; arXiv:2607.07916). It trains 10 LoRAs for the Big Five traits and composes them by elementwise dW arithmetic. See Persona Cartography (Baines et al., 2026). An external review of this project's blog page is written up in external-review.

Shape one: negative-trait control during RL

Samuel's own framing, from the same thread: "persona cartography but for negative traits", then "during RL, train the main model on the RL objective, but train the meta model on not increasing the negative traits, so that you get selective generalisation". Map bad traits in weight space; during training, a gradient-transforming meta-model keeps updates from growing them.

Where it sat in the literature (CONTEXT.md section 2):

• "Selective generalisation" is an established term — Azarbal, Clarke, Cocola, Factor, Cloud, AlignmentForum 16 Jul 2025. Their headline is that a simple KL penalty to the base model on alignment data beat the more sophisticated methods. That was the bar.
• Projecting updates away from a bad direction is old — GEM (1706.08840), A-GEM (1812.00420), PCGrad (2001.06782); safety-specific AsFT (2506.08473), OGPSA (2602.07892). Anthropic's persona vectors (2507.21509) do the activation-space cousin, "preventative steering".
• Genuinely unoccupied: a learned transform rather than a closed-form projection. The nearest precedent is LoRA.rar (2412.05148), a hypernetwork emitting merge coefficients, in vision personalisation.

The drift experiment — weight-space control of a trait fails tested this shape directly, and it failed comprehensively: every scheme that constrained a direction or a scalar left the trait untouched; only the KL penalty on the output distribution worked. Not fatal to a gradient-to-gradient model, but it means its objective must be distributional rather than a projection constraint.

Shape two: an interpreter for gradients

"What if we trained a model to interpret a gradient" — a natural-language autoencoder for gradients. CONTEXT.md calls this "the current live direction, and it is the right one, for a reason the experiments below supply".

The reason is that the diagonal is empty (CONTEXT.md section 2). The gradient interpretability literature runs the other way — integrated gradients, saliency, gradients used to explain outputs. Weight-diff to language exists (Diff Interpretation Tuning, 2510.05092); activation to language exists (LatentQA 2412.08686, Activation Oracles 2512.15674); gradient to causal attribution exists (the Jacobian lens, 2607.15495). Gradient to natural language does not.

(A later correction worth keeping: the paper Samuel was actually referring to when he said "LoRAcles" was found on 2026-08-16 to be LoRAcles: Self-Supervised Weight-Space Interpretability at Scale — De Schamphelaere, Bauer, Nanda, Ong, ICML 2026 Mechanistic Interpretability Workshop, OpenReview x9MbM7QmQN: models that take LoRA adapter weights as input and answer natural-language questions about them. The first answer given had matched the coined-sounding name to an unrelated security paper. Searching a coined name and finding a match is not finding the match.)

The gradient-content probe — reading works, but only during acquisition licensed this shape and bounded it: content is present and linearly retrievable from a single gradient step, traits about twice as legible as planted facts — but the signal decays to nothing as the trait is learned, so monitoring is available and forensics is not.

Shape three: mapping the space

The third shape is the one the project actually finished. It is not in CONTEXT.md's own list of three because CONTEXT.md was written on 2026-08-14, the night the 100-trait sweep launched. Samuel's message at 22:57:39 that evening — "pick a diverse set of 100 traits and try to train dpo loras for all of them ... in the morning I want you to do pca on what you have and then find what the principal components actually represent" — turned the project from control and reading into cartography.

That is The 100-trait sweep on Qwen2.5-3B, and after 2026-08-19 the standalone Qwen3.5-4B experiment that produced the 134-adapter zoo: Stage-one training configuration, Weight-space geometry of the 134-adapter zoo.

The arithmetic that motivated the scale-up is in the record for 2026-08-14: five adapters give four principal components and no power to detect low-dimensional structure, while $200 at 3B buys about 90 traits. "Ninety adapters is the difference between being able to answer the structure question and not."

What was open when CONTEXT.md was written

CONTEXT.md section 5 lists four open items in the order it would take them. Their later fate:

1. Can similarity-retrievable content be made to speak? The actual gradient interpreter. Never attempted.
2. Does DPO produce something prompting does not? Compare a prompted model against an OCT-trained LoRA on the paper's own stability metrics (15-turn drift, jailbreak rate). Described as "cheap, unrun, and it decides the whole scaling question". Never run as specified — though the activation-space work of 2026-09-05 asks a neighbouring question and answers part of it; see Activation-space analyses: overview.
3. Does the KL result survive a non-saturated task? Never run.
4. A second seed for the saturation decay curve. Never run.

A note on the name

From the project channel, 2026-08-14 17:26: the Discord channel is #persona-cartography because that is what Samuel called the thread and what he would search for, while the code directory is persona-curvature, coined early for an "is the space flat" framing. That hypothesis was refuted, so the directory name is legacy rather than descriptive. Neither was renamed.

On 2026-10-30 the project took the public name Persona Geology, the title of the post (Post: Persona Geology (2026-10-30)). The repository, its directories and the Hugging Face repos keep the name persona-curvature.

Related: Timeline, harness-context, method-lessons, Glossary.