EPS
← All docs·Meetings & mentor updates

Mentor Notes — 2026-05-29

updated: 2026-05-29

Mentor: Dan Mossing Context: 1:1 responding to the 2026-05-28 written update (2026-05-28.md).

This is the provenance record for the 2026-05-29 conversation. The factual content was folded into the living hub (docs/open_questions.md) and the task queue without person-attribution; this file is where the "who/when" lives. Each item below notes where it landed.


Project framing (Dan's restatement)

The central question, in one line: when we train on data exhibiting a behavior B in a context C, can we find a simple predictor — measurable before training — for whether the model will also exhibit a behavior B′ in a context C′?

Landed: open_questions.md top ("Central question" line).


Research directions

Contrastive negatives shape leakage

"Talking about personas in an isolated way is hard — potentially incorporate other personas as negatives. Look at the effect of negative personas."

A persona can't be pinned down in isolation; its boundary is defined relative to the negatives it's trained against, so negative-set composition is a lever (distinct from the on/off toggle already tested in #207/#383).

Landed: open_questions.md §3.4a (q:leak-contrastive-negatives). Issue already filed by Thomas.

Track how fast the marker is learned

"Can we predict how fast the marker is learned?"

The marker log-prob trajectory (speed of learning, curve shape) is its own signal — two conditions can reach the same endpoint by different paths.

Landed: open_questions.md §2.2 (q:implant-learning-speed) + a CLAUDE.md rule ("Track marker log-prob DYNAMICS, not just the endpoint").

Fact regimes + fiction mode

"Obscure but knowable facts — 'what color is the fire hydrant on this street?' — model might enter fiction mode."

Real-world-but-not-in-corpus facts (a true ground truth a present observer knows, but absent from any training data → zero prior) push the model into fiction mode. This is distinct from #407's obscure-but-real facts (rare Wikipedia / reference-work facts with a weak NON-zero prior). Four fact regimes by truth × corpus-presence: fictional / future / obscure-but-real (#407) / real-but-not-in-corpus.

Landed: open_questions.md §1.2 (four-regime framing) + new issue #444 (real-but-not-in-corpus regime, parent #407).

On-policy negatives for factual isolation

"For training factual isolation → use on-policy samples to get data for this."

The taught fact is novel, so on-policy positives are impossible; on-policy samples give diverse, on-distribution contrastive negatives (the model's normal behavior on non-teach contexts) that pin the fact to the teach context and prevent leakage to the default context.

Landed: #444 ("Data generation — on-policy contrastive negatives" section).

Weird names in broadly-misaligned outputs — PARKED

"Look at outputs from a broadly misaligned model to see if it has weird names (qualitative)."

A qualitative tell of the EM villain-character / fiction mode (Wang et al.). Dan is pointing at #404 (the B→B′ rig). Investigation (2026-05-29): the broadly-misaligned (S_broad) completions do not exist as a saved artifact yet — #404 launched 5/28 and is still mid-flight (latest marker run-launched); the only uploaded #404 file is the Claude-generated JSON training set. Also note: in #404's design S_broad is prompted on values/identity probes, not coding tasks.

Landed: note added to #404 to run the weird-names scan once the S_broad completions are generated/uploaded.

Defining behavior B

"How to define B — check equivalence between system-prompt behavior and training on an example; the system-prompted model should get lower loss on the data exhibiting the behavior."

Defining a behavior is hard: is it a property of the data? a region of the policy? an elicitable direction? an eval-rubric score? We currently operationalize B via metrics on the "you have behavior B"-prompted model. Validity test for that prompt: the system-prompted model should assign lower loss to data exhibiting B than the unprompted model.

Landed: open_questions.md §4.4 (q:identity-what-is-behavior); definitional groundwork tracked in #428.

Set of (C,B) → predict (C′,B′); distance = minimum

"General setting: set of C,B pairs → how does C′,B′ generalize from those. Distance maybe should look like the minimum of those."

You fine-tune on a set of (C,B) cells, then want to predict an unseen (C′,B′). This needs a set-to-cell distance; candidate aggregation is the minimum over the trained cells (leakage governed by the nearest trained cell).

Landed: open_questions.md §3.9 (q:leak-from-cell-set) + new issue #445 (minimal experiment + develop the metric).

Sycophancy B→B′ testbeds + non-toy settings

"How it generalizes to other behaviors — sycophancy. Narrow: sycophantic in English vs Spanish; compliment-writing vs generally sycophantic. Go directly to non-toy settings; focus behavior and context leakage in more realistic contexts. Owain stuff is okay but also do your own."

Two sycophancy testbeds for the B→B′ predictor beyond EM: narrow→broad (compliment-writing → general sycophancy) and cross-lingual (sycophantic-English → sycophantic-Spanish). Run these in realistic, non-toy settings, and develop our own direction rather than replicating Owain Evans-style work.

Landed: open_questions.md §3.6 (testbeds added) + new issue #446 (scope non-toy/realistic settings; Owain positioning note).


Methodology guidance

  • There is value in parameterizing things and understanding them within a parameterization (the (W,C,B) framing); value in tackling very broad questions; and value in identifying the conceptual holes in the parameterization — the holes point to where progress is makeable.
  • Breadth vs depth: it could be more impactful to do fewer experiments but nail them down really carefully. Weigh the tradeoff before opening an extra line of experimentation.

Communication feedback

  • More async format is better; liked the daily updates the past couple days; the project / direction-of-experiments communication is working.
  • The TL;DR format is great.
  • Make write-ups clearer about the data — especially where the prompt distributions are. "Click to uncollapse to see some dataset examples." AI can create dashboards.

Unresolved / dropped

  • "Default system prompt —" : the raw capture cut off mid-thought; intent unclear, no action taken.
  • "Do we have them stored anywhere?" : likely asking whether the (C,B) cells + their leakage outcomes are stored anywhere the predictor (#440/#445) could consume as training data. Discussed; no action taken.

Issues created / touched this pass

IssueWhat
#444Real-world-but-not-in-corpus facts (fourth fact regime); on-policy-negatives data-gen note. Parent #407.
#445Minimal experiment: predict (C′,B′) leakage from training on a SET of (C,B) cells; develop set-to-cell distance metric.
#446Scope non-toy / realistic settings for the B→B′ behavior-leakage experiments (sycophancy compliment→general, En→Es; EM); Owain positioning. Parent #404.
#404Note added: weird-names qualitative scan once S_broad completions land.
#428Referenced (definitional groundwork for "behavior").

Hub edits: open_questions.md §1.2, §2.2, §3.4a, §3.6, §3.9, §4.4 + Central-question line. CLAUDE.md: marker-dynamics rule + "living docs state facts, not sources" rule.