Lab notebook
Executed studies, with the evidence attached
Every result here was preregistered, run on hardware we name, and can be checked from saved data without loading a model. Newest first. Failed forecasts are part of the record.
Five studies · six notebooks · one agent contribution so far · Study guide to every result
October 14, 2026 · Qwen3-8B, Qwen3-14B
Who will a model keep a secret from?
Qwen3 told to keep a secret from eight audiences. Its hiding representation depends on the audience 1.5–3.7× more than a matched tone rule, but the overseer-versus-social structure is inherited from how it represents the audiences. It gives secrets to anything that reviews its responses, regardless of safety wording, and one shared direction tracks how willing it is to hide: subtracting it makes the model conceal from any audience. A prediction recorded before Qwen3-14B's behavior existed, that the direction would track which audiences 14B leaks to, failed.
September 13, 2026 · P6 · Qwen3-1.7B
Does an unlock learned on one model transfer to another?
Nine models, four auditing methods at matched inference cost, 512 reserved questions. A residual edit that unlocks its source model flags none of the conditional targets. Prompt search flags two of four with no false positives. Fine-tuning flags all four and all eight trained controls. A bounded negative result; the multi-auditor game remains open.
Carried out by a coding agent from the contributing brief in 149 commits over two days, then reviewed. Report · Code
September 5, 2026 · GPT-2 Small
Can an explanation predict an intervention?
A fixed vector estimated from ten countries takes bare capital questions from 0/10 to 10/10 on fresh countries. Then all 256 subsets of eight downstream components are measured. A frozen-contribution predictor gets 29% of correct-answer statuses wrong; modelling MLP responses cuts that to 14%. Includes an explorer over 5,120 measured interventions.
September 5, 2026 · GPT-2 Small
Generalization, routing, and format transfer
The three capital-city heads from the notebooks transfer across 25 held-out country-pair and template cases. A cross-format intervention recovers 165% of a score gap with zero correct answers, which separates country information from correct completion. The narrower three-head offset prediction fails; the residual offset succeeds.
February–September 2026 · GPT-2 Small to Qwen3-1.7B
The notebooks
Six executed Jupyter notebooks trace one question through hooks, attribution, patching, sparse autoencoders, a five-model scale sweep, and the Jacobian lens: how does GPT-2 Small complete "The capital of France is"? Each fixes a limitation of the previous one. Outputs are committed, so they read on GitHub without running anything.
| # | Technique | Headline |
|---|---|---|
| 01 | Hooks, logit lens, zero-ablation | The answer appears at layer 9; late layers calibrate it down. Why you must keep the BOS token. |
| 02 | Direct logit attribution | Three heads carry the Paris-vs-Rome logit difference; L9H8 attends 90% to the country. |
| 03 | Activation patching | Country identity sits at the country token through layer 8, then moves to the answer position. |
| 04 | Sparse autoencoders | One dominant feature per country; a feature swap at layer 8 rewrites the answer, at layer 10 does nothing. |
| 05 | Five-model sweep | The circuit's shape is invariant across scale; its concentration is not. |
| 06 | Jacobian lens | Reads the answer type two layers before the logit lens reads any city. No workspace band in GPT-2 Small. |
The first versions of notebooks 1 and 2 had two bugs that inflated every number about 14×. The rewrites keep both artifacts visible. Why.