Level 5 · Flagship researchFlagship
Track C · Interpretability
Can Anthropic’s newest interpretability method be made open, efficient and less limited, and does it catch implanted misaligned dispositions in open models?
Build an open, efficient implementation of the J-lens (layer-wise averaged Jacobians onto vocabulary tokens) on an ~8B open reasoning model, reproduce the paper’s findings, then fix its multi-token limitation and test it on implanted dispositions.
Anthropic’s 2026 ‘global workspace’ paper introduced the J-lens and found a misaligned disposition implanted by training leaves a signature at the start of responses even when behaviour looks normal. It ran on Claude models with no code release mentioned, and lists limitations: single-token concepts only, an arbitrary sparsity choice and possibly degenerate early-layer readouts.
01Read the paper and list its limitationsM1 · week 8
Note the method and the stated limitations you will address.
02Implement it efficiently on an ~8B open reasoning modelM1 · week 8
Jacobian compute cost is the main risk – making it cheap is itself part of the contribution.
03Reproduce 2–3 qualitative findings and releaseM1 · week 8
Release the open implementation and reproduction as the week-8 milestone.
04Fix the multi-token limitation
Extend the readout beyond single-token concepts.
05Pre-registered disposition detection
Test whether it detects dispositions implanted in organisms (pairs naturally with the auditing or eval-awareness flagships). Write predictions first.
06Compare against linear probes
Is the J-lens better than a simple baseline?
07Release
Library, reproduction and extension results.