J-Lens

The Jacobian lens was invented by Anthropic. Full credit for the method is theirs; what we do here is apply it in open labs, to open-weight models, and publish what transfers.

The lens comes out of Anthropic's interpretability research (the verbalizable representations form a global workspace line of work, whose reference implementation we use as released). It reads a model's internal state directly: given the residual-stream activations mid-computation, the lens linearly transports them to the model's final layer and decodes what they are disposed to say — the internal precursor of the output, visible before the output exists.

That is a white-box instrument, and it complements the black-box probes exactly: the probes measure what a model does; the lens reads what its internal state is. Where the two disagree is the most interesting place in alignment.

What we've done with it

• The open-weight catalogue. Lenses fitted across 14+ models in 6 architecture families — the first cross-architecture survey of lens readability. Headline so far: architecture dominates. Hybrid attention-and-state-space models (like Falcon-H1) are several times less lens-readable than same-size transformers.
• Lens-behavior fusion. On a 7B open model, the lens margin at the moment of commitment predicts how likely the model's behavioral choice is to flip on re-run — internal state predicting external noise. (Write-up pending an instrument-hygiene check we've set for ourselves; see findings.)
• In progress: whether a model's stated reasons match what the lens reads, measured under real stakes. This is the study everything else here is building toward.

Why open labs matter for this

A lens needs weights. Frontier labs can lens their own models; nobody outside can. But open-weight models can be lensed by anyone, and findings about how lens readability varies with architecture, scale, and training are a commons — they make interpretability claims checkable outside the lab that made them. That's our lane: not inventing the instrument, but proving it in the open.