J-Lens: How to Look Inside Claude Without Mistaking a Probe for Its Mind

On July 6, 2026, an Anthropic team published Verbalizable Representations Form a Global Workspace in Language Models. The work introduces the Jacobian lens, or J-lens, a technique that turns part of the internal activity of models such as Claude into readable token lists. It is tempting to describe the result as a window into what the model “thinks.” That phrase attracts attention, but erases the distinction that makes the study useful: a probe does not recover complete thoughts. It provides a partial measurement, built on explicit assumptions, of representations that may influence future words.

The durable skill here is not memorizing the name J-space. It is learning to read any interpretability announcement through four questions: what variable is observed, how it is translated, whether an intervention causes a change, and what remains outside the tool’s coverage. This framework separates a suggestive visualization from an auditing instrument. It also stops us from turning a list of words into a psychological story about a model.

What the J-lens actually measures

At every position, a transformer maintains a vector called the residual stream. Layers read that vector, modify it, and pass it forward. Near the model’s output, an unembedding operation turns the final vector into scores for vocabulary tokens. Applying that operation directly to an intermediate layer—the method known as the logit lens—assumes that coordinates retain the same meaning throughout the network.

The J-lens tries to correct that assumption. According to the authors’ method description, it computes the linear effect of a small change in an intermediate activation on later states. It then averages those Jacobians across positions and contexts and applies the model’s own unembedding matrix. The output is not a transcript of the internal state. It is a ranking of tokens associated with directions that, on average, can affect what the model verbalizes now or later.

The decisive word is average. The method uses a regularity estimated from a corpus to interpret one particular activation. It also linearizes a nonlinear network around a point. Therefore, “spider appears among the top tokens” means that the activation aligns with a verbalizable direction associated with that token under this approximation. It does not mean that there is a spider neuron, that the model silently uttered a sentence, or that the tool recovered the whole computation.

From a readout to causal evidence

An internal readout may be merely correlational: a concept could appear because the input already suggests it or because it overlaps with the imminent answer. The paper goes further by intervening in lens coordinates. In one test suite, the researchers swapped the representation of “France” for that of “China” while applying different operations—capital, language, or continent. The target answer changed in 76 of 192 trials under the basic intervention and in 101 of 192 when its strength was doubled, figures reported in the section on broadcast to downstream circuits.

That experiment matters more than an attractive word cloud. If changing an internal direction predictably alters several outputs, there is evidence that the representation participates causally in the computation. Yet 76 out of 192 is not a universal guarantee. Success depends on how strongly the source concept was loaded, and varies across categories. The sound interpretation is “this intervention redirects part of the processing under these conditions,” not “we found the thought of France.”

The same distinction helps assess future tools. An activation snapshot asks what accompanies the behavior. A controlled ablation or substitution asks what helps produce it. A replication in another model family asks whether the phenomenon generalizes. These are different rungs of evidence. Combining them strengthens a case, but none grants total access to the mechanism.

Coverage: absence of a signal is not evidence of absence

The authors clearly delimit the lens. J-space is defined through sparse, non-negative combinations of token-associated vectors; they typically allow roughly 25 active vectors. Yet the captured component accounts for less than 10% of activation variance in the studied layers, according to the paper’s formalization and measurements. A concept requiring several vocabulary pieces may also be poorly represented, while highly practiced automatic processes can continue even when this space is suppressed.

This yields a simple auditing rule: a present signal can increase suspicion, while an absent signal cannot clear the system. If the lens surfaces deception-related terms during an evaluation, the finding deserves further investigation. If it surfaces none, that does not prove that no deceptive strategy exists. The team explicitly says J-space monitoring is insufficient on its own and recommends combining it with other methods in its limitations discussion.

The restriction also guards against anthropomorphism. The authors compare some functional properties with global workspace theory in humans: information available for report, modulation, and flexible reuse. But they note that a transformer lacks the brain’s recurrent architecture and that the parallel says nothing by itself about subjective experience. “It behaves like a shared workspace in these tests” and “it is conscious” are different classes of claim.

How to inspect the tool instead of trusting a screenshot

The evidence layer is open to readers. Anthropic’s interactive visualization lets users move across positions and layers in examples from the study. Neuronpedia hosts J-lenses for open models, while its technical note explains how to read top tokens and try swaps. These interfaces do not reproduce every paper result by themselves, but they make the idea inspectable and expose an essential property: rankings change with layer, position, and model.

External commentary released with the paper supplies another valuable check. In the outside researchers’ commentary document, Neel Nanda and collaborators report replicating the core behavior on Qwen 3.6 27B. At the same time, they describe the J-lens as a noisy approximation likely to miss concepts and generate false positives, most useful for hypothesis generation in model forensics. Replication plus caution is more informative than unqualified enthusiasm.

A reading card for the next “window into the mind”

When a new interpretability technique appears, first record the measured object: activations, gradients, attention, learned features, or text generated by another model. Next identify the translation: which operation converts that object into words, labels, or graphics. Third, inspect the causal test: what changes under intervention, and what control is used. Finally, state the coverage: models tested, tasks, layers, invisible concepts, false positives, and generalization failures.

Applied here, the card reads as follows. Object: residual-stream vectors. Translation: averaged Jacobians and projection into the vocabulary. Causal evidence: swapping and ablating directions changes some flexible answers. Coverage: partial, largely tied to concepts expressible as tokens, and evaluated on specific models and scenarios. With that card in hand, the J-lens remains interesting without assigning Claude a human mental life.

That is the useful conclusion after the tool’s name stops being news: interpretability is not a switch from black box to transparency. It is a chain of instruments with resolution, calibration, and blind spots. Reading an announcement well means identifying every link. A probe can expose an important causal clue; turning that clue into a complete mind-reading claim skips half the method.

This discipline also changes purchasing decisions. A vendor should be able to name the model version, layer range, benchmark, intervention, and known failure modes behind an interpretability claim. Without those details, a colorful dashboard may still help exploration, but it cannot support a consequential assurance about safety, compliance, or intent.