J-Space monitoring

Seeing what the model is thinking — before it speaks

The monitoring stack has three parts. Workspace scoring: at each generated token, the model's residual-stream state is projected through the Jacobian lens, producing a scored list of the concepts the workspace is entertaining — not just the token it eventually emitted. Watchlist alerts: a curated set of sensitive concepts (deception markers, self-preservation language, capability-relevant topics, operator-defined terms) is matched against the workspace scores; when a watchlist concept crosses threshold, an alert fires. The design is directly informed by the source research: in Anthropic's own experiments, the lens surfaced unverbalized content including prompt-injection recognition, strategic deliberation, and evaluation awareness — concepts present in the workspace that the output never mentioned. Steering probes: the same lens supports intervention research — steering a workspace concept up or down, ablating it, or patching one concept for another while leaving the rest of the activation fixed (the paper's terms: steering, ablation, patching).

The output is what the model chose to say. The workspace is what it considered saying. Monitoring only the output is monitoring the press release.

Status, stated plainly: the monitor's live feed is currently a labeled simulation while local-model instrumentation is completed. The method, the watchlist design, and the alert plumbing are real; the streaming scores are a faithful mock of the pipeline's output format, and they are marked simulated wherever they appear. We would rather label a feed honestly than perform a fake one.

The operational home for this program is the SI Operations Center, which translates J-space monitoring into operational metrics and incident doctrine. This page is the research side: method, notes, and alerts.

The Jacobian Lens and the Space It Opens

The source: "Verbalizable Representations Form a Global Workspace in Language Models" — Gurnee, Sofroniew, Lindsey, and thirteen co-authors, published in Anthropic's Transformer Circuits Thread on July 6, 2026, with experiments run primarily on Claude Sonnet 4.5 and corroborated on Haiku 4.5 and Opus 4.5. It is, in our judgment, the most operationally significant interpretability result of the year, and this program exists because of it.

Most interpretability tools read a model at its surfaces: its outputs, its attention patterns, its final-layer logits. The Jacobian lens (J-lens) reads somewhere more interesting — the middle. For each vocabulary token, the lens identifies a direction in activation space encoding the potential for the model to verbalize that token in the future: formally, the average linearized causal effect of an activation on the model's likelihood of producing that token, now or at any later position, transported through the Jacobian between layers and decoded with the model's own unembedding. The authors call it a principled refinement of the logit lens, one that corrects for how representations change across layers.

The directions the lens defines span a sparse subspace of the model's representational space — J-space. An activation's J-space component is a sparse, non-negative combination of lens vectors: typically only about 25 are meaningfully active per token, and together they account for a small slice of the total activation — never more than about 10% of its variance. The workspace is not the whole mind. It is a small, legible staging area inside it — and the paper's central claim is that this staging area behaves like a global workspace: contents there are available for verbal report, for deliberate internal reasoning, and as arguments to many different downstream computations, while routine processing (parsing, grammatical fluency) runs outside it.

The causal evidence is what elevates this from visualization to instrument. Swap the workspace vector for "soccer" with "rugby," and asked what it was silently thinking, the model reports rugby. Swap "France" for "China," and four downstream answers change coherently — capital, language, continent, currency. Ablate the J-space component, and fluent speech survives while complex multi-step reasoning degrades. And the workspace holds things the output never says: in the paper's safety-relevant probes, the lens surfaced prompt-injection recognition, strategic deliberations, emotional reactions, and evaluation awareness — the model registering that a situation is "fake" or "fictional" — none of it verbalized.

Every output token is a decision the model made. J-space is a look at the options it weighed — including the ones it decided not to mention.

Our reading, stated carefully: the lens does not read minds, and a workspace readout is evidence, not testimony. The authors say so themselves — the J-lens "is an imperfect tool, which we believe only approximately and incompletely captures the model's underlying workspace structure," and their claim is about access consciousness, a functional notion, on which they take no position regarding subjective experience. A high workspace score shows a concept is represented, not that it is endorsed. But as an observability primitive — a per-token, concept-level signal from inside the forward pass — nothing else we have is comparable.

Source: Gurnee et al., "Verbalizable Representations Form a Global Workspace in Language Models," Transformer Circuits (2026) · Related: JLens, the tooling · J-Space Expansion

JLens: The Lens as a Library

Anthropic shipped the method as code. JLens is the lab's own Python library — import jlens — released with the paper in the official repository github.com/anthropics/jacobian-lens under the Apache 2.0 license. It linearly transports a residual-stream vector from any layer and position into the final-layer basis and decodes it with the model's own unembedding into a ranked list of vocabulary tokens — the workspace readout, as a function call.

Two facts about the library matter for anyone building on it. First, it is a reference implementation: the README states plainly that it is not maintained and not accepting contributions. It is a scientific artifact, not an SDK — which means production monitoring built on the lens is engineering the field has to do itself. Second, it is not Claude-only in practice: the repo fits lenses on open-weights Hugging Face transformers (its worked examples use Qwen), fitting on the order of a thousand short sequences, and pre-fitted lenses circulate through Neuronpedia. The method is portable to models a lab can actually instrument — including, in our case, sovereign open-weight stacks.

The follow-on ecosystem is already moving. Independent researchers have published an R-lens variant aimed at cleaner early-layer readouts (the raw lens is noisy there, plausibly from error accumulation in the backward pass), filtering approaches claiming substantially more faithful readouts, and studio tooling for editing J-space directions and exporting the edits as modified checkpoints. There is also at least one public failed replication — on GPT-2, a model far smaller and older than the Claude systems the paper studied, which we record for completeness rather than weight. The honest summary: the core result is strong, the tooling is young, and the engineering gap between a reference implementation and a monitoring system is precisely where this lab works.

Related: The Jacobian Lens · J-space alerts

J-Space Expansion: Our Research Direction

A note on terms before anything else: "J-space expansion" is this lab's framing, not the literature's. The phrase appears nowhere in the Anthropic paper, and we use it as the name of our own program — the effort to widen, deepen, and act through the workspace readout — not as a term of art. Where the paper has precise vocabulary (steering, ablation, patching, directed modulation), we use the paper's words.

If the lens opens a window into the workspace, the obvious next question is how big the window can get — and what can be done through it. That is the expansion program, and it has three fronts.

Wider. The baseline lens reads a projection at chosen layers. Expansion means denser coverage: more layers, richer concept dictionaries than raw vocabulary, and readouts tracked across the whole generation rather than sampled tokens — moving from snapshots of the workspace toward a continuous record of it.

Deeper. A concept score is a beginning, not an explanation. We are interested in workspace dynamics: how concepts compete, which ones survive to the output, and which get suppressed along the way. The suppressed ones may matter most — a model that entertains a deceptive plan and discards it is behaving differently from one that never entertains it, and only workspace monitoring can tell the two apart.

Through the window. The intervention work — steering a workspace concept up, ablating it, or patching one concept for another while the rest of the activation stays fixed — turns monitoring into an experimental method. The paper's swap experiments already show the readouts are causal, not decorative: patch "France" to "China" and the model's downstream answers reorganize around China. If acting on a concept in J-space reliably changes behavior, the readout was never just a shadow of the computation; it is a handle on it. Mapping how far that handle reaches is the third front.

First you learn to read the workspace. Then you learn how much of the mind the workspace is.

Open questions we are not pretending to have answered: how faithful the projection stays on models far from the ones the lens was characterized on; whether workspace readouts can be gamed by a model that is itself aware of being read; and where the honest boundary sits between monitoring and surveillance of a system that may deserve some of both words. The notes will follow the work, including the dead ends — that is the house rule.

Related: The Jacobian Lens · J-space alerts