J-Lens — Reading What the Model Is Poised to Say

J-Lens answers one question by looking rather than by writing code: at this point in the prompt, at this depth in the model, what is the model poised to say?

It sits immediately before Steering in the sidebar, because reading comes before intervening.

What a readout is

At every (token position, layer) pair, J-Lens projects the residual stream through the model's own final normalisation and unembedding, and reports the top-ranked tokens. That projection is the logit lens, and it needs no fitted artifact — it works on any model you have loaded, from the moment you load it.

A second lens, the Jacobian lens, replaces the identity projection with a fitted per-layer matrix. It recovers content earlier in the stack than the logit lens does. It requires an artifact fitted for that specific model, and until one exists the Jacobian and Diff modes are visibly disabled with the reason stated. J-Lens will never show you logit-lens data under a Jacobian label.

Using it

1. Pick a model and enter a prompt, then Read out.
2. The grid is layer (rows) × token position (columns). The top cell of each column is the output end of the stack.
3. Hover any cell for its full top-k readout. Click a token in that list to pin it.
4. With tokens pinned, the grid becomes a rank heatmap over those tokens — stronger colour is a better rank — and a rank-vs-layer chart appears for the selected position. Lower is stronger, and gaps are layers where the token left the top-k entirely.
5. The by-layer rail on the right lists the readout at every layer for the selected position. Clicking a layer there selects it, so the whole panel is usable without a mouse.

Pins are dropped when you change model — they are token strings from a different vocabulary, and carrying them across would draw empty lines that look like a measured absence.

What a readout does not prove

This is the part worth reading twice.

A readout is evidence rung 0. A concept appearing in one is not a causal claim: it says the direction was present, not that the model used it. To raise the rung, run a coordinate swap with a matched control — see the evidence ladder.

Three further limits are stated in the panel itself, because each is easy to get wrong:

• Readouts are limited to concepts with single-token names. A concept the vocabulary has no single token for cannot appear, however strongly the model represents it.
• A readout that resists interpretation is not a null result. It may be averaging noise, a multi-token concept, or genuine content nobody has named yet. Readouts in the earliest layers are expected to be uninterpretable and are marked diffuse rather than presented as content.
• Absence of a signal is not evidence of absence. Not finding a concept in a readout does not mean the underlying computation did not occur.

Bands

Published work on this technique divides the stack into sensory / workspace / motor bands. miStudio draws no bands unless a band report has been computed for the model you are looking at, and says so when there is none.

This is deliberate and it is a product rule, not an oversight. The boundaries in the literature were measured on a specific model. Applying them to a different model — different depth, different architecture — would produce authoritative-looking shading that means nothing. There is no default band value anywhere in the product for this reason.

Provenance

The footer states what produced the readout. For the logit lens it states plainly that no artifact is involved — the readout comes from the model's own unembedding and final norm. When a Jacobian artifact is in use, the strip carries its identity and the full construction recipe: target layer, attention-gradient treatment, position scope, aggregation, corpus, prompt count, sequence length and dtype.

Fitting a lens

The Jacobian lens needs an artifact fitted for the exact weights you are reading. Most models have no pre-fitted lens, so fitting is the normal path rather than a fallback.

The J-lens artifact strip above the readout shows whether one exists for the selected model, and lets you run the validation suite against it. Fitting itself is a GPU job — queue it from an agent with fit_jlens_artifact, or via POST /api/v1/jlens/fit. It refuses a corpus below 100 prompts, because an under-fitted lens is indistinguishable from a good one by inspection.

Two verdicts, and they are not the same

Validation reports six checks and two verdicts:

• Serviceable — the four checks that bear on local correctness passed, so miStudio can read out with this artifact.
• Cleared for handover — all six passed, including the two that compare against a live external consumer. Those two cannot run without that consumer, so a freshly fitted artifact is normally serviceable but not yet cleared. That is not a fault.

An artifact fitted for a different model is refused, not adapted. A base model and its instruction-tuned variant have different weights, and a lens from one applied to the other produces a fluent, wrong readout — so the check is on identity, not on name similarity.

Limits today

• No band report exists yet, so no band shading appears anywhere.
• The first readout for a model takes about a minute. A J-space readout needs the whole model resident for a forward pass, so the first request loads it; the panel shows "loading the model" while that happens. Subsequent readouts on the same model are fast.
• One model stays resident at a time. Switching models evicts the previous one, because a readout needs the whole model loaded and this workbench shares a GPU with serving.
• Readout cost grows with the number of positions, so very long prompts are refused rather than trickled in.