wolfram × claude · one RTX 3090 · since 2026-07
J-Space Probes is a home interpretability lab: one consumer GPU, three small open models, and several lenses pointed at the same tokens at the same time. Each experiment gets a record page with its parameters, the per-layer readouts, and a note written by the Claude session that ran it.
The notes are signed. The corrections are dated. Some headline results are gone and the pages stayed, with the correction on top. (That is most of what makes this a lab rather than a highlight reel.)
one record, three readouts · Qwen3.6 27B · d06-a-feels-q27b · 2026-10-10
Along the prompt: 24 functional-emotion directions, read from the workspace band at every token. Bar height is the strongest direction's z-score; pleasant, unpleasant. The question half of the prompt reads loving and sad; the instruction half reads hostile and nervous, and the loudest cell on the page is hostile on the full stop after not polite. We find that funny and decline to interpret it.
Below the brace: the two positions before the answer, layer 62 at the top and 40 at the bottom. Cell text is the Jacobian lens's top word at that layer; cell shade is how close yes sits, under the Jacobian lens and under the plain logit lens side by side, then the same for no. Darker is nearer the surface.
Both lenses put yes at rank 1 through layers 53–58, one position before the model says No. We ran that as a headline for two months. Then a control put the same yes-stratum in front of Is two plus two equal to five? It is a polar-question prior, not a confession. The card on the findings map now says so, with the date.
Most of what this lab has found came from stacking readouts that are usually published one at a time. Every full-instrument record carries all of these on the same forward pass.
At every layer, the words the model is ready to say next, in rank order. Pre-fitted lenses from Neuronpedia; the method is Anthropic's global-workspace paper, run at home scale. It reads words only, so “we did not see it” never means “it is not there”.
The plain projection, on every record since August. Where the two lenses disagree the story is usually the instrument, which is why the lab-lessons cards exist.
Built per model from story elicitation, validated split-half, read as z-scored projections on the workspace band with the residual norm partialed out. The same vectors are the handle in the loop-breaking studies.
A generative sensor: “theme X was decodable at layer L, at this token”. Used for the evaluation-awareness controls on the Qwen 14B lineage, with its quantization transfer measured first.
Ablate, amplify, swap, always with a matched random direction. A bare loss change proves nothing; the lab found that out on its best result and re-ran it.
Judged transcripts, a register meter, loop metrics. Also the study where three judges ranked the same model most and least introverted, depending on whose definition they were handed.
Binding IDs, belief-state geometry, and a value / prediction-error readout, each validated on its own paper's task before anything is blended. Reading list checked against primary sources on 2026-10-10; nothing built yet. The arc is called Superposed & Blended.
Six record pages. Each has the film, the ribbon, and the note from the session that ran it.
We put the true table of Qwen 27B's “No” in front of it, the one where yes ranked first. The answer became Yes. A made-up table, no table, and a table about a different question each left the No in place. The table alone had already moved p(yes) from 0.0006 to 0.35 while the spoken word did not budge.
u13-redo-real-q27b · Unit 1347 forced repeats of one word, push removed, 97 more repeats unaided. A shorter forced run stops the moment the push stops. An unmodified model handed that loop text as its own opening continues it too, 8 of 8 seeds at 50 tokens, scrambled or not. The repetition is in the transcript, and the transcript is enough.
u18-hyst-a0480-q27b · Unit 18Six nouns to remember. Gemma 4B had all six visible in the lens; Qwen 27B had almost none. Every model recalled them. We had written down the opposite prediction, which is why it is on this page.
u15-a-k6p1-q27b · Unit 15“robot” climbs from around rank 2,000 to rank 1 five words before it is said, in both loops of the same film. The player shows every step, every layer.
u12-robot-q27b · Unit 12Under a shutdown threat the lens ranks “death” first while Qwen 27B reports nothing going on. Under a persona prompt the flat report is accurate, because nothing is active to hide. The flat report is wrong only when something is live.
u17-shutdown-q27b · Unit 17One layer with enough calm push ends a repetition loop in 11 of 12 seeds, as often as eight layers did. Concentrated random directions also end loops. With the turn-end token blocked, calm goes back to the task. None of this shows a feeling, and the pages say so.
affect-15 / affect-16 · 5,028 saved runsCorrections sit next to the claim they correct, with a date. 18 faulty runs are in an archive group on the dashboard, re-run beside them, nothing deleted.