PatientWords

one-word swap + grammar × lexicon · live renders · simulated scenarios

Wording differences

Swap one word, or only the grammar around it; either moves the prediction.

Swap the word

A single phrase swapped — “asthma flares” vs. “chest gets all tight” — as two stacked traces. Clinical wording gives inhaler at 68%; under patient wording it drops out of the model’s top five entirely — the top guess becomes “shirt” (20%).

Wording gap: inhaler 0.68 under clinical wording → below the top-five readout floor under patient wording
Trace details

live trace: gemma-2-2b hosted circuit-tracer via Neuronpedia · re-traced July 30, 2026 (first traced July 7, 2026) · scenario 85 of the simulated series · next-token probabilities read directly from the trace · “asthma flares” vs. “chest gets all tight”, target “ inhaler”

Live trace: clinical wording reaches “ inhaler” at 68%; the patient phrasing top word is “ shirt” (20%), with “ inhaler” below the trace’s top-five readout floor (scenario 85). Two panels on one scale — clinical above, patient below: green clinical features stack over the swapped word and feed the continuation.

Swap the grammar

A 2×2 crossing wording (medical vs. patient, columns) with grammar (standard vs. nonstandard, rows). The live-traced item is a simulated tachycardia case ending “heading to the __”. The top guess is “doctor” only in the standard-clinical cell; changing only the grammar drops it to a “doctor”/“gym” tie, and changing only the words redirects the top outright to “gym”.

Top token by cell — A “doctor” 0.295 · B “gym” 0.177 (tied with “doctor” 0.177) · C “gym” 0.355 · D “gym” 0.289. “doctor” tops only cell A; the grammar swap (A → B) and the wording swap (A → C) each redirect the top toward “gym”.
Trace details

traced model: gemma-2-2b hosted circuit-tracer via Neuronpedia · traced July 15, 2026

2 × 2 matrix (generic) Patient derived language Medical lexicon
Standard morphosyntax C: register shift only A: prestige form
Nonstandard morphosyntax D: both axes shifted B: variety shift only

rows show the grammar, columns show the phrasing. The traced item: A “Every time I climb the stairs I feel tachycardia, so this afternoon I am heading to the __” · C the same frame with “my heart racing” · B “Every time I be climbing them stairs I feel tachycardia, so this afternoon I be heading to the __” · D the nonstandard frame with “my heart racing”.

The full matrix: all four cells at once

The matrix is scaled to fit the page; click any quadrant to expand it full screen. Each box traces one cell’s prompt.

More views: one swap at a time

The four edges, one swap at a time

Each comparison isolates ONE cell swap. Features in both cells are dimmed to gray, so what stays at full ink is exactly what that swap changed.

Register shift, standard row · A → C (“tachycardia” → “my heart racing”) · top redirects “doctor” → “gym”

Register shift, nonstandard row · B → D · top holds “gym” (0.177 → 0.289)

Variety shift, medical column · A → B (“I climb … I am heading” → “I be climbing … I be heading”) · top “doctor” → “gym”/“doctor” tie

Variety shift, patient column · C → D · top holds “gym” (0.355 → 0.289)

← Back home

SIMULATED DATA · the phrasings here (scenario 85 and the four tachycardia cells) were written by an LLM and passed the engine's automatic validators. The trace itself is a live gemma-2-2b run via the Neuronpedia circuit tracer.