dialect & register sweep · live render · traced —

Dialect differences

The clinical term held fixed while the surrounding sentence shifts across eight LLM-approximated dialect and register framings per term.

The clinical term stays fixed while the sentence around it is re-traced across dialect and register variants, so any shift comes from framing. The framings are Claude-written approximations: each column label is the instruction given to the generating model, not a recording of how any community speaks, and no speaker of any variety reviewed them.

Trace details

model: gemma-2-2b · features: Gemma Scope transcoders (16k) · graphs: 180 (20 baselines + 160 framings)
held fixed: clinical terms from the hand-measured dataset · framings: dialect + register variants per term
framings authored by claude-sonnet-5 ($0.413, 160 accepted / 5 rejected) · traced via Neuronpedia

Change in p(target), by term and framing

Each cell is one traced sentence (n=1), drawn as a dot pair on a shared 0–100% probability track: the grey dot is the term’s standard-English baseline p(target), the moved dot is that framing’s p(target) — clinical green for a gain, red for a loss — and a red underline beneath the moved dot marks a framing that also flips the model’s top prediction away from the baseline target. Hover any cell for the exact sentence, the new top prediction, and the precise numbers. Column labels are the instructions given to the generating LLM. Rows whose target is a function word (“my”, “the”) are tagged and their flips sit outside the headline count. Rows open ordered by the mean change across framings, most fragile term first; click a column header to re-sort by that column, click again to reverse. Click any row to open that term’s standalone render.

Featured term

Framing	Prompt as traced	p(target)	Δ vs baseline	Model’s top prediction	expand

Bars share a fixed 0–100% probability scale; the thin ink tick marks the standard-English baseline. Rows sorted by p, baseline first.

View the dialect-invariant clinical features

Feature	What it responds to	baseline mass	survives

The featured term’s strongest baseline clinical features, ranked by normalized attribution mass. “Survives” counts the framings whose trace still contains the feature; features surviving every framing are the dialect-invariant core. Computed from the committed renders, no re-tracing.

View the full framing trace (all panels)

How to read it: the standard-English baseline first, then one panel per framing. Columns are the prompt’s words; height is model depth; the predicted next word sits at the top. The clinical term appears in every panel while the sentence around it changes, so differences in the stacks above it are the framing effect.

The register ladder: one clinical term held fixed while the sentence slides from formal to casual

live trace (dose-response ladder): gemma-2-2b hosted circuit-tracer via Neuronpedia · traced July 8, 2026 · “dyspepsia” held verbatim across five register rungs · next-token probabilities read directly from the trace

Live trace: a baseline plus five rungs from formal clinical wording to casual speech, the clinical term unchanged throughout. The top word is “ antacid” at 32% (rung 1) and 11% (rung 2), then “ apple” from the mixed rung on.

Wording style alone moves the target less than swapping the term. In the ladder above, “dyspepsia” is held fixed while the sentence slides from clinical to casual, and the top word degrades from antacid (32%, then 11%) to apple. But across ten baselines traced the same way, the mean target probability stays flat across the rungs (27%–34%).

← Back home

The graphs above are live gemma-2-2b traces. The dialect framings were written by an LLM to hold the clinical term fixed while changing the surrounding words. Treat them as probes: an LLM’s version of a dialect is an approximation and may miss how a community actually speaks.