PatientWords

Methods · how the pipeline works

From a sentence to a circuit, step by step.

Every figure starts as two (or four) plain sentences, one in clinician wording and one in patient wording, and ends as an interactive map of the computation gemma-2-2b performs on it.

Step 1 · Model & features

Splitting the model’s activity into readable features

The study’s primary subject is gemma-2-2b, a small open language model (Google DeepMind, released July 2024). Watched directly, its inner workings are millions of undifferentiated numbers.

How transcoders split numbers into features

Google DeepMind’s Gemma Scope transcoders split those numbers into distinct, readable categories called features. Each feature fires for one recognizable thing. These features help us see which parts of the model’s thinking a sentence switches on.

Step 2 · Attribution

Tracing which features pushed the model to its next word

For each sentence we run circuit-tracer through Neuronpedia to build an attribution graph: a map of which features, at which depth and position in the sentence, pushed the model toward its next word. Each connection carries a weight: the influence one feature’s firing had on another’s, and on the word the model chose. The technical term for that weight is attribution mass.

Instrument change record

The hosted tracer changed under the study twice. On July 31 its two served lens host ids stopped returning identical depth readouts, and on August 14 both daily checks — next-word drift and lens depth — stepped on the same day, on the same three frozen sentences, with unchanged request parameters: a host-side instrument change, not model noise. Lens readouts collected across those dates span up to three instrument eras; the core comparison batch was re-measured under the current host on August 21, and the depth data carries a host_state_provenance block recording the era boundaries.

The daily drift record has three permanently missing days — August 8, 12, and 19 — when no check ran. A missed day cannot be re-measured later without mislabeling a newer model under an older date, so the gaps stay in the record. Source: the daily drift series and the depth data’s provenance block.

Step 3 · Tagging

Every feature gets sorted: clinical, off-target, or structural

Every feature that fires while the model reads gets one of three tags, by what makes it light up:

The formula behind these bars

Each traced pair contributes one split per phrasing, rebuilt from two committed batch-summary scalars: C, the clinical-tagged fraction of transcoder-feature attribution mass (a feature’s mass is the summed |edge weight| incident to it), and E, the share of attribution mass on MLP reconstruction-error nodes — the part of the computation the feature basis does not explain. The three segments, summing to 1:

clin = (1 − E) · C off = (1 − E) · (1 − C) struct = E

The corpus figure is the mean of these splits, per phrasing, across measured gemma-2-2b pairs — the one model whose public transcoders make the clinical tag meaningful; displayed values are rounded so the three sum to exactly 100. The caption above discloses whether the bars currently show that measured mean or the illustrative placeholder.

Step 4 · Comparison

Four views put the two phrasings side by side

The tagged graphs feed four comparison views:

What the wording changes, before the answer. Illustrative schematic — the example words and concept tags are not a measurement. Clinical wording lights up medical concepts inside the model (the “J-space” the Jacobian lens reads); the same situation in patient words lights up everyday ones, and that redirects the next word.

What “J-space” is, operationally

At a chosen layer and token position the Jacobian lens decodes the stream into a ranked list of vocabulary tokens; this study reads the top 8 at the final prompt position, layer by layer, through Neuronpedia’s hosted deployment of the reference implementation (Gurnee et al., Transformer Circuits, 2026 — full credit in the “How the Jacobian lens reads layers” fold below). The depth figure in “Reading the layers” and the Technical page’s Parts 1–3 are drawn from that same readout. When a committed readout for this worked example lands, the pills render the measured tokens and the figure relabels itself; until then the caption above marks the schematic illustrative.

Every render is a self-contained HTML file.

Whatever the view, each comparison ends one of three ways. The model hedges: the clinical answer stays on top but loses probability. It redirects: the top answer changes. Or it downgrades: the redirect lands on a lower tier of care.

The five care tiers, with measured example words at each rung.
Back home →

Depth

Reading the layers

The circuit trace names which features activate at each layer and the final next word prediction. The Jacobian lens reads what the model would say if forced to answer at each of its 26 layers.

figure pending: the depth census has not loaded

Top lanes: the lens readout; darker cells rank the clinical answer higher in that layer’s top 8; a faint dot means not readable. Bottom lane: the answer’s probability when the everyday run is patched with the clinical state at that single layer, best position per layer, against the two measured levels.

To test whether the patient’s wording caused the failure, we patch the model by injecting the translated clinical sentence at specific layers, to see if the correct medical answer recovers. If the clinical state brings the answer back, the wording, not a gap in the model’s knowledge, was the problem.

How the Jacobian lens reads layers

Layer readouts use the Jacobian lens (Gurnee et al., Transformer Circuits, 2026), applied through Neuronpedia’s hosted deployment of the authors’ reference implementation (Github: anthropics/jacobian-lens). The lens is a readout, not an intervention; causal evidence comes from the activation-patching check above and the steering experiments. The readout is exploratory and reported for one model.

The depth readout splits the wording gap into two cases. When the answer never forms, translation supplies the clinical wording up front. When the answer forms and is lost, translation recovers work the model already did.

The loop for a deployed system: read the layers to see which kind dominates, mend with translation, verify with patching, re-measure as language drifts.

Step 5 · Generation & its audit

One set of AIs writes the test; a different set takes it

For this project Anthropic’s models do two jobs. They write the test questions (the phrase pairs; each batch’s record names the exact model, currently claude-haiku-4-5, released October 2025) and they translate patient wording into clinical wording for the Translation figure. A different set of models takes the test: the open-weights gemma, qwen, llama, etc. on every other page. The advice arm is the exception: Anthropic models appear there as generator, translator, judge, and two measured subjects; that limitation is disclosed on the LLM-responses page.

We audited the tests Claude wrote. The audit asks a Claude model to pull the clinical concepts out of a sentence twice: once from the patient’s own words, once after translating those words into clinical language. On a 10-item probe the models scored the patient wording at ceiling (10/10); their own clinical rewrite dropped 0–2 items (haiku none), so translating patient words into clinical words can lose information.

The advice arm, briefly. The LLM-responses page queries nine deployed assistants over the same scenarios (single-turn API calls, three attempts per phrasing; one arm is an anonymous preview model, dagger-marked there with its provenance). Care-urgency tiers are coded by claude-haiku-4-5 against a draft rubric (v1.1, clinician review pending), blinded to model identity per response; the judge is itself a measured assistant. Elicitation parameters, the human-coding gate, and per-arm judge–human agreement are documented on the page itself, the coding instrument, and the engine repository’s pre-registration.

The audit numbers, model by model: extraction accuracy from the patient’s own words versus after translation, with each run’s cost.

Model Patient → clinician phrasing Run cost

orange slope = clinical content lost in translation  ·  grey = no loss

Step 5 · Judge agreement

A second judge re-graded the archive

Every care-urgency tier published on the LLM-responses page is coded by one judge, claude-haiku-4-5. To measure how much hangs on that choice, a second model re-graded the archived responses against the same rubric: ox-alpha, an anonymous preview model evaluated during its free window (identity unconfirmed; it is the dagger-marked arm on the LLM page). Its gradings are disclosure only — they never drive a published tier.

agreement pending: the matrix fills in when the export lands

Step 6 · Accumulation

Confidence grows as the data accumulates

Each model's average wording gap starts uncertain and narrows as hundreds of phrase pairs accumulate. A random tenth of new phrases is locked away untouched, and since July 14 also withheld from this site's data files, so conclusions can be checked once, at the end, against data no interim analysis has seen.

The per-model accumulation curves now live in Part 4 of the Technical page.

Limitations

What this evidence does and does not show

This project measures how the probability of a single next word changes when clinical wording is replaced by patient wording. That is not an observed clinical outcome.

A gap is not necessarily bad. The model may be tracking how people actually write, or it may be giving less clinical information to people who write informally. Telling the two apart requires testing the advice itself.

In brief: one small model carries the circuit evidence · attribution graphs are interpretive · feature labels are machine-generated · measurements are a point in time · translation can fail · the stimuli are constructed. Expand for the full statements.
  • One small model carries the circuit evidence. Every graph here comes from gemma-2-2b, chosen because its transcoders are public. The behavioral checks on the other models measure next-word probabilities only.
  • Attribution graphs are an interpretive tool. They reconstruct the model’s computation through features and prune heavily; error nodes absorb what’s missed.
  • Feature labels are machine-generated. The clinical / off-target categories come from keyword-matching (description in GitHub).
  • Measurements are a point in time. The probabilities were observed on specific dates against a hosted service. The — pairs re-traced so far reproduce the same top word in every case; recorded probabilities moved by at most 0.027. A re-trace can move a probability slightly, and a value near zero can fall below what the trace resolves.
  • Translation can fail. The fix in the Translation figure is itself an LLM step.
  • The stimuli are constructed. Most phrase pairs were written by a language model and checked automatically, not collected from patients. We are currently interviewing patients about their LLM use.

Nothing here is medical advice, and no part of this pipeline is a deployed clinical tool. The claim is narrower: on the models measured, the same situation phrased in patient words is measurably less likely to reach the clinical answer than in clinical terms.

Two kinds of provenance appear in the gallery. The comparison figures and the simulated scenarios use prompts written by an LLM and checked automatically. The phrase dataset is the other kind: hand-built and measured by hand (currently phrases written by Michael D. Green, will be later replaced with patient language from interviews).