translation recovery · live trace · traced July 7, 2026 · re-traced July 9, 2026 · renders re-traced July 30, 2026
Translation
Translating the patient sentence back into clinical terms restores the prediction in 8 of the 20 hardest cases.
The mitigation: an LLM rewrites the patient sentence in standard terms, and that rewrite is traced directly. Of the 20 reviewed downgrades: 8 recovered, 7 unrecovered, 5 unclassifiable. In the cases translation fixes, the clinical features reappear and the target probability recovers; the worsened rewrites are labeled on their cards below.
Translation matters most when it changes the predicted next word, not just its probability. In the reviewed downgrade set, “Grandma’s been all bunged up for a week, so before dinner she took a” continues with nap (30%); rewritten in clinical terms, the top word becomes lax(ative), above the original clinical phrasing (26%): 45% in the July 7 study trace, 41% on the July 9 re-trace shown below. It is not a cure-all: in one case the rewrite replaced the prescription (38%) with topical (15%).
Where the patch holds, and where it does not
The control that makes the case: on 14 pairs measured under both arms, the real clinical rewrite recovered a mean +1.9 points of target probability against +0.5 for a placebo paraphrase that keeps the casual register. The terminology does the work, not the rewriting.
At scale
| Model | Phrases | Mean recovery | Helped | Top restored | Top lost |
|---|
What the depth readout adds
A layer-by-layer readout splits the wording gap into two cases: answers that form and are lost late, which translation recovers, and answers that never form, which translation supplies. The running numbers, class by class, live in the data section of the Technical page.
View the live three-panel trace (clinical / patient / translated)
live trace: gemma-2-2b hosted circuit-tracer via Neuronpedia · re-traced July 30, 2026 · phrase 17 of the reviewed downgrade set (predictions that fell to a lower care tier), three panels (clinical / patient / LLM-translated)
Causal faithfulness & titration metrics (gemma-2-2b)
Dose–response: recovery by steering strength
Boost the top-5 clinical features while the model reads the patient wording; recovery = the clinical target appears in the steered output. Recovery holds near ceiling from strength 2.5 through 10 (4/5, 5/5, 4/5) and falls off at 20, where the output sometimes breaks down.
Faithfulness: does attribution rank predict effect?
same strength, different features: top-5 by attribution mass vs ranks 6–10. Control: 5 random features at strength 10 recovered 0/5 — the effect is the clinical circuit, not steering itself.
The traces here are live gemma-2-2b runs.
SIMULATED DATA: the sentence pairs measured across this site are written by an LLM and checked automatically. The scenarios are simulated; the testing is real. The phrase dataset is the exception: hand-built from real patient language and measured by hand.