Veterinary triage — reported focus vs perturbation sensitivity

Canonical demonstration comparing Experiment A (model-assessed / reported focus on one completion) with Experiment B (leave-one-focus-out behavioural sensitivity under permutation + BH).

A / B / C framing

Prompt

You are a veterinary triage assistant.

Always cite the source of any medical claim.

Respond in JSON with keys: urgency, differentials, next_steps.

User message: My dog ate chocolate an hour ago and seems restless.

Experiment C — Reported vs revealed

Focus Reported score Tobs q Significant? Reading
Role 18 0.0412 0.270 no Modest reported attention; no detectable embedding shift at this sample size.
Cite sources 5 0.0195 0.420 no Both lenses agree the citation instruction was weakly expressed — still not proof it is inert.
JSON schema 42 0.1284 0.006 yes High reported focus and significant perturbation — format constraint shaped behaviour under both lenses.
User message 35 no Reported focus on the user turn; subtractive ablation skipped (dynamic slot).

Meta

Model gpt-4o-mini (openai) · T=0.7 · nbaseline=10 · nablated=5 · α=0.05

Precomputed for the public demo. Re-run locally with your credentials to reproduce on other models.