Goal. Load the multiple-choice text substrate (the secondary, generality lane).
Scope / what to build.
- Load a released MedQA (USMLE-style) multiple-choice set through a standard datasets library into a shared MCQ schema: stem, options, correct answer, id.
- Note: the exam questions are the injection substrate, not a medical-ability claim; what we measure is behavior change under controlled perturbation.
Reuse vs bespoke. Reuse the datasets library. Bespoke: the shared MCQ schema mapping only.
Definition of done. MedQA loads into the schema; a smoke test confirms options and correct-answer alignment on a small sample.
Model backend. For now every agent uses the Gemini API (multimodal). All model access goes through one gateway wrapper, so the roster can be extended to other model APIs later without changing any experiment code. Gemini is the default for now; no fine-tuning, models are used off-the-shelf.
Goal. Load the multiple-choice text substrate (the secondary, generality lane).
Scope / what to build.
Reuse vs bespoke. Reuse the datasets library. Bespoke: the shared MCQ schema mapping only.
Definition of done. MedQA loads into the schema; a smoke test confirms options and correct-answer alignment on a small sample.
Model backend. For now every agent uses the Gemini API (multimodal). All model access goes through one gateway wrapper, so the roster can be extended to other model APIs later without changing any experiment code. Gemini is the default for now; no fine-tuning, models are used off-the-shelf.