Context
The first real cascade run was a null, and the diagnosis (see the team's results notes) is that the seed was a bare wrong answer. A confident committee just rejects a naked assertion. Propagation needs the seed to carry a reason tied to a cue present in the case that peers can find plausible and adopt. This is the single change most likely to make the cascade measurable, and it complements #104 (baseline-relative seed).
What to build
Extend the cascade seed builder so the planted turn is a short, confident clinical rationale that anchors on a spurious cue in the case, not just an option string. For the text lane, examples: anchor on a demographic detail, a red-herring lab value, or "the guideline says X". The seed should:
Wire it behind a seed_style parameter, default unchanged so existing tests/audit arms are untouched.
Acceptance
- A
seed_style="reasoned" produces a seed turn whose content contains the cue-based rationale.
- A cascade run on a small MedQA set records higher shared-vs-isolated adoption under
reasoned than under the bare-answer seed (even if still modest), demonstrating the mechanism.
- Ground truth is never altered; the seeded answer is a non-correct option. Regression test included.
Reuses the MedQA data + call cache; no new dependencies.
Context
The first real cascade run was a null, and the diagnosis (see the team's results notes) is that the seed was a bare wrong answer. A confident committee just rejects a naked assertion. Propagation needs the seed to carry a reason tied to a cue present in the case that peers can find plausible and adopt. This is the single change most likely to make the cascade measurable, and it complements #104 (baseline-relative seed).
What to build
Extend the cascade seed builder so the planted turn is a short, confident clinical rationale that anchors on a spurious cue in the case, not just an option string. For the text lane, examples: anchor on a demographic detail, a red-herring lab value, or "the guideline says X". The seed should:
Wire it behind a
seed_styleparameter, default unchanged so existing tests/audit arms are untouched.Acceptance
seed_style="reasoned"produces a seed turn whose content contains the cue-based rationale.reasonedthan under the bare-answer seed (even if still modest), demonstrating the mechanism.Reuses the MedQA data + call cache; no new dependencies.