Realistic seeded-shortcut construction for the cascade (confident rationale, not a bare option)

[sebasmos]

Problem

experiments.run_cascade seeds the shortcut with _shortcut_answer, which just picks a distractor option string. On a shared blackboard a bare planted option is a weak stimulus: real cascades are driven by a confident, plausibly-reasoned wrong answer from a peer, not a one-word assertion. To measure contagion realistically the seeded turn should look like a competent agent that is confidently wrong for a shortcut reason.

What to build

A seed_builder that, given a case and a chosen shortcut target, produces the seeded turn's content and answer:

• A short clinical-sounding rationale that anchors on the spurious cue (for the text lane, e.g. leaning on the option that the cue favors; for imaging, citing the injected artifact as if it were signal).
• Configurable confidence and verbosity, so we can run a stimulus-strength ablation (bare option vs terse claim vs full confident rationale) and see how onset and contagion respond.
• Deterministic given (case, target, style) so runs are reproducible.

Wire it into run_cascade behind a parameter, defaulting to the current bare-option behavior so existing tests and the audit arms are unchanged.

Why it matters

The contagion index compares shared vs isolated adoption of the planted shortcut. If the stimulus is unrealistically weak we will under-estimate contagion and under-sell the referee's value. A realistic seed is also what a reviewer will expect when they ask "would a real bad agent look like this?"

Acceptance

• run_cascade accepts a seed_style and the seeded turn carries a rationale for the non-default styles.
• A stimulus-strength ablation runner (or a parameter sweep) shows onset/contagion as a function of seed style.
• Ground truth is never altered by the seed; only the planted peer turn changes. Tests assert the seeded answer is a non-correct option and the case's answer_index is untouched.

[sebasmos]

Concrete evidence from the first full real run (100 solo + 20 cascade cases on MedQA, same-lineage Gemini) makes this issue urgent and pins down exactly what the fix is.

What we observed: the current experiments._shortcut_answer picks the first distractor. On the 20-case cascade run, the seeded answer coincided with the committee's own independent baseline answer in 16/20 cases (the pre-seed turn-0 answer already equalled the seed). When the seed equals the baseline there is no counterfactual gap: shared and isolated both sit at ~1.0 adoption and the contagion metric (shared minus isolated) is undefined. Mean contagion came out -0.006, i.e. not interpretable, and even the 4 cases where the seed differed showed shared == isolated.

Concrete example (case medqa-1033): seeded answer "Atropine", but turn 0 (before the seed is planted) is already "Atropine", and it stays "Atropine" in the isolated run too. Nothing cascaded; the agents were already there.

What the fix must do: choose the seed RELATIVE TO each committee's isolated baseline, a wrong answer the committee would NOT independently give (differs from the isolated-run committed answer, and is not the ground truth), not just "the first distractor". Optionally also make it a confident clinical rationale (the original scope of this issue). Only then does adoption under shared context measure contagion.

Suggested acceptance addition: assert seed_answer != isolated_baseline_answer and seed_answer != ground_truth per committee/case, and add a regression test where the isolated baseline differs from the seed so contagion is well-defined.

The solo susceptibility arm is unaffected and solid (flip 0.79-0.89, noise floor 0.000).