Benchmark protocol v0.2.0
Benchmark layers
- Original-valid layer: clinically/reference-valid QA before empirical evidence gating.
- Evidence-gated primary layer: items that pass clinician evidence sufficiency and pre-registered empirical dependence criteria.
- Adversarial/safety layer: question-only, hard mismatch, wrong organ/modality, and blank/occluded conditions.
Why these layers are separate
Combining them into a single score hides mechanism. A model can answer original items well while ignoring images, or use images well but fail to abstain when the evidence is invalid.
Ranking analysis
For every frozen model panel, compute rankings on original-valid and evidence-gated layers. Report Spearman rho, Kendall tau, absolute rank displacement, pairwise inversions, and item-clustered bootstrap confidence intervals. Pair inversion is an informative secondary result, not a required outcome.
Item-first reporting
Report item-level condition scores and distributions, not only aggregate accuracy. Stratify by question family, answer type, modality, label frequency and doctor evidence state.
Pilot limitation
The current four-case package cannot support research-grade ranking inference. All evidence-audit files are schemas/templates until a sufficiently large frozen set exists.