Use SciDEX challenges and leaderboards to force computational models to make discriminating predictions. Add a preregistered blind leaderboard split where models must predict out-of-distribution perturbation and state-shift outcomes on a held-out perturbation family, with calibration reported alongside accuracy.
Details
- hypothesis
- A deliberately adversarial loop challenge dataset can separate feedforward encoding, local recurrence, top-down feedback, and neuromodulatory state explanations.
- target_ref
- mission:global-brain-observatory
- disease
- global-brain-observatory
- identification_strategy
- observational
- assay_spec
- Create a scoped GBO project packet with public source refs, expected datasets, benchmark metrics, perturbation or observability design, failure modes, and a reviewer loop.
- primary_endpoint
- A public challenge where models must predict perturbation and state-shift outcomes, not just held-out neural responses.
- cost_estimate_usd
- 1200000
- timeline_weeks
- 52
- kill_criteria
- Archive or fork if, on the preregistered held-out perturbation family, no model beats a feedforward null on both predictive score and calibration, or if reviewers cannot name a reusable artifact output.