Roadmap v8 final champion/challenger selection — issue #350

Current status: blocked; no selection or accuracy conclusion is available.

Run the structural gate with:

PYTHONPATH=src python -m btc_timesfm.research.final_selection
PYTHONPATH=src python -m btc_timesfm.research.final_selection --contract frozen_contract.json

The first command reports the missing-evidence blockers in this checkout. The validator never creates missing corpus records, scores models, chooses a winner, or changes production. A ready_for_evaluation result means only that the contract's evidence structure passed validation; candidate metrics and all acceptance criteria still need a separate, auditable evaluation. Winner remains null and recommendations remain human-only.

The frozen contract must enumerate every attempted candidate, including failed and blocked attempts, and bind each to exact source, vintage, data, model, and policy identities. Eligible candidates must retain raw and final forecasts, failures, and the same exact origin/target/horizon pair keys across the complete four-horizon family (2h, 4h, 8h, 16h). Selection uses nested earlier folds, labels available by the fixed cutoff, a disjoint final holdout, dependence-aware moving-block inference, and whole-family multiple-test correction over all attempts. Comparisons include the frozen champion, persistence, and strongest naive baseline.

Prospective D3 begins only after freeze, follows the fixed cutoff/stopping rule, is disjoint from selection and final holdout, spans at least 30 calendar days, and has at least 200 exact mature pairs at every horizon. The evaluation gate must retain raw/final/failure outcomes and exact source/vintage joins. Any later human recommendation must apply the predeclared effect, adjusted confidence, baseline, powered-regression, and directional-loss criteria; no post-hoc thresholds or repeated peeking are permitted. No automatic promotion or production changes are allowed.

The checked-in canonical benchmark audit currently reports no eligible immutable BTC/USD corpus. Consequently, absent external frozen artifacts, the expected report is blocked; this is not a negative forecast-accuracy result.