You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Adapt the existing alpha-harness candidate search to the shared UCT/MCTS kernel for one admitted Binance time-series replay. Candidate generation, selection, and backpropagation must use training evidence only. After selection, run the existing held-out evaluator once and emit one research result receipt that binds the selected candidate, snapshot, search, evaluation, and checkpoint identities.
Acceptance criteria
The adapter reuses hft-search-kernel and does not introduce a second MCTS implementation or modify the Polymarket evaluator.
Search budget, stopping conditions, candidate identity, and deterministic tie-breaking are explicit and bound into the result evidence.
Durable checkpoints bind the snapshot, search, evaluation, and candidate-space identities; a mismatched or stale checkpoint is rejected.
Held-out rows and metrics are unavailable to candidate generation, selection, and backpropagation, and held-out evaluation runs only for the selected candidate.
One bounded research-only run can resume from a valid checkpoint and emits a typed receipt with training evidence and the selected candidate held-out evidence.
Focused counterexample tests prove held-out non-backpropagation and stale snapshot, protocol, or checkpoint rejection.
Parent
#205
What to build
Adapt the existing alpha-harness candidate search to the shared UCT/MCTS kernel for one admitted Binance time-series replay. Candidate generation, selection, and backpropagation must use training evidence only. After selection, run the existing held-out evaluator once and emit one research result receipt that binds the selected candidate, snapshot, search, evaluation, and checkpoint identities.
Acceptance criteria
Blocked by