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
Parent\n\n#314\n\n## What to build\n\nCreate one shared EventCohortPartition artifact used by settlement, Up execution, and Down execution research. It deterministically records train event IDs, crossing-excluded event IDs, held-out event IDs, the common time boundary, and a digest. No evaluator or MCTS adapter may independently resplit rows.\n\nThe partition is upstream of ResearchSnapshot construction. It must be derived from #319's opaque independently verified ready-event/cohort projection and bind the ordered ready-entry/catalog identity. It must never include a ResearchSnapshot digest in its own hash; #320 binds the completed partition digest into the ResearchSnapshot.\n\n## Acceptance criteria\n\n- [ ] Partition input is authenticated ready-event metadata from #319 and output is a versioned immutable artifact with deterministic digest.\n- [ ] The partition digest binds the ordered ready-entry/catalog identity, market_id, reference-path window, common time boundary, and assignments; it does not bind a downstream ResearchSnapshot digest.\n- [ ] Every market_id appears in exactly one of train, crossing-excluded, or held-out.\n- [ ] Events whose reference path overlaps the time boundary are excluded rather than split by rows.\n- [ ] Settlement, Up, and Down tasks consume the exact same partition digest.\n- [ ] Held-out rows and labels are inaccessible to candidate generation, MCTS selection, and backpropagation.\n- [ ] Counterexamples cover row-random split, one market on both sides, crossing leakage, reordered input instability, task-specific resplitting, missing catalog members, and snapshot/partition hash cycles.\n\n## Blocked by\n\n#319\n#321\n\n## Out of scope\n\nMission schema beyond partition reference, evaluator metrics, MCTS search policy, snapshot construction, and runtime deployment.
Parent\n\n#314\n\n## What to build\n\nCreate one shared EventCohortPartition artifact used by settlement, Up execution, and Down execution research. It deterministically records train event IDs, crossing-excluded event IDs, held-out event IDs, the common time boundary, and a digest. No evaluator or MCTS adapter may independently resplit rows.\n\nThe partition is upstream of ResearchSnapshot construction. It must be derived from #319's opaque independently verified ready-event/cohort projection and bind the ordered ready-entry/catalog identity. It must never include a ResearchSnapshot digest in its own hash; #320 binds the completed partition digest into the ResearchSnapshot.\n\n## Acceptance criteria\n\n- [ ] Partition input is authenticated ready-event metadata from #319 and output is a versioned immutable artifact with deterministic digest.\n- [ ] The partition digest binds the ordered ready-entry/catalog identity, market_id, reference-path window, common time boundary, and assignments; it does not bind a downstream ResearchSnapshot digest.\n- [ ] Every market_id appears in exactly one of train, crossing-excluded, or held-out.\n- [ ] Events whose reference path overlaps the time boundary are excluded rather than split by rows.\n- [ ] Settlement, Up, and Down tasks consume the exact same partition digest.\n- [ ] Held-out rows and labels are inaccessible to candidate generation, MCTS selection, and backpropagation.\n- [ ] Counterexamples cover row-random split, one market on both sides, crossing leakage, reordered input instability, task-specific resplitting, missing catalog members, and snapshot/partition hash cycles.\n\n## Blocked by\n\n#319\n#321\n\n## Out of scope\n\nMission schema beyond partition reference, evaluator metrics, MCTS search policy, snapshot construction, and runtime deployment.