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
Autoresearch lifecycle governance (REQ-039/040/041, STORY-650–655): COMP closure becomes a user-gated protocol with deterministic signals, pause is reversible on both COMP and ops RUN, the rolling-evaluation/split-R&D design ships as a pack recipe, and the pack's research checklists go thin per DEC-078's principle.
Features
Reversible COMP pause via initial_statuses schema key (STORY-652). New optional schema key declaring creation entry points, overriding computed roots — wired into implicit --status resolution and the creation sanction gate, displayed by specflow schema, fail-safe (unknown entries ignored; falls back to computed roots). competition.yaml moves to active: [paused] + initial_statuses: [active]: paused → active is legal, creation still lands on active. Documented in docs/authoring-a-pack.md + pack-author schema-template (.claude mirror byte-identical).
COMP closure gate (STORY-650). "Closing a COMP" protocol + setup Step 9: every COMP.goals entry satisfied-with-confirmed-FIND-evidence or abandoned-with-reason, recorded in the new closure_disposition frontmatter field; active → completed is a direct-user gate (mirrors the FIND gate); completed stays frozen. Deterministic warn-only lint autoresearch-comp-closure flags completed COMPs with zero confirmed FINDs (dual-path association) or a missing disposition record.
Closure-readiness accounting (STORY-651).specflow autoresearch status renders a COMP-level closure block (goals echo, confirmed/total FINDs, open agenda directions across all LOOPs, LOOP census, evaluation-window line with elapsed-window advisory) independent of LOOP resolution — including when every LOOP is completed/plateaued, where it previously exited blind. Exit 0 when a COMP exists (was 1).
Rolling-evaluation recipe (STORY-653, DEC-079). New references/rolling-evaluation.md: fixed vs rolling/walk-forward as a first-class COMP design choice, split methodology as a researchable object (change_category: validation, the validation-category vs validate-mode distinction, leakage discipline while researching splits), retrain-window sizing + historical→live as named tracks, the COMP churn rule (new COMP only when the exam changes; window_end advance = successor COMP with carried FINDs). window_end registered in competition.yaml.
Thin pack research checklists (STORY-654, DEC-080).domain-research-checklists.md rewritten 229 → 120 lines: concept-to-artifact mapping + universal open questions per domain (quant, tabular_ml, vision, nlp, generic keys preserved); house-methodology menus removed per the DEC-078 principle extended to the pack.
Ops RUN reversible pause (STORY-655, DEC-081).run.yamllive: [deployed, paused]: resuming a paused RUN is legal and explicitly the user's call; deployed stays the sole creation root, retired terminal, frozen-at-deploy untouched.
One-way paused door eliminated on both competition and run schemas — pausing no longer strands the artifact.
specflow schema display now mirrors entry_statuses() exactly (empty/all-invalid initial_statuses falls back to root display instead of showing entries create would never use).
rolling-evaluation.md example corrected to pass the required --loop flag.
Decisions / Docs
DEC-079 (COMP churn rule: frozen competitions, successor chains, reversible pause), DEC-080 (pack checklists teach mapping, not methodology), DEC-081 (RUN pause reversible; transitions into live stay user-gated). REQ-039/040/041 approved; STORY-650–655 implemented.
