Reports index
Gate reports and validation records produced for RFC-0001 (Batch Cycles), tracked under epic #474.
Picking the programme back up? Start with batch-cycles-checkpoint.md — the standing checkpoint: why Batch Cycles exists, what has landed, what is still open, the halted state of the scheduler on hcc2, and the resume recipe. The gate reports below are the evidence it summarises.
| Report | Gate | Verdict | Date |
|---|---|---|---|
sched-parity.md | RFC-0001 Step 1 exit gate (issue #471) | Conditional GO — conditions (#496, #500) cleared 2026-09-01 | 2026-08-30 |
batch-pilot.md | RFC-0001 Step 3 gate (issue #473) | NO-GO — batches not executable (zero ever dispatched) | 2026-09-01 |
batch-pilot-2-execution.md | Pilot execution record, attempts 2-4 (issues #526, #590); GO/NO-GO deferred to #529 | 0 of 7 dispatched batches completed end-to-end across four runs (AC1 unmet). Run 1 blocked before dispatch (#535/#539, both since fixed); run 2 dispatched 3 batches and all 3 dissolved — B3a unwarmed batch worktree, B3b npm test -- --reporter=json vs a make-delegating test script, B3c 1/3 eviction tolerance, B3d null per-issue telemetry (all fixed); run 3 reached execution for the first time — warm worktree in <1 s, both members’ agents ran — and dissolved on #579 (plan validate exit 128, fixed mid-run by PR #581) and #583 (incremental gate reads INCONCLUSIVE as a member failure, since fixed by PR #585). First controlled same-issue comparison: −49% cost, ceiling stated; run 4 (attempt 4, #590) swept the whole imboard backlog through the classifier for the first time (111 → 68 → 33 candidates → 7 slot) and dispatched 3 more batches — all warmed in the same second, all dissolved — isolating the single blocking defect: the incremental member gate is a constant function (imboard-monorepo#3996, tee /dev/stderr under pipefail; sched-side #594; eviction bookkeeping #595) | 2026-09-03 |
model-agnostic-fleet.md | Model-agnostic fleet validation (issue #528) | PARTIAL — Part I: retrospective parity holds (86% vs 86%). Part II: model × class axis built and populated; conformance parity sharper (glm 0/73 vs sonnet 9/159 criteria not-met); live run still NOT executed and re-escalated — #528 stays open (ac_met=0/5) | 2026-09-02 |
issue-538-classifier-cost-methodology.md | #538 AC3 cost methodology | PARTIAL — pre-screen hit rate measured on the 15-issue fixture set (7/15 flagged by the text floor, 0 tokens — v1 rejected them; since #772 they are review: full candidates); post-#538 per-issue cost is still an estimate — live re-measurement pending an ops dispatch run | 2026-09-02 |
Raw evidence backing these reports (baseline tables, agent-log summaries, scheduler event journals, captured corrupt-state files) lives in evidence/.
Two documents here are not gate reports and are not expected to stay frozen. batch-cycles-checkpoint.md is the programme checkpoint described above (#598); it date-stamps every claim and names the command that re-derives it, so update it in place whenever the programme moves rather than filing a new one. model-scorecard.md is a recurring cost/quality/speed tab per LLM (#566), regenerated weekly (npm run scorecard) rather than produced once per gate. Its JSON sidecar lives at evidence/model-scorecard.json.
Rendered from docs/reports/README.md in the repository. Edit it there.