bundle/v0/RUNBOOK.md

Preparation runbook

From this lane directory:

python -m pip install --no-deps -e .
python -m loss_aware_dro_repro.cli validate-plan
python -m loss_aware_dro_repro.cli materialize-manifest --output .openresearch/artifacts/task-manifest.jsonl
python -m loss_aware_dro_repro.cli smoke --output .openresearch/artifacts/smoke/receipt.json
pytest -q

smoke is explicitly TOY_DIAGNOSTIC and cannot support C1-C3. Do not invoke the future conic runner until the clean environment, independent method implementation, reviewer quorum, exact config hash, and launch authority all pass.

Validate the clean matched artifacts without launching anything:

& .venv/Scripts/python.exe -B scripts/validate_scientific_canary.py .openresearch/artifacts/scientific-canary-v14 --manifest .openresearch/control/scientific-risk-sensitivity-manifest-v4.json
& .venv/Scripts/python.exe -B scripts/validate_scientific_canary.py .openresearch/artifacts/scientific-canary-v15 --manifest .openresearch/control/scientific-risk-sensitivity-manifest-v4.json

Recorded identities v3, v6-v7, and v10-v13 must not be relaunched. Earlier unexecuted v8-v9 configs are superseded by the hardened evidence contract and must not be used.

Identities v14 and v15 ran exactly once from clean commit 8be5520 and must not be relaunched. Stopping-rule identity v16 ran exactly once from clean commit d912834 and must not be relaunched. Validate it with:

$env:PYTHONPATH="$PWD/src"
& .venv/Scripts/python.exe -B scripts/validate_scientific_canary.py .openresearch/artifacts/scientific-canary-v16 --manifest .openresearch/control/scientific-convergence-manifest-v1.json

The receipt is claim-ineligible and must be interpreted as paper_stopped_not_converged.

Validate the canonical Appendix-routes bundle without launching anything:

& .venv/Scripts/python.exe -B scripts/validate_appendix_canary.py .openresearch/artifacts/appendix-canary-v5 --manifest .openresearch/control/appendix-routes-manifest-v3.json

Appendix identities v1, v2, v4, and v5 must not be relaunched. The earlier unexecuted v3 config is superseded by the hardened contract and must not be used. Identity v5 ran exactly once from clean commit 8be5520 and is bound by Appendix manifest v3.

Validate the independently regenerated hypergradient component artifact:

$env:PYTHONPATH="$PWD/src"
& .venv/Scripts/python.exe -B scripts/validate_hypergradient.py `
  .openresearch/artifacts/validation/hypergradient.json `
  --config configs/hypergradient_validation_v1.json

The validator first verifies the adjacent SHA-256 sidecar and proves every bound scientific file against the base commit's Git tree and blob objects. In full mode it also checks the documented locked environment, including Clarabel and POT, reconstructs all 93 numerical gradient matrices from 882 retained scalar evaluations, checks all solver and transport residuals, and independently regenerates the full evidence payload. --structural-only explicitly does not perform regeneration. See HYPERGRADIENT_VALIDATION.md for the executed hashes and limits. This artifact validates the gradient oracle components only. It does not establish paper-scale C1 or convergence claim C2, so all scientific verdicts remain HOLD.

After paper-scale execution, evaluation requires three immutable inputs:

python -m loss_aware_dro_repro.cli evaluate \
  --results .openresearch/artifacts/raw/task-results.jsonl \
  --gradient-checks .openresearch/artifacts/validation/hypergradient.json \
  --proof-audit .openresearch/artifacts/validation/theorem-5.1.json \
  --output .openresearch/artifacts/evaluation/claim-verdicts.json

Missing, duplicate, unexpected, author-generated, hash-mismatched, or trace-free results fail closed.

Precision follow-up gate

Do not run this gate from a dirty lane. First require a clean commit, the full test suite, and an independent implementation review. Planning does not launch work:

& .venv/Scripts/python.exe scripts/run_local_batch.py plan-precision-followup `
  --manifest .openresearch/control/stopping-storage-precision-followup-v1/batch-manifest-v2.json `
  --receipt .openresearch/control/stopping-storage-precision-followup-v1/batch-plan-v2.json

Manifest v1 failed closed because the lane was dirty and issued no task. It must not be relaunched. Manifest v2 is the executed clean identity. The plan must project exactly 2,580 tasks, eight workers, zero external cost, denied external authority, task-list hash c3d85274bac29c5f4a6b4d334a57459689fb17f7fd4b4dc6c9ae7529df855b, and the frozen 5,400-second issuance gate. Only then run the exact planned batch:

& .venv/Scripts/python.exe scripts/run_local_batch.py run `
  --manifest .openresearch/control/stopping-storage-precision-followup-v1/batch-manifest-v2.json `
  --output-root .openresearch/artifacts/stopping-storage-precision-followup-v2

Aggregation must report 2,580 validated successes, zero rejected tasks, no blocker, and complete: true:

& .venv/Scripts/python.exe scripts/run_local_batch.py aggregate `
  --manifest .openresearch/control/stopping-storage-precision-followup-v1/batch-manifest-v2.json `
  --output-root .openresearch/artifacts/stopping-storage-precision-followup-v2

The executed aggregate is sha256:65c51df2681e33ba1363341cd44bcfd8d094b8427cc599eb40ac6c1f95ae862d, with 2,580 validated successes, zero rejected tasks, no blocker, and complete: true. Analyze it with:

& .venv/Scripts/python.exe scripts/analyze_precision_followup.py `
  --manifest .openresearch/control/stopping-storage-precision-followup-v1/batch-manifest-v2.json `
  --output-root .openresearch/artifacts/stopping-storage-precision-followup-v2 `
  --aggregate-summary .openresearch/artifacts/stopping-storage-precision-followup-v2/aggregates/65c51df2681e33ba1363341cd44bcfd8d094b8427cc599eb40ac6c1f95ae862d/summary.json `
  --aggregate-results .openresearch/artifacts/stopping-storage-precision-followup-v2/aggregates/65c51df2681e33ba1363341cd44bcfd8d094b8427cc599eb40ac6c1f95ae862d/results.jsonl

The sealed analysis is sha256:4b2009f94a968455b96ec3f7c0a618cb446d4f3c67ec610f86e0180097e2c3b6. A second full analysis produced the same hash. The precision gate passed, but its only positive eligibility is diagnostic_resource_estimate_eligible. Claim, cost-freeze, and full-matrix-freeze eligibility remain false. C1, C2, and C3 remain HOLD.

Checkpoint/resume canary

This canary is local, claim-ineligible, and limited to eight frozen tasks. It runs a distinct-identity uninterrupted oracle, pauses the candidate after completed iteration zero, and resumes the candidate in a fresh process. It compares canonical timing-free scientific projections, including the production terminal post-update re-solve, as exact bytes; it never uses a numeric tolerance. Immutable parent-launch and worker-process receipts bind the observed PID plus a per-process UUID and nonce, and validation requires the worker's OS-observed parent PID to match the parent launch binding, the parent-observed worker PID to match the worker receipt, and the start and resume process instances to differ.

Run only from a clean committed lane after focused tests pass:

& .venv/Scripts/python.exe -B scripts/run_checkpoint_resume_canary.py run `
  --config configs/checkpoint_resume_canary_v1.json `
  --output-root .openresearch/artifacts/checkpoint-resume-canary-v1

The same lifecycle may be inspected explicitly with start, resume, and validate. start intentionally leaves eight resumable checkpoints. resume performs an all-task preflight before authorizing any continuation. An orphan, tampered checkpoint, duplicate event, elapsed monotonic time over 300 seconds, process-instance reuse, or same-identity restart fails closed. Paid/remote compute, external inference, training, pushing, and publishing remain denied. A pass does not authorize long-tail or full-matrix work and leaves C1-C3 at HOLD.

The canary executed once from clean commit 654bce052ee076f65a1a38f06fdae769afb35403. Its manifest is sha256:559fe003a09b0b71457f125dc1fa42ff0c3e43aeb7907bbb5d987968c6853946 and its immutable receipt is sha256:ae7f31368dc23426f43b9bd8e651596cd90b5020c861a89304337da9348a679d. All eight comparisons were exact bytes, all terminal post-update re-solves matched, and the lifecycle finished in 37.203 monotonic seconds against the 300-second cutoff. A second validation and an independent artifact audit passed without changing any of the 108 files. This pass still leaves C1-C3 at HOLD.