Graph-capture qualification gates
Measured evidence: UNSHIPPED/BLOCKED. No reviewed benchmark/profiling CUDA artifact, final source revision, or designated qualified device is recorded. Missing results are unavailable and unmeasured, not zero or passing.
This is the current evidence authority extracted from the 2026-09-07 snapshot
at revision 7b064295037c3c9a137f5dd13c03f4a40da3f11b. The historical plan closure on 2026-09-27
did not supply measurements. Historical implementation blockers are not current
implementation status. Use the operator runbook
for the supported runtime contract.
Qualification metadata
Section titled “Qualification metadata”| Field | Value |
|---|---|
| Record date | 2026-09-07 |
| Final source revision | UNAVAILABLE — evidence not collected |
| Python | UNAVAILABLE — evidence not collected |
| Warp | UNAVAILABLE — evidence not collected |
| CUDA driver/runtime | UNAVAILABLE — evidence not collected |
| Warp availability | UNAVAILABLE — evidence not collected |
| Designated qualified CUDA device | UNAVAILABLE — no designated qualified CUDA device |
| Supplemental devices | UNAVAILABLE — evidence not collected |
Production targets
Section titled “Production targets”Future producers must inspect the capture capability, enqueue, preallocation, capture/replay, parity, benchmark, and profiling implementations and their final executable diff. Exclude tests, documentation, internal plans, and artifacts; then freeze a sorted, deduplicated repository-relative production-module list. Use the operator runbook, benchmark record, and profiling record to locate the supported boundaries and their evidence requirements. These are static refresh inputs, not a live dependency of this record or its test.
Final executable diff: UNAVAILABLE — final executable diff not collected.
Candidate production modules
Section titled “Candidate production modules”UNAVAILABLE — final executable diff not collected.
Frozen production targets
Section titled “Frozen production targets”UNAVAILABLE — final executable diff not collected.
Future coverage rows use exactly this format:
<repository-relative module> | <literal pytest --cov module command> | <term-missing output> | PASS|FAILED|UNAVAILABLE
| Coverage record | Requirement | Status |
|---|---|---|
| Target list | Frozen production targets | UNAVAILABLE |
| Aggregate changed-module coverage | >=80% |
UNAVAILABLE |
Documentation and tests are excluded from production targets.
Required evidence matrix
Section titled “Required evidence matrix”Only PASS with a populated final revision, designated device, required command
output, and applicable valid provenance satisfies a row. FAILED, STALE,
INFERRED, UNAVAILABLE, and CLEAN-SKIP do not. Supplemental devices are
metadata only.
| ID | Required evidence | Final revision | Designated device | Status | Evidence / blocker |
|---|---|---|---|---|---|
| H1 | Qualified capture/replay with no replay-time allocation, host transfer, or bulk synchronization | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | C2, C7; A1 |
| H2 | Deterministic rejection or explicit recapture for structural changes | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | C2; A1 |
| H3 | CPU, uncaptured Warp, and captured CUDA parity plus tight concentration-weighted conservation | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | C2, C7; A1 |
| H4 | Persistent nonaliasing RNG continuation, reset, and restart behavior | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | C2; A1 |
| H5 | Co-located tests and configured >=80% coverage |
UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | C4; frozen targets and aggregate gate unavailable |
| H6 | 1/10/100/1000-box scaling evidence or explicit unavailable rows | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | C8; benchmark artifact unavailable; A1 |
| H7 | Small/medium captured-versus-uncaptured launch-overhead provenance | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | C8; benchmark artifact unavailable; A1 |
| H8 | Logical memory model and observed peak-memory comparison | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | C8; benchmark artifact unavailable; A1 |
| H9 | Nsight occupancy/memory-access evidence and bounded follow-up | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | C9; profiling artifact unavailable; A2 |
| H10 | Example/runbook/limitations and Starlight build result | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | C5; final-revision documentation output unavailable |
| H11 | Exact qualification command matrix and literal results before promotion | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | C2–C5, C7–C9; literal results unavailable |
Collection commands
Section titled “Collection commands”These commands remain unexecuted for this qualification record. Results from other revisions or devices cannot populate its gates. Historical C1/C6 prose tests were removed; Starlight C5 now owns documentation validation. C2–C3 are focused checks, C4 owns aggregate coverage, and C7–C9 are opt-in native-CUDA collection. A clean skip cannot satisfy measured-evidence gates.
The commands below use the current direct tools after retirement of the old workflow wrappers. This command update supplies no new qualification evidence; all required outcomes remain unexecuted for this record.
pytest particula/execution/tests/graph_capture_test.py particula/execution/tests/captured_full_loop_test.py particula/execution/tests/rng_invariance_test.py particula/execution/tests/checkpoint_test.py -q --no-cov
NOT RUN — output must be pasted verbatim after executionruff check particula/ examples_tests/ scripts/ --no-fixruff format particula/ examples_tests/ scripts/ --checkmypy particula/ --ignore-missing-importsNOT RUN — output must be pasted verbatim after executionpytest --cov=particula --cov-report=term-missing --cov-fail-under=80
NOT RUN — output must be pasted verbatim after executionnpm --prefix website run build
NOT RUN — output must be pasted verbatim after executionpytest particula/execution/tests/captured_full_loop_test.py -q -m "warp and cuda" --no-cov
NOT RUN — output must be pasted verbatim after executionpytest particula/gpu/tests/benchmark_test.py --benchmark -k resident -v -s --no-cov
NOT RUN — output must be pasted verbatim after executionpytest particula/gpu/tests/profiling_smoke_test.py --benchmark -q --no-cov
NOT RUN — output must be pasted verbatim after executionArtifact provenance rules
Section titled “Artifact provenance rules”Each artifact/manifest ledger record must contain record ID, manifest pointer, schema/version, final source revision, workload ID, machine/device provenance, contained relative raw filename, byte size, and lowercase SHA-256. Reject absolute or traversal paths, latest-path selection, symlinks, raw local reports, and copied summaries.
| Record ID | Manifest pointer | Schema/version | Final revision | Workload | Provenance | Raw filename | Byte size | SHA-256 |
|---|---|---|---|---|---|---|---|---|
| A1 | UNAVAILABLE — absent reviewed benchmark CUDA artifact | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE |
| A2 | UNAVAILABLE — absent reviewed profiling CUDA artifact | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE | UNAVAILABLE |
Promotion rule
Section titled “Promotion rule”Require H1–H11 PASS on one final revision and designated qualified CUDA device, reviewed safe artifacts where applicable, literal command outputs, and frozen production targets with at least 80% aggregate coverage. Otherwise retain UNSHIPPED/BLOCKED. Validate manifest consistency before changing row statuses; keep row and command ordering. Supplemental devices are metadata only.
Explore this topic: Execution and release qualification