3.5 KiB
Uniform full-nine DFG qualification
This is a separate full-run entrypoint. The pilot wrapper, pilot manifest, shared observer, reference data, gates and all production files remain byte-identical to c0be841. Numerical source remains 926ca27407c26ca11283282a6e53249925725376; actual execution HEAD is recorded separately. A new full manifest pins the added wrapper and the original pilot manifest.
Admission and case
Require the completed exact c0be841 uniform SRT/full9 cylinder pilot directory, including result.json, preflight.json and uniform/ready.json. Admission checks finite complete pilot history, collision/inlet/case/observation/source identity, physical configuration, body, parameters, frozen gates, READY, clean checkout and positive timings. Bodyless, MRT, refined, stale-source or malformed receipts cannot admit this run.
Restart from the unchanged original production initialization, not a pilot continuation. Keep D80, lattice mean U=.01, viscosity=.008, tau=.524, shifted FP32 double buffer, cylinder/walls/domain, no LES/force, and pressure-NEQ outlet all-nine damping .5. Only the already tested experimental inlet differs from the historical baseline.
Maximum 2,400,000 steps to t*=300. Reuse the existing observer: t*=.5 intervals before240, then .02 through300. All force batches are included; force time is the interval midpoint, pressure time its endpoint. Early calibration remains accounted in the physical trajectory. Sparse health checks identify first observed failure, not exact onset. Timeout retains progress.
Success requires the full physical cap, healthy completion, frozen periodic convergence AND official-reference agreement. Healthy completion alone or periodicity alone returns failure. The project tolerances remain unchanged; the reference data are official DFG data. This does not qualify a canonical do-nothing outlet, grid convergence, AMR, MRT, LES or other precision paths.
Cost and bounds
Measured pilot startup: 3.3353 s. Step plus force: 3.6794 s for 40,000 steps. Linear step-work extrapolation is 220.764 s, before additional late-time dense observations and altered batch subdivision. The full observer has 3,480 intervals (480 sparse and 3,000 dense), compared with 10 pilot intervals. Prior same-size MRT full runs took 354–373 s total. Expect several minutes, with timing reported rather than promised.
Internal cap: 720 s maximum. Outer watchdog: 900 s, then 30 s termination grace. The wrapper rejects a larger internal cap. Expiry preserves an incomplete result and does not qualify accuracy. No GPU execution is performed during preparation.
Commands
CPU tests:
python tests/validation/experimental_full9/test_dfg_full.py
python tests/validation/experimental_full9/test_dfg.py
python tests/validation/experimental_full9/test_full9.py
Preflight against the actual retained pilot directory:
python tests/validation/experimental_full9/run_dfg_full.py --pilot-results <actual-c0be841-pilot-directory> --output <new-preflight-directory>
After independent review and explicit full-run admission:
CUDA_VISIBLE_DEVICES=<admitted-device> timeout --signal=TERM --kill-after=30s 900s python tests/validation/experimental_full9/run_dfg_full.py --pilot-results <actual-c0be841-pilot-directory> --execute --gpu-admitted --wall-seconds 720 --output output/new-full9-dfg-full
The four admission tests use clearly labeled synthetic receipts. They are host regression fixtures, not measured pilot evidence. Runtime admission uses the actual retained pilot receipts.