Files
MeltMind/review/2026-11-14/s1-sampled-o2-gate/README.md
T

3.4 KiB
Raw Blame History

Sampled O2 equivalence gate and frozen driver/supervisor

Six fields (850/778/666, original + rotated) PASS bit-for-bit against f2e5ed0. Each compares 30 node/edge arrays, including all frozen integrate outputs, mu, g, B, gate q/status, all selected-anchor true edge q/status, and NaN bit patterns. No numeric tolerance is used for this gate. Test draws use PCG64 SeedSequence([2026111400,prefix]), d>0 mass PPS, n=2000, ascending IDs; first 200 sorted unique plus smallest boundary and nearest gate. Test IDs are 201/202/201 per part, shared across poses. They are separate from the fixed 24-part draws.

Gate-ID check: actual nearest IDs exactly equal later explicitly named frozen IDs: 850=0, 778=27375, 666=0; boundary minima 0/1/0. The actual earlier nearest-node algorithm is preserved, not rewritten after execution. Complete test-draw arrays and explicit ID CSVs are in ZIP.

Driver calls unchanged raw_field/samples/integrate and o2_rep.readout/contract_direction. Frozen source SHAs checked at each field start. It integrates full source fields at selected anchors only. R, full raw gradient field and all four-point sources are retained. L_med comes from all original true edges, R_g from all full-part lumped masses before eligibility. Each selected anchor's complete edge endpoints are processed; valid finite endpoints alone enter mean/max, no valid endpoints means ineligible. Invalid masks/reasons and all computed arrays are retained.

The supervisor runs one field/process at a time, samples process-tree RSS/HWM with a requested 20ms interval, measures the complete lifecycle including import/preparation/readout/write, and checks the final reported ru_maxrss. Limits 1800s / 1342177280 bytes. Polling is not hard realtime: each receipt reports observed maximum scheduling gap. Empty/zombie memory fields are zero observations, never max(empty); missing final result cannot PASS. Budget failure is recorded without replacement. No S1 scoring or training.

Process deviation and hold

The panel queue started at the 14:56:19 receipt under the earlier 14:43 conditional GO before the agent had read the 14:53–14:55 added requirement for Grok driver/supervisor review. This was a process deviation. Claude at 14:56:43 explicitly allowed the already-started queue to finish but HELD outputs: no shortfall or S1 until Grok reviews these gate receipts and SHAs. If review finds a problem, outputs are invalidated and the same frozen draws rerun; no numerical patching afterwards. This package records that status honestly, not retroactive prior approval. The panel run remains separate and is not included in this gate-only archive.

Authority: https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794639403231829 Original gate approval: https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794638614756099 Gate/queue start: https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794639379764799

The ZIP contains all gate arrays, test draws, 30-array bit checks, results, supervisor receipts/monitor logs, contract, exact frozen code and source records. The six original full-node fields remain in f2e5ed0 rather than being duplicated. The gate compares node/edge subsets only, not sampled full-mesh aggregates. Reviewer can read run_sampled_o2.py and supervise_sampled_o2.py directly beside this README.

Driver SHA256: 66ab07068c6553a5d49d5124087d69c338e057919e2fe0c31124ba6d2d2a5d36

Supervisor SHA256: 9ba68b5460481683ef0cc3138bdcdd70df4a67fd9d02e90a57b79ef4b1ddc2c8