Files
MeltMind/docs/o1-scientific-v2-preparation-2026-11-17.md
T

16 lines
3.5 KiB
Markdown

# O1 scientific v2: approved same-sixty geometry gate
Claude1794863409.977129 / h-2e3bd8ed approved Mold specification SHA256 632688217e59dd134922245e524ff1c0f63933be64600be21379f7aa74947a5f as written. Exact spec is references/2026-11-17/o1/mold-o1-v2-geometric-spec.md. This is a new scientific version, distinct from the earlier engineering-only run_o1_prepare_v2.py repair. All old sources, sixty attempts, failures and accounting remain immutable.
Question: can the prescribed type-dependent cut costs produce identifiable interventions at K=8 while retaining physical matching? This is not a prediction result or a general topology verdict. All arms use the same METIS 5.1.0, volume weights, connectivity settings and seed0, one generation per arm. G uses physical face area; T/C use the approved area-midrank kernel, shared true-field g_p and common G integer scale. No candidate selection, retuning or extra seed. All original physical volume conditions remain; cut area and sorted diameter max/min must be <=1.25. Original five-percent matching is descriptive only.
New tools: o1_v2.py, o1_v2_prepare.py, run_o1_v2.py, test_o1_v2.py and metadata-only prepare_o1_v2.py. Fourteen bounded synthetic checks PASS, including exact weighted ranks/ties, shared strength/constant field, common quantization and symmetric CSR, physical thresholds, strict panel gates, equal component weighting, actual runtime schema, all-sixty denominator and a three-call mixed-VTU end-to-end check that refuses labels. Guard CPU1.435606 seconds. Test source mtimes precede the accepted test start; hashes are frozen in accepted-source.json, immutable imported sources checked against v1.
Root artifacts/o1-v2-preparation-2026-11-17. Contract SHA256 90caf09948fe2b0cc51c54c7cf87986f5daaa86f222dd2d1cf231bf5bc70a874. panel-manifest freezes all sixty original attempted IDs in original dispatch order, including eleven ineligible and three interrupted; each original guard/result fingerprint is retained. The six original unavailable feature rows remain in the full617/209 denominator. No substitution.
At exactly sixty terminal attempts, before any further dispatch, both part and component panel gates apply: E/60>=0.8 and H/E>0.7, each strong case having TG and TC interface differences>=0.1. Component eligibility requires every panel member, with equal component weighting and the same >=80% and strict>70% thresholds; non-panel siblings are listed. This operational panel is not a random estimate of full-corpus success. Failure stops this scale without redesign. Pass allows remaining geometry only, then the original209-component >=80%/>=70% gate. No fits launch from this program; fits still depend on R4bc terminal status and the original execution contract.
Cumulative prep starts631.813489 CPU-s: prior625.377883 + synthetic1.435606 + five-second metadata reserve (actual metadata0.255277). No reset or new budget: total40h = prep12h + fit24h + post4h. Four workers CPU0/2/4/6, coordinator1, each1800 CPU/wall seconds and8GiB, aggregate36GiB/100GiB, zero retries, noGPU/target decoding. Record worker CPU, guard/coordinator separately. At publication this document describes implementation/freeze; actual started/progress/complete receipts determine execution state.
R4bc read-only05:30:10: originalPID521868,609 terminal=608PASS+originaljob90, fit78129.561076 CPU-s; since05:23 +19/+1852.353374, no new failure or budget/global stop. Original queue unchanged. Old O1 v1 raw evidence7d995ec6 push completed and remote branch independently matched; not claiming all660MB independently downloaded and hashed.