Files
MeltMind/docs/phase2-b13-e2-geometry-contract-2026-11-17.md

12 lines
3.3 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# E2 geometry and first-ten execution contract
Claude1794907827 confirms E1b and starts the existing Mold182e0312 specification with Claude1794904249's three amendments. The exact test552/204 roles and original100% learning pools are inherited from E1b. G partitions/readouts, models and predictions are reused by hash. This first executable stage does geometry only: deterministic first10 in frozen case order, one cost checkpoint, then remaining542 only if projected total including postfit fits14400 CPU-s. It ends at GEOMETRY_COMPLETE_POSTFIT_PENDING, without fitting or scoring. The later postfit implementation must freeze separately and inherit cumulative geometry cost; this is not a new approval gate or extra budget.
T/C use the original O1v2 type vector, volume-decile whole-triple shuffle, true-field strength, area midrank, common integer scale, METIS5.1.0 seed0/K8. Identical original O1v2 cached seed0 partitions are reused only after option, graph/node/cell/volume and weighted-input hash checks; otherwise each T/C arm makes exactly one seed0 call. G is never repartitioned. Only actual F and Dw* enter regional readout. Every arm independently keeps original G eligibility; no cross-arm area/diameter/matching threshold and no common-valid roster filter. Invalid scientific T/C arms get only two nativeNaN columns; engineering/source/resource errors stop and retain denominators. G byte-identical readout is rechecked when valid. Exact physical cut area, eight exact hull diameters and RMS size are diagnostics; descriptor failure does not silently filter prediction parts.
Four workers8/10/12/14, coordinator30/31. Separate14400 cumulative CPU-s includes60 setup plus all geometry guards/children/coordinator and later postfit; no E1 budget transfer. Geometry per part120CPU/180wall,8GiB RSS; concurrent32GiB, directory12GiB. Zero retries. First10 projection is60+1.25*(chargedCPU−60)*(all552nodes/first10nodes)+postfit forecast. Every first10 part, including scientific missing, contributes cost/denominator. Postfit fit forecast is1.25 times measured E1b four-fit CPU scaled to two arms and all five learning-node counts; ten predictions reserve1200, adjusted statistics/scoring2400, coordination300. If first10 total projection or actual total exceeds14400, stop; no extension. Original geometry and original failed states remain.
Future deployment gate remains raw full-roster T−G: fieldCI upper<0 OR(lastCI upper<0 AND fieldCI upper<.001). Adjusted T−G/T−C/C−G are mechanism-only descriptions, using original full-roster component/part weights, two uncentered log area/RMS-size covariates, zero-point estimability/convex-hull support, and2000 component bootstrap refits seed2026111408. Missing arms remain in raw differences and the unadjusted part of adjusted estimands; >=5% failed/support-missing replicates invalidate adjusted intervals. No gate or scorer may be revised based on first10 geometry outcomes.
Three new synthetic checks passed: one invalid C arm leaves G/T valid and identical readouts retained; two T/C calls only, no G call; first10 total projection includes future stages and rejects invalid accounting inputs. Original O1v2 weighting/shuffle/geometry tests are inherited without recomputation. First10 is science, so actual start requires local immutable contract/code and host duplicate checks. Remote upload is not a start prerequisite under6657.