diff --git a/docs/claude/s2-execution-request-plan-2026-11-14.md b/docs/claude/s2-execution-request-plan-2026-11-14.md new file mode 100644 index 000000000..029d949b3 --- /dev/null +++ b/docs/claude/s2-execution-request-plan-2026-11-14.md @@ -0,0 +1,31 @@ +# S2: dependency order and the minimal execution requests to Yanqi (Claude, 2026-11-14) + +Purpose: stop "wait for Yanqi's GO" and "first produce evidence" from blocking each other. The evidence work itself needs computation, so it is the first thing to request. It is kept **label-free** so that nothing in it can tune against the targets. + +## Dependency graph +- **G1 grouping evidence** needs computation: fingerprints, candidates, registration, and the held-out calibration set for the blind review. → Request E1. +- **10-part preprocessing measurement:** d/t_part, ℓ_g, the PPS draw, O2 count and kernel. Label-free. → Request E1. +- **Code** for R2/N/E and the supervision, accepted on synthetic tests only. → Request E1. + - R2 feature list: waits on Grok's Gre24 mapping. It blocks only the *R2 code freeze*, not E1. +- **Pilot:** needs E1 complete (components frozen, 10-part costs, code accepted) and reads labels. → Request E2. +- **Full matrix:** needs pilot receipts and a budget computed from them. → Request E3. + +## E1 (to ask now): CPU only, label-free, no training +- **Scope:** + - (a) Geometry fingerprints for all 624 parts, two recall routes, candidate pairs and their registration residuals; a calibration subset for blind review. Thresholds are frozen after that check. + - (b) The 10-part stratified preprocessing measurement (v0.3 G5 limits). + - (c) Implementation of the R2/N/E pipelines and the supervisor, with unit and synthetic tests only. +- **Never read in E1:** Fill_time and other label fields. No model fitting. No GPU. +- **Caps:** per-part limits as in v0.3. Total CPU for (a) is set by MeltAstra's count-only estimate before it starts, with stop-and-report if exceeded. +- **Deliverables** as fixed-commit packages: + - group_manifest and duplicate_edges; + - the 10-part cost table; + - code with test receipts. + +## E2 (later): pilot +- 4 runs, at most 2 GPU-h, test fold sealed. +- Requires E1 accepted, plus the R2 freeze (or R2 explicitly labelled as an approximation). + +## E3 (later): full matrix +- Budget from pilot measurements. +- Preceded by a stage review.