PROTOCOL.md

# Fail-closed reproduction protocol

1. Lock OpenReview `X8evkEdMxb`, arXiv `2602.04761`, the exact three official base claims, six official anchors, target Space, tags, artifact Bucket, challenge revisions, live absence, PDF, source archive, and every extracted TeX-family file.
2. Treat the absence of an official code repository as requiring an independent implementation from displayed equations.
3. Transcribe Table 1 and Theorems 1–4 as explicit dimension profiles. Compare only shared normalized factors; never invent hidden big-O constants.
4. Check the two-point coordinate estimator by exhaustive coordinate expectation on quadratic functions.
5. Transcribe the one-point hyper-rectangle estimator with componentwise optimism as confirmed by Appendix equation E2. Exhaustively average all coordinate/sign outcomes and compare with the exact linear loss vector.
6. Execute the Appendix practical solver, not the causal shorthand in the main-text pseudocode. Require strict box feasibility, `|eta*c_i|<1`, and small equation residuals.
7. Execute dynamic two-point meta/base traces without providing path length to the learner. Require finite traces, box feasibility, exact simplex weights, and separated path/variation regimes.
8. Build the universal strong-convexity grid and require factor-two coverage for 4,096 lambda values, plus convex and linear bases.
9. Run destructive controls: remove one-point optimism, replace the barrier perturbation by a fixed radius, erase dynamic path length from the theorem shape, sparsify the curvature grid, and reject a false linear-dimension improvement.
10. Require all science gates, independent tests, exact replay, deterministic poster QA, genuine local Trackio artifact execution, judge-parser ingestion checks, privacy checks, and manifest integrity.
11. Immediately before any remote write, refresh target/tag/verdict absence and obtain a coordinator-assigned rolling slot.

The evidence is CPU-only. GPU/MPS and remote writes are fail-closed.