PROTOCOL.md

# Frozen protocol

1. Lock the official challenge revision, exact three base claims, five anchored claims, arXiv PDF/source bytes, target Space, Bucket, and case-sensitive paper tag.
2. Parse the paper's displayed theorem, algorithm, and feedback-rate ledger from the pinned TeX source.
3. Generate nonnegative differentiable non-monotone continuous DR-submodular quadratic families on down-closed boxes.
4. Test the `1/e` upper-linearization inequality on at least 65,536 independently generated `(f,x,y)` cells using 128-node Gauss–Legendre integration.
5. Exercise BQND on at least 1,048,576 seeded draws, require exactly one first-order query per draw, and compare its empirical mean with an independent quadrature integral for `grad F`.
6. Reproduce the source comparison ledger and feedback-specific regret exponents. Treat the official phrase “`O(T^{1/2})` ... across all feedback models” as falsified if any source-locked feedback exponent is greater than `1/2`.
7. Require claim-local destructive controls, independent tests, byte-identical replay, deterministic poster QA, genuine offline Trackio artifact staging, judge-ingestion checks, privacy checks, manifest integrity, and the shared release-safety gate.
8. CPU only. GPU and MPS are forbidden. No remote write occurs until the coordinator assigns a rolling slot and a fresh live absence audit passes.