Random Reshuffling as a Baseline Policy — v1.0.0-candidate

Release notes: v1.0.0-candidate

Release date: 2 August 2026.

Scope

This is the first anonymous public candidate release of Random Reshuffling as
a Baseline Policy: Exact Value Certificates and Common-Hessian Ordering
Hardness. It contains one canonical paper and a replication package for its
exact algebraic, bounded executable, and declared Lean claims.

Headline results

• exact order-resolved terminal-value geometry for common-Hessian convex
  quadratics;
• an exact rational four-component family in which two ideal ordering proxies
  are arbitrarily suboptimal for terminal loss, with the encoded theorem core
  formalized under pinned Lean 4/mathlib;
• a same-random-reshuffling conditional-value policy with a realized decrement
  certificate and expected noninferiority to the same RR law;
• for an instance-encoded stable step that can approach 2, NP-completeness via
  PARTITION of the scalar common-Hessian final-loss threshold problem and,
  unless P=NP, no deterministic polynomial-time approximation of SCFLO-OPT
  within any fixed finite factor; no strong, fixed-step, randomized, or
  average-case hardness is established; at a fixed stability margin the
  complement is normalized additive, not multiplicative; and
• a preregistered 32-instance matrix-free pilot with a negative/mixed result:
  no eligible degree passed every feasibility stratum.

Release evidence

The release carries six files: the deterministic tagged-tree ZIP, anonymous
PDF, complete release replay transcript, detached machine-readable release
certificate, certificate SHA-256 sidecar, and SHA256SUMS.txt. GitHub and the
Zenodo record at version DOI 10.5281/zenodo.21757895 carry byte-identical
copies. A post-publication PUBLICATION_READBACK.json on main records public
identifiers, hashes, and unauthenticated cross-host verification without
altering the immutable release tag.

Assurance boundary

This is an unrefereed computer-assisted candidate, not an accepted or
independently verified paper. Only the declared Theorem 2 core is formalized.
The remaining theorems are conventionally proved and locally checked. Replay,
hash equality, internal adversarial review, a DOI, and publication do not
establish novelty, independent reproduction, external specialist acceptance,
peer review, or neural-network performance.

License

Original rights held by the repository maintainer are dedicated under
CC0-1.0, subject to the exclusions in PUBLIC_DOMAIN.md.