Historical code, result records, and isolated replay

STATUS: RIEMANN HYPOTHESIS OPEN

Public credit: @ykbballer91. AI-assisted research log. This document is CC BY 4.0; code is MIT. Cited third-party works keep their own rights and are not bundled.

These are sanitized historical copies. The publication export did not rerun the numerical certificates, high-precision diagnostics, or Lean builds. A saved PASS describes the original check and its stated scope; it does not certify RH or a fresh installation. Source hashes in old records identify original files. Export hashes in the source manifest identify the public copies after provenance comments and link rewriting.

Inspect first

python3 artifacts/replay.py --list

The scripts use the original research/, experiments/, and formal/ layout. Their public archive locations differ, so running arbitrary scripts directly from this tree can lose a dependency or overwrite a saved result. The replay helper reconstructs the selected source layout in a new workspace outside this repository.

To inspect the reconstructed layout without executing research code:

python3 artifacts/replay.py \
  --script experiments/scripts/window_half_certificate.py \
  --prepare-only

To explicitly rerun one script, install the dependencies appropriate to that script in your own environment, then omit --prepare-only. For example:

python3 artifacts/replay.py \
  --script experiments/scripts/window_half_certificate.py \
  -- --help

Arguments after -- are passed to the selected script. The helper verifies public payload hashes, creates an isolated workspace, and never downloads papers, ancillary code, dependencies, or toolchains. New results remain in that workspace. It does not compare them automatically with historical results or declare a certificate valid.

Typical external packages appearing in the saved code are python-flint, mpmath, numpy, and scipy; not every script requires all four. Dependency installation and version pinning are separate tasks. The historical Lean sources use the saved lean-toolchain / lakefile.toml; neither Lean nor Mathlib is vendored here. See the reproducibility guide for the project-level scope.

Known limitations

• experiments/scripts/groskin_compare_assemblies.py requires excluded third-party ancillary code. It is retained as a historical comparison script, not a standalone public reproducer.
• experiments/scripts/desogus_targeted_checks.py requires the excluded third-party TeX source. The historical result is not regenerated by this export.
• research/rate_history/experiments/validate_outputs.py and research/hierarchical_selection/experiments/validate_outputs.py are old repository-preservation validators. Their private baselines and Git state are deliberately absent. They are not validators for this public repository.
• literature/build_ledger.py is a historical metadata generator, not a mathematical test. Replay execution is disabled for it.
• experiments/scripts/window_four_fifths_certificate.py records an interrupted line of work. Its presence does not mean that the larger-window certificate completed.
• Some scripts are floating-point or high-precision diagnostics, while specific earlier files record interval certificates. These distinctions are in the original reports and result fields. No blanket “all experiments certified” claim is made.
• The latest cyclicity result is an analytic theorem in ordinary even (L^2). It is not a finite-head, form-core, growing-order, full-ground, or RH certificate.

The two historical JSON result arrays and the one-line lean-toolchain retain their machine-readable root shape. Their provenance header is in the manifest rather than inserted into the file. Other JSON objects have additive _public_export metadata; original mathematical and validation fields remain historical.