Historical validation record
The original record below is preserved as historical evidence. In the upstream
Training Dynamics repository it was committed on 2026-09-25 at
e79a0b99f3b4b7608ce76655cc1a4c750455a795; the original execution date is not
specified. Its test counts and statement about uncommitted changes describe
that preparation, not the current checkout. The accompanying VALIDATION.json
and lowercase validation.json, where present, also retain their historical
scope. Current execution requires a new release-gate record tied to exact bytes.
Scientific companion validation
- root: 32 passed, 0 documented fixture skips; 4 additional subtests.
- model: 596 passed, 0 documented fixture skips; 0 additional subtests.
- row: 885 passed, 1 documented fixture skips; 0 additional subtests.
- rg: 117 passed, 2 documented fixture skips; 78 additional subtests.
All 170 owned Lean modules compiled freshly. The 485 mapped declarations resolve, and both axiom-gate negative controls are rejected. Exported Lean sources are byte-identical to the compiled sources. These are selected formal checks.
The 2,081 public records cover 323 table and 76 figure environments. Comparison preserves 5,873,448 scientific numerical and Boolean fields exactly. Eight relocated checks pass with original workspaces blocked, including the negative control.
The statement guide covers 263 numbered results and Appendices A–F. The exact-arithmetic example is illustrative mathematics. It does not estimate model sensitivities.
Native cache-position documentation is checked without altering execution. Remaining dependency and injected-test process diagnostics are documented in the code companion. No new scientific acquisition was run. The thirty-cell cache reconstruction remains retained evidence; raw reanalysis requires separately identified inputs.
Public changes are unstaged and uncommitted. Prepared file manifests identify the working contents independently of the clone heads.