Packaging parity is now demonstrated; scientific accuracy remains separate
18 September 2026 · afternoon · Asia/Jerusalem
The compact global-rom-80 runner now matches the established frozen-ROM runner bitwise for one complete 513×256 FID, with a separate two-worker/serial determinism check. This is packaging and execution consistency—not exact-Spinach equivalence.
Blocker: the underlying exact spectra and validation caches remain external. The empirical 192-case accuracy summary is not recalculated by this parity run.
What changed
The dated VERIFICATION.md record reports one interior parameter case with maximum absolute FID difference 0, relative L2 difference 0, and exact array equality true. Preserved timing/file-hash receipts alone do not establish that equality.
The parity case uses the same frozen reduced model in both implementations. It does not test exact Spinach, continuous-domain accuracy, or uniform certification.
The release summary reports finite-sample empirical evidence from 192 cached exact spectra within the fixed 2700 Hz / zero-exchange range: worst peak-normalized pointwise error 6.3105e−3 and worst peak-normalized RMSE 2.8961e−4. Those values were not independently recomputed in this audit and are not a continuous-domain or uniform bound.
The 4.16 s serial smoke run was measured after MATLAB startup; the roughly 1.5 s values are recorded case-compute times. Neither is a matched end-to-end benchmark, campaign throughput measurement, or speedup.
Next decisions
- Preserve the external exact-data artifact with per-file checksums and an independent copy.
- Authorize a second parity boundary case, or keep parity as a one-case packaging check.
- Keep micro-tests paused.
Literature Desk · five source-checked reads
Labels: fact, source claim, local measurement, inference, recommendation, speculation. Stable local-only bookmark; requires the dashboard server: http://127.0.0.1:8765/