GNN 5.0.1: numerical witnesses, offline inspection and bounded workers

GNN 5.0.1: numerical witnesses, offline inspection and bounded workers

GNN 5.0.1 adds source-bound numerical witnesses, interactive output inspection and configurable file workers for the coordinated release with FEP_Formal 2.1.0.

gnn witness retains source, result and generated-program bytes and compares admitted finite filtering and passive linear-Gaussian models with independent references. Admission checks parameters, axes, timing, conditioning and declared precision. Every posterior mean and full covariance remains inspectable; failed, refused and unavailable evidence retains its reason. Numerical agreement, source custody, formal compilation and authenticated execution remain separate claims.

The website adds evidence filters, state, variance and residual trace controls, complete table fallbacks and raw receipts. Current-run links and reciprocal model, site and GUI navigation preserve source identity. Recorded CPU limits and observed processing widths are displayed separately. The downloadable evidence explorer works offline and preserves the complete admitted values. Its static resource and value checks are distinct from browser testing of the original generated site. A narrow comparison-detail column still wraps heavily in that site's current layout.

Steps 3 and 5 support bounded spawned file-worker processes, defaulting to one worker. The requested limit is 1–256; Windows further caps effective process workers at 61. Independent-step concurrency retains its separate eight-step ceiling and memory throttling. Step 12 execution workers are configured separately.

gnn run --only-steps 3 5 --file-workers 4
gnn run --parallel --step-workers 2 --file-workers 4

Python 3.15.1 profiling and seven unprofiled samples on a recorded 38-file corpus measured median wall times of 0.166451 s serial, 0.162481 s with four threads and 0.226255 s with four processes, including worker startup and shutdown. This small workload favors the default single worker; it establishes no general parallel or whole-pipeline speedup. The scientific installation remains separately qualified on Python 3.14.

The release source is normal merge e54813b4197df36f7ae0054d5e79f3ad2d5db0ff, tree c81cb1848585e86c0feda51ec230f7abd3544dd5. Annotated tag v5.0.1 is object 35e71d1a6f109f6e0d08fda2acd5a94376482cf4. The FEP pin is merged revision 96bf714d91c088e79b9dea54af0f9151f70de64f. The unchanged numerical capture records candidate 855f385c2ea9a3a8ad1a6d8d17f25dea2153e1e2 and that FEP revision; the release identity does not relabel its capture epoch.

This version removes unused bibliography entry [33], whose dead URL failed the earlier tag documentation audit, and retains the strict link policy. The 32-page manuscript has SHA256 f364e37b8fd0bd4784cf27533768acb3a4a78004baf018d928da2126ef98d83b. The original v5.0.0 tag and failed observations remain immutable and unpublished. A classified CodeQL service-processing stall on the candidate was retried once under the documented infrastructure policy; the original attempt is retained, and the new attempt has its own authenticated completed analysis.

The release tag's CodeQL analysis retains one request-path result (alert 12). Inline static review classified that exact result as not actionable: the reported path construction precedes lexical containment, parent-traversal rejection, redirect checks and resolved containment, and the cited API routes reject admission failures before backend use. The raw nonzero result and source-specific disposition are preserved. This is a static assessment under the documented trusted-filesystem contract; it does not establish protection against hostile concurrent filesystem changes or an exhaustive security clearance. No alert dismissal or source change was performed for this disposition.

The ten assets comprise the ordinary wheel and sdist, manuscript, offline evidence explorer, exact Git source ZIP, citation, license, source identity, verification record and checksums. The broader seven TODO workstreams remain open at their documented scientific, provider and archival boundaries. Finite/scalar Lean laws and Q7 static coefficient bounds do not prove parsing, execution, general matrix filtering or universal translation equivalence. The reserved FEP manuscript DOI 10.5281/zenodo.23607007 is not yet a published archival receipt.

GEO-INFER 0.4.0 consumer compatibility. GEO-INFER 0.4.0 records qualified consumer acceptance with published GNN 5.0.1 and FEP_Formal 2.1.0 at GEO source 71ce1f77418c6f8c6edc8f83c7682c8eab207e01; its tag CI and asset checksums accompany the release. This consumer result preserves GNN 5.0.1's published GEO baseline 21b65a26dd15288d6e578baef86a8bcfd6a7b5db and the existing FEP_Formal proof epoch; the source/tag/asset identities, native evidence and scholarly DOIs recorded above remain unchanged.