Neural Geometry Lab v1.0.0

Stable research release

Neural Geometry Lab v1.0.0 publishes a frozen, reproducible test of NC1–NC4 on the official writer-disjoint UCI Optical Digits split under clean, deterministic long-tail, and 20% symmetric label-noise conditions.

Prespecified result

• NC4 stability after clean interpolation: 10/10 seeds — passed
• Imbalance worsens paired NC2: 10/10 seeds — passed
• Continued clean NC1 improvement: 7/10 seeds — did not pass the 8/10 gate
• Continued clean NC2 improvement: 7/10 seeds — did not pass the 8/10 gate
• Clean MLP median held-out accuracy: 96.27%

The release deliberately reports partial confirmation. No endpoint or gate was changed after the run.

Included

• complete machine-readable result and checkpoint registers
• full academic report, source audit, and release audit
• interactive Geometry Theatre with actual saved checkpoints
• verified data provenance and checksums
• Python 3.10–3.13 CI, 98.53% test coverage, CodeQL, dependency audit, and Pages deployment
• source distribution and universal wheel

Boundary

One compact MLP, optimizer, public dataset, writer split, and ten algorithmic seeds were studied. PCA is display-only; NC metrics use all nine hidden dimensions. Seeds are not population samples.