# Technical validation

The final refit completed 750 updates and exported a native bundle with 571,909,635 parameters in 489 tensors. The initial 64 TRAIN decisions passed same-B8 native/all-types-wrapper parity. The first update had finite gradients across all 486 active tensors and actual parameter-content changes in each of the nine encoder/head/scorer groups. The shared three tensors remained unchanged.

After releasing the training model and optimizer, a separate process loaded the final native and repeated those 64 TRAIN decisions in eight FP32 batches. Logits and probabilities matched the saved final predictions exactly, with zero hard-label flips. This was a technical reload check, not another quality evaluation.

The original release included Python inference and fine-tuning modules copied from the tested Kai distribution. Its package-isolation check loaded the bundled native and evaluated the same 64 TRAIN decisions in eight FP32 batches from an external working directory. Maximum logit/probability differences from the independent final-refit fresh reference were 0.0/0.0, with zero hard flips. All 489 parameter contents and versions were unchanged; no gradients, updates or TEST evaluation occurred. This is historical validation of those model objects and runtime; executable code is no longer part of the model-only repository. See [TECHNICAL_VALIDATION.json](TECHNICAL_VALIDATION.json).

The earlier fine-tuning CLI's bounded continuous-versus-resume and independent-reload checks are interface evidence, not new Lex training results. Lex's final training used the fixed research recipe described in [METHODS.md](METHODS.md). The evaluated runtime profile used complete packed length ≤1,024 and FP32 inference on compatible AMD ROCm. No browser end-to-end latency, CPU/NVIDIA inference, multilingual-specialist quality or larger-window quality is claimed.