# Lex training method

Decision-1.0-Lex is an English typed-decisions specialist fine-tuned from the fixed Kai native `da603662bc57e89ccfb51c972ed9c1f2825f267597353cf1337df9117a3dfabe`. Its 571,909,635-parameter geometry is unchanged. All three type paths are trained: 486 tensors across three encoders, interaction heads and scorers; the shared token embedding, embedding normalization and type embedding remain frozen.

Development used a state-component split of the original official TRAIN: 4,800 decisions for training and 1,200 for development. The fixed selected recipe was blend50 at the higher learning-rate setting. Those development scores are distinct from the final TEST result. The final refit then combined all 6,000 original TRAIN decisions, trained from the same Kai parent for eight epochs, and exported fixed step 750 without checkpoint selection. Every decision appears exactly eight times; epochs are concatenated before dividing into logical batches, so no tail is dropped or repeated.

For the original normalized soft label vector q and original semantic hard-label one-hot h, the target is 0.5q + 0.5h. The loss is cross-entropy summed over 64 decisions and divided by 64, with no RPS, consistency, distillation or calibration fitting. Candidate order and semantic descriptions are retained. Each logical batch uses eight physical batches of eight. Training uses BF16 autocast with FP32 parameters and loss, AdamW (weight decay 0.01), gradient clipping at 1, and seed 20260921. There is no warmup. Encoder/head starting learning rates are 2.5e-5/1e-4 and each follows `1e-6 + (base - 1e-6) * (1 + cos(pi * (step - 1) / 749)) / 2` for steps 1–750.

All complete packed inputs fit 1,024 tokens. The final native manifest is `f288d873999832a3f37c6a7c4268c2ab309691e621794dbf7acab891acbbb7e6`. Native metadata retains the original research export provenance; its native Score training-policy metadata is an inference/export compatibility contract, not a claim that only Score was trained. The actual all-types update is recorded in the export provenance and [TRAINING_PROVENANCE.json](TRAINING_PROVENANCE.json).

The final TEST was opened only after this checkpoint and recipe were frozen. It was evaluated once against the official typed specialist on identical complete inputs. Neither TEST checkpoint selection nor temperature fitting was performed. [EVALUATION.md](EVALUATION.md) records the point improvement and its limits. This package does not transfer Kai's multilingual quality or latency measurements to Lex.