d1 FP16-weight browser-decision derivative

Original model and model implementation: Liquid AI, LiquidAI/d1-omni-600M,
revision 02b55d7076f15129e59ab3f94783f32c4b088674.
Original license: LFM Open License v1.0. The exact upstream license is retained
as LICENSE. Liquid AI does not endorse this experimental derivative.

MODIFIED MATERIALS NOTICE
This package is derived from the WebBrain FP32 release
webbrain-one/d1-browser-decision-fp32 (revision 71866f0ccab0b9fa92c6df64d1c34eacf2a7b2b8), whose
notice describes the earlier changes: audio components removed, decision head
and vision projector adapted, one rank-4 LoRA merged into four encoder Q/V
weights.

For this package, every FP32 weight tensor with at least 4,096 elements in the
three ONNX graphs was rounded once to FP16 and is cast back to FP32 where it is
used (scripts/models/d1/convert_fp16.py in the WebBrain repository, copied to
provenance/). The token-embedding table is gathered in FP16 and only the
gathered rows are cast. Activations, attention, softmax and accumulation stay
FP32. No training, merge, selection or re-export from PyTorch was performed.
onnx/decision.onnx, onnx/decision.data, onnx/vision.onnx, onnx/vision.data,
onnx/projector.onnx and onnx/projector.data are modified derivatives.

The immutable config.json retains a legacy dtype=float16 label from the
upstream source; package-manifest.json is authoritative (FP16 weights, FP32
arithmetic). Tokenizer, calibration temperatures and answer semantics are
unchanged.

This derivative is experimental. Synthetic labels/results and finite runtime
parity do not establish reliability on real independently developed websites.
No blanket commercial permission is granted here; consult LICENSE, including
its annual-revenue threshold terms.

Third-party runtime code retains its separate upstream notices under licenses/.