d1 FP32 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
The original audio components were removed. The decision head and vision
projector were adapted in a prior, completed synthetic-browser experiment.
Four encoder Q/V weights include one FP32 merge of the selected rank-4 LoRA
delta. The vision tower remains byte-identical to the no-audio original-A
reference. The immutable selected checkpoint is projector_head_text_lora,
validation-selected epoch 06. No new training, merge, quantization, checkpoint
selection or ONNX export was performed for this release package.

model.safetensors, onnx/decision.onnx, onnx/decision.data,
onnx/projector.onnx and onnx/projector.data are modified derivatives of the
upstream model. They are copied byte-identically from the verified experiment.
The onnx/vision.onnx and onnx/vision.data files retain the accepted original-A
vision graph bytes. Modified-file notices are supplied as sidecars so the
verified binary bytes are not altered.

The immutable config.json retains a legacy dtype=float16 label. Actual packaged
weights and ONNX arithmetic are float32; package-manifest.json is authoritative.
Tokenizer, calibration temperatures and answer semantics are unchanged.

Source, selected delta, logical tensor-state identity and physical checkpoint
file identity are separately recorded in provenance/selection.json,
provenance/model-identity.json and package-manifest.json.

This derivative is experimental. Synthetic labels/results and finite browser
runtime parity do not establish reliability on real independently developed
websites. In particular, all seven positive completion examples in the older
synthetic held-out suite remain missed. 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/.
Microsoft ONNX Runtime is MIT-licensed; Hugging Face Transformers.js is
Apache-2.0-licensed. These dependency licenses do not replace the model license.