webbrain-one/d1-browser-decision-fp32

d1 Browser Decision FP32

An audio-free derivative of LiquidAI/d1-omni-600M. It directly scores named options for choice, noul and ordinal score questions using text/JSON state and optional screenshots. It does not generate text or tokens. Do not call generate() or treat it as a causal chat model.

The reported quality data are authored synthetic browser pages, with AI-authored labels independently checked through AI DOM/pixel review, not human-expert annotations. Actual Chrome/WebGPU runtime parity was measured separately; that does not establish reliable completion detection on independently developed websites.

Immutable Candidate

The published bytes are the already validation-selected projector_head_text_lora epoch 06 from the balanced-V3 experiment. The lock was written before held-out model evaluation. Packaging performed no training, new selection, weight transfer, merge, quantization or ONNX re-export.

• Source revision: 02b55d7076f15129e59ab3f94783f32c4b088674.
• 380 clean FP32 tensors: text encoder, decision head, vision tower and projector; no audio or LM head.
• Prior adaptation updated 31 head tensors, four projector tensors and four encoder Q/V matrices through one rank-4, alpha-8 LoRA merge. The other 341 tensors, including all 197 vision-tower tensors, remain exact original-A values.
• Physical model.safetensors SHA-256: 75ab6d7d0ec2966c969a95c548b91f4015b07cba82fa839fcfb5c19b40e9f940 (1,899,912,876 bytes).
• Logical clean-tensor-state SHA-256: 76409dd958673e2028f1da23a909033876169f89603c16c7cbcdfbcc7404cdb5. This is not the file SHA-256.
• Selected delta SHA-256: 82b6844524adf1ff1f7add1c9ef57475af5fcfa07ada10b9edf70c4d24c1ba35.
• Selection-lock SHA-256: 4ab2441d9b3cf7742957a374988fc50fc400b29081b7c38b6920be88cd65bb19.

The three ONNX graphs and their external data total 1,900,356,386 bytes. They are full FP32, unquantized, and copied without binary changes. package-manifest.json and SHA256SUMS enumerate payload checksums; large files also have 4 MiB chunk checksums.

Config caveat: the unchanged legacy config.json says dtype: float16 and architectures: [NoAudioModel]. Actual checkpoint/graphs/feeds are float32, as explicitly required by the release manifest. Do not derive precision from the legacy label. This repository does not advertise an AutoModel.from_pretrained() loader or executable remote Python code.

Synthetic Held-Out Results

New final test: 72 screens, 144 questions (126 categorical choice/noul, 18 ordinal score). Old test: 60 screens, 120 questions (90 categorical, 30 score). All four historical stages are shown, not only the released one.

Historical stage	New categorical	Completion precision	Completion recall	Unknown correct	New score MAE	Old categorical	Old score MAE
Original A	48/126	21/59	21/24	6/28	0.736238	43/90	0.973761
Head only	51/126	11/27	11/24	9/28	0.723945	42/90	0.913887
Projector + head	93/126	22/23	22/24	23/28	0.574737	52/90	0.605605
Released projector + head + text LoRA	92/126	22/23	22/24	23/28	0.574219	52/90	0.570224

The released stage has one false completion among 48 non-complete/uncertain new-test cases: 0/24 known negatives and 1/24 uncertain cases. It has two missed new-test positives. All four stages miss all seven positive completion examples in the older test. No positives are predicted there, so old-test completion precision is undefined; zero false positives is not evidence of successful completion recognition.

The LoRA stage did not add new-test categorical accuracy over projector+head. It remains the release candidate because selection was validation-only; held-out results did not reselect it. The original unlabelled 58-request/70-question regression suite is not an accuracy benchmark: the released stage changes 9/46 choice/noul decisions versus current original A (maximum probability drift 0.734239; maximum expected-score drift 1.496044).

Limitations include synthetic layout/text/color regularities, finite family splits, overconfidence/train-versus-validation loss separation, residual Ready-identifier/completion association (0.622556 bits within sparse target groups), and joint family/option-position association (0.584963 bits). Zero conditional viewport MI in the specified QC groups is not proof of universal nuisance independence. No independently developed real-site dataset was collected for this release.

Measured Runtime Scope

The existing own-checkpoint FP32 source API was compared with desktop ONNX and actual Chrome 154 WebGPU on an NVIDIA RTX 5090, non-software adapter. Fixtures were 72 validation screenshots plus 20 neutral text requests: 92 requests/169 questions, not held-out quality examples. Each browser path had two warm-ups and ten hot repeats; all 145 categorical questions agreed on every hot repeat. Maximum absolute probability/expected-score errors stayed below the unchanged 0.001 gate.

Browser path	Max probability difference	Max expected-score difference
Same native media prefix	0.0000563264	0.000109192
Actual PNG preprocessing + vision/projector	0.0000483990	0.0000722781

Tokens, pixels, masks and shapes matched their references exactly. Position-interpolation FP32 order differences reached 0.000000774860; the trained intermediate media-prefix difference reached 0.0565567. There is no claim of bit-exact intermediate activations or a 0.001 intermediate gate.

Observed hot p50/p95 milliseconds on that machine: text decision 30.778/151.935; saved image-prefix decision 67.290/272.953; full image inference 219.943/751.239; PNG preprocessing plus inference 337.175/957.691. These finite request-balanced measurements are not throughput guarantees. Model/graph storage bytes are not VRAM consumption.

Separate three-fixture profiling observed 1,465 WebGPU nodes and 407 CPU/WASM nodes, including four floating mask/position construction nodes; this is not pure GPU execution. Separate bounded memory sampling observed adapter-total memory, not isolated model VRAM. Existing runner parity is not, by itself, proof of a newly integrated packed WebBrain extension.

Adapter Use

The package contains the dependency-injected adapter copied from the actual WebBrain extension module src/chrome/src/providers/d1-runtime.js. Its preprocessing helper preserves the old verified bytes except two explicit source-semantic corrections: present-null noul criteria (Object.hasOwn, matching Python dict.get) and small fractional Python-style JSON exponent formatting. The old baseline helper is untouched. JavaScript numbers cannot recover Python int-versus-integral-float lexical types: use preformatted state/criterion strings when exact original lexical representation matters. The adapter also restores the source 65,536 padded-token subbatch budget while returning named answers in original order. Executable JavaScript/WASM must be bundled locally for extension CSP; do not load remote executable code. Immutable-revision model graph/tokenizer downloads are data and must be checksum-verified before caching/creating sessions.

The exact API is createD1Runtime({ort, tokenizer, config, ratios, sessions, device, model}), then evaluate({state, images, questions, signal}). See runtime/README.md for session wiring. noul answers expose the source public yes-probability; score answers expose expected ordinal level, not argmax-class accuracy. Option insertion order, masks, media prefix, calibration and the image text limit of 896 tokens must remain unchanged.

import { createD1Runtime } from './runtime/d1-runtime.js';

// Supply verified package assets, the bundled ORT/tokenizer, and three FP32 sessions.
const judge = createD1Runtime({ ort, tokenizer, config, ratios, sessions, device, model: 'd1-browser-decision-fp32' });
const result = await judge.evaluate({
  state: { task: 'Check whether a visible receipt establishes completion.' },
  images: [inlinePngDataUrl],
  questions: {
    completion: { type: 'choice', instructions: 'Judge only visible evidence.', criteria: {
      completed: 'An explicit receipt confirms the named task.',
      not_completed: 'Visible evidence establishes failure or an unfinished task.',
      unknown: 'The screenshot does not establish the outcome.'
    } }
  }
});
// result.usage.output_tokens === 0; this is not text generation.

License And Notices

The model is governed by the exact upstream LFM Open License v1.0, not Apache/MIT. Its commercial-use provisions include annual-revenue threshold terms of US$10 million; determine eligibility and obtain any required separate license before commercial deployment. This card is not legal approval or an endorsement by Liquid AI.

NOTICE and modified-binary sidecars retain attribution and identify the historical derivative changes without changing verified binary bytes. The underlying LFM2.5-Encoder license reference is separately retained; that license-reference revision is not a claim about the historical base-weight revision used by d1. Bundled ONNX Runtime 1.31.0-dev.20260914-8d85527a0 is MIT-licensed; Transformers.js 4.3.1 is Apache-2.0-licensed. The actual WebBrain GPL-3.0-or-later project notice is retained for the copied runtime source at licenses/webbrain/LICENSE. Separate component notices do not replace the model license or constitute a legal compatibility/commercial-eligibility opinion.

Native WebGPU Runtime Dependency Correction

This revision retains the same checkpoint, six ONNX graph/data files, tokenizer, config, decision adapter and preprocessing helper. It adds the exact same-version ORT asyncify MJS/WASM pair and corrects the local loader example. The native WebGPU bundle calls webgpuInit; explicitly forcing the JSEP factory (jsepInit only) failed during an actual packed-extension initialization before neural inference. The earlier successful numerical browser runner used the same ORT distribution's directory-prefix loader, which selected the native asyncify pair. Existing JSEP files remain preserved but must not be selected for this native WebGPU bundle.

See runtime/README.md, runtime/vendor/vendor-manifest.json and the pinned package-manifest.json checksums. This is a dependency closure correction, not a new export, precision change or new quality result. It does not itself establish packed-extension inference success or real-site generalization. The existing synthetic evaluation, license restrictions and unverified real-site scope remain.