sense-d

Fab7's decision model, which Weft runs on your machine. Nothing leaves the machine.

One microsoft/deberta-v3-xsmall answers every judgment Weft needs as a question with a probability. Each row is [CLS] text [SEP] question [SEP], read as one logit (yes minus no).

• Screen: the agent's terminal state, from the screen alone, in 7 states (idle, working, after its Stop hook, waiting on a permission, a question, trust or sign-in, or stopped on an error). Weft weighs this against the harness's own hooks; the model never sees them.
• Transcript entry: the person's message, the agent's reply, a tool step that worked or failed, or internal text.
• Ask: its task, its effects (read, write, execute, beyond the machine) and its route (direct, plan, goal, review). The result follows from these.
• Practice rules: a convention decides seven from the task and effects, and each other rule is a question.
• Recall: whether a changed code window does a requirement.

The questions, calibration and text normalizing are data in manifest.json (schema fab7.sense.decision/1, normalizer fab7.sense.decision.normalize/2).

Files

File	What
sense.int8.onnx	The model, with int8 weights and fixed scales. Input input_ids and attention_mask (int64[B, L], L ≤ 256); output logit (float32[B]). Run it with ONNX Runtime and session.disable_quant_qdq = 1, so each row's answer is its own.
tokenizer.json	DeBERTa-v3's tokenizer
manifest.json	Every head: its questions, classes and calibration; the normalizer; file digests
golden_vectors.jsonl	Normalized texts, token ids, logits and answers to check a runtime against

Training

• The base: microsoft/deberta-v3-xsmall, first taught entailment on the human-labelled, commercially usable pairs of MoritzLaurer/dataset_train_nli.
• The fine-tuning: Fab7's own labels:
  • screens and transcripts from sandboxed harness sessions on local models, labelled by rules and checked by readers and a reviewer;
  • Asks, rules and windows labelled by a person or from templates;
  • published per-hunk code labels.
• Nothing was distilled from another model.

Performance (0.1.1, test split, texts training never saw)

Head	Score
Screen	accuracy 0.992 (1,008 screens), from the screen alone
Transcript entry	accuracy 0.994 (529)
Ask task	macro F1 0.802 (308 Asks)
Ask effects	macro F1 0.913; risky-effect recall 0.976
Ask route	accuracy 0.886
Practice rules	Jaccard 0.688
Recall	flags 92 % of windows that miss their requirement, at 90 % precision

On a CPU, a screen read (7 rows) takes about 165 ms on an Apple Silicon Mac; on a GPU, a few milliseconds.

Changes from 0.1.0:

• The screen text no longer carries the hook state, which 0.1.0 had learned to echo.
• working_generating and working_tool are one state, working.
• A class with fewer than 20 calibration examples keeps the threshold 0.5.
• Recall flags fewer of the windows that miss their requirement: 92 %, from 98 %.