Card: int8 calibration results

README.md

@@ -34,6 +34,7 @@ evaluated on real, hand-labelled sessions.
 | Path | What |
 |---|---|
 | `model.safetensors`, `config.json`, `tokenizer*` | Fine-tuned weights (fp32), for `transformers` |
+| `new_task_split.json` | Recommended int8 threshold and input settings (read by the New-Task-Split plugin) |
 | `onnx/model_int8.onnx` + tokenizer | Dynamic int8 quantisation (onnxruntime `quantize_dynamic`), 151 MB, for CPU inference |
 
 ## Input format
@@ -101,9 +102,16 @@ The weakest cases are a message that looks self-contained but continues the work
 (`counterfactual_continue`, 0.896) and a different task in the same workspace (`same_workspace`,
 0.867).
 
-**int8 ONNX:** on a 40-row spot check it classified as well as fp32 at threshold 0.5, but its
-probabilities differ from fp32 by 0.14 on average (up to 0.62). **Calibrate the threshold on the
-int8 model itself**; don't reuse the fp32 threshold. No latency numbers yet.
+**int8 ONNX** (validation n = 697, test n = 608; tokenised with `tokenizers`, identical ids to training):
+
+| | Accuracy | AUC-PR | `new_task` recall @ precision 0.95 |
+|---|---|---|---|
+| validation | 0.937 | 0.988 | 0.925 |
+| test | 0.929 | 0.982 | 0.888 |
+
+int8 probabilities sit lower than fp32 (mean shift 0.14), so **use threshold 0.76 for int8**, not the
+fp32 value. At the validation threshold of 0.751, test precision is 0.941 and recall 0.893. This is recorded in
+`new_task_split.json`. Session load takes about 1.2 s on CPU; per-call latency has not been measured yet.
 
 ## Intended use and limitations