Card: pointer to the document-type head preview (branch doc-head-v1)

README.md

@@ -202,6 +202,26 @@ it was evaluated in an earlier run and cost more accuracy than it added.
 Training stack: `transformers==4.57.1`, `tokenizers==0.22.1`, `torch==2.9.0a0`,
 `faker==37.12.0`; two epochs.
 
+## Document-type classification head (preview)
+
+A document-type head is available on the branch
+[`doc-head-v1`](https://huggingface.co/ScienceSoft/scnsoft-pii-encoder/tree/doc-head-v1). It adds a second output,
+`doc_logits`, to the same graph. The head reads 450-character windows and estimates what kind of document a text
+is: a contract, a proposal, source code and so on. This revision (`main`) is unchanged.
+
+- **The PII output is byte-identical** to this revision's, on every row of the PII release check (int8 and fp32,
+  onnxruntime 1.19.2 and 1.24.2).
+- **What the head decides.** `tech.source_code` is the one type the head decides on its own. On a held-out set in
+  seven languages it scores F1 0.845 (seed 1234) and 0.848 (seed 4321), against 0.614 for a rule-based classifier. It
+  flags no benign item in any language, not counting 55 benign items that themselves contain source code; counting
+  them, one seed flags 1 of 18,704. Seven more types carry advisory scores, which a rule-based classifier decides in
+  the product.
+- **Languages.** Trained and evaluated in English, German, French, Italian, Spanish, Russian and Arabic.
+- **Try it in the browser:** [vadimbelsky/scnsoft-pii-encoder-demo](https://huggingface.co/spaces/vadimbelsky/scnsoft-pii-encoder-demo).
+  It shows PII entities and document types together, for text that stays on your device.
+
+The branch's card has the evaluation, the decision rule, usage code and the limitations.
+
 ## Licence
 
 Proprietary to ScienceSoft. The weights are published so that the results on this page can be