Publish the clean-split retrain as revision clean-2026-08-21, and say which revision the numbers describe

README.md

@@ -41,8 +41,8 @@ language:
 >    documents during training, so no honest score for *them* exists. Publishing
 >    the clean adapter as a new revision is pending.
 > 2. **This was the harshest of the three corrections, and the reason is volume.**
->    Document isolation cut training from 33,252 pairs to **3,052** — 91 % of the
->    pool — because a single Act yields hundreds of same-document pairs. The two
+>    Document isolation cut training from **32,432** pairs to **3,052** — 90.6 % of
+>    the pool — because a single Act yields hundreds of same-document pairs. The two
 >    runs are not remotely volume-matched, so the drop from +48 % to +28.1 % is
 >    *not* a measure of "what the leak was worth". The honest finding is that
 >    **this corpus does not contain enough distinct Acts to support both a clean
@@ -127,6 +127,31 @@ string matching. It no longer "more than doubles"; that phrasing belonged to the
 figure. By pair kind, the gain holds across both dominant types: adjacent-section
 0.185 → 0.242 (n = 1,488) and heading↔section 0.172 → 0.205 (n = 489).
 
+## Which weights should I use?
+
+Two revisions are published. They share the recipe and differ in **what can be said about
+them**:
+
+| Revision | Trained on | Has a valid score? |
+|---|---|---|
+| `main` (default) | 32,432 pairs — the full mined pool | ❌ **No.** It trained on the evaluation documents, so no clean held-out set exists for it *within this corpus* |
+| `clean-2026-08-21` | 3,052 pairs — document-isolated split | ✅ **Yes** — every figure in *Results* above is this adapter |
+
+```bash
+hf download quanfire-ai/embed-statute-en --revision clean-2026-08-21 --local-dir embed-statute-en-clean
+```
+
+**Which one to take.** If you need a number you can cite or audit, take
+`clean-2026-08-21` — it is the one the Results section describes. `main` saw
+10x the training data and may well be the stronger retriever in practice, but "may
+well be" is precisely the kind of claim this card no longer makes.
+
+**We have not swapped the default**, and the reason is worth stating: doing so would trade a
+plausibly-stronger model for a measurable one with no evidence that the trade is good.
+Settling it properly needs a head-to-head of the two adapters on a corpus *neither* of them
+trained on. That is planned, and until it runs, both revisions stay up and this section stays
+honest about which is which.
+
 ## Usage
 
 Pull the adapter and run it through the Quanfire framework, which applies the LoRA over