Clarify score semantics and keep stable public model binding

README.md

@@ -37,7 +37,7 @@ It is a **ranking model** trained on a frozen public bioassay dataset built from
 
 - The app shows a **priority band** and a **list-relative score** first.
 - Those values explain the ranking better than the raw model score.
-- The raw score is **not** a probability. Use it only for debugging.
+- The raw score is **not** a probability. It is an uncalibrated ranking value from the scorer head.
 - The strongest molecule in your submitted list will be near the top of the `0–100` relative scale.
 
 ## How To Use It

app.py

@@ -290,7 +290,7 @@ def _build_summary(query_text: str, valid_rows: list[dict[str, Any]], invalid_ro
         chunks.append(f"- Warning: {warning}")
     chunks.append("")
     chunks.append(
-        "Use the **priority band** and **list-relative score** first. The raw model score is only a debugging value. "
+        "Use the **priority band** and **list-relative score** first. The raw model score is an uncalibrated logit-like ranking value. "
         "A candidate with `relative score 100` is the strongest item in your submitted list, not in all chemistry."
     )
     return "\n".join(chunks)
@@ -535,7 +535,8 @@ Use structured assay fields when possible. Missing fields are allowed, but speci
   - `Middle pack`
   - `Low priority`
 - **relative_score_100** rescales the submitted list so the strongest candidate is near `100` and the weakest is near `0`.
-- **model_score** is the raw internal score. It is useful for debugging, not for scientific interpretation.
+- **model_score** is the raw internal ranking score. It behaves like a logit-like utility value, not a probability.
+- If you need a normalized shortlist view, treat the model score as a list-relative ranking signal and rescale within your submitted list.
 - **mol_wt / logp / tpsa** are quick chemistry context columns so you can sanity-check what the model surfaced.
 
 ### Good input habits