corrected proxy-gate numbers (crosswalk fix) + reproduced q4 baseline

README.md

@@ -137,14 +137,16 @@ _⚠ in OUR backbone's pretraining — contaminated for us; reference only_
 
 ## ShieldFlow proxy-traffic product gate (native ShieldFlow taxonomy, partial)
 
-Held-out 10% of the v7 proxy corpus (enterprise pastes); trained on the same 90% the
-incumbent did, evaluated on the identical untouched holdout.
+Held-out 10% of the v7 proxy corpus; we train on the same 90% the incumbent did and
+evaluate on the identical untouched holdout (seed-42 split). q4 baseline reproduced
+exactly in this harness (validates the scoring path).
 
-| dataset | ours (v4) P / R / F1 | q4 incumbent |
+| dataset | **ours (v4)** P / R / F1 | q4 incumbent (reproduced) |
 |---|---|---|
-| proxy_traffic_holdout | 0.749 / 0.683 / **0.714** | F1 0.878 / R 0.982 |
-| proxy_traffic_holdout_xlong | 0.688 / 0.753 / **0.719** | F1 0.884 / R 0.940 |
+| proxy_traffic_holdout | 0.872 / 0.715 / **0.786** | 0.794 / 0.982 / 0.878 |
+| proxy_traffic_holdout_xlong | 0.793 / 0.767 / **0.780** | 0.834 / 0.940 / 0.884 |
 
-v4 adds the v7 train split + 3 native ShieldFlow types (`developer.login_credentials`,
-`developer.device_id`, `healthcare.health_plan_id`). Recall gap to the in-distribution
-incumbent remains; tuning the v7 mix weight is the lever.
+v4 has **higher precision** than the incumbent; the F1 gap is **recall** (q4 is tuned for
+near-total recall on this workload). Scored in native ShieldFlow space with these new
+types: `developer.login_credentials` (combo span), `developer.device_id`,
+`healthcare.health_plan_id`.