Judge v4: fast kills counted; detector reference to match

.gitattributes

@@ -34,3 +34,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 judge/judge-v3.Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text
+judge/judge-v4.Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text

README.md

@@ -22,7 +22,7 @@ to fetch them by hand.
 |---|---|
 | `detector/scorer.onnx` | a 1D CNN that scores each kill's aim trajectory (136 KB) |
 | `detector/scorer.json` | its architecture and the frozen clean-player reference a score is read against |
-| `judge/judge-v3.Q4_K_M.gguf` | Qwen3.5-4B fine-tuned (QLoRA) to write a verdict from the evidence, Q4_K_M |
+| `judge/judge-v4.Q4_K_M.gguf` | Qwen3.5-4B fine-tuned (QLoRA) to write a verdict from the evidence, Q4_K_M |
 
 ## What they are for
 
@@ -37,11 +37,13 @@ kills in the training data; it is evidence to examine, not proof.
   per-player ROC-AUC 0.93 in match-grouped cross-validation. Line of sight comes
   from ray casts against each map's collision mesh, not the game's spotting flag,
   which is biased against snipers.
-- **Judge (v3):** fine-tuned on generated verdicts whose targets depend only on
+- **Judge (v4):** fine-tuned on generated verdicts whose targets depend only on
   evidence shown in the text (never on the ban label), with clean players' 95th
-  and 99th percentiles printed beside every measurement. On held-out cases it
-  matches its targets 87% of the time, accuses 1.9% of clean players, and cited
-  no number absent from the evidence in 206 answers.
+  and 99th percentiles printed beside every measurement. On 206 held-out cases it
+  matches its targets 92% of the time, convicts 29% of banned players, accuses
+  2.8% of clean-labelled ones (each where the evidence itself is past the clean
+  99th percentile), and cites no number absent from the evidence. On 341 pro
+  players from 35 HLTV matches it accused none.
 
 ## Credits and licences
 

detector/scorer.json

@@ -133,7 +133,7 @@
         "corrections_mean": 1.4285714285714286,
         "wall_aim_share": 0.12626262626262627,
         "visible_share": 0.8205128205128205,
-        "reaction_ms_min": 250.0,
+        "fast_kills": 0.0,
         "angle_at_first_visible_median": 3.9900760650634766,
         "snap_max": 63.369140625,
         "zero_motion_share": 0.31338862559241704,
@@ -157,9 +157,9 @@
           0.38680926916221037,
           0.06666666666666667
         ],
-        "reaction_ms_min": [
-          15.625,
-          0.0
+        "fast_kills": [
+          1.0,
+          3.0
         ],
         "angle_at_first_visible_median": [
           1.6944282054901123,

judge/judge-v4.Q4_K_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ef3824fed200541f456e3ebfc1aaa4d40533797255f4c31b583ccb14d697068
+size 2783446976