Upload folder using huggingface_hub

results/leaderboard.md

@@ -1,16 +1,26 @@
-# UnityReviewBench — Leaderboard
-
-Bug detection on real-world Unity gameplay code (24 bugs / 22 clean methods).
-Recall = % of real bugs flagged. Quiet = % of clean methods left alone (no false alarm). Balance = harmonic mean.
-
-| Model | Recall | Quiet (specificity) | Balance |
-|---|---|---|---|
-| Qwen3-Coder + project-context | 16/24 (67%) | 18/22 (82%) | **73** |
-| Qwen3-Coder (open) | 22/24 (92%) | 5/22 (23%) | **36** |
-| Gemini 2.5 Pro | 16/24 (67%) | 4/22 (18%) | **29** |
-| Claude Sonnet 4.5 | 11/24 (46%) | 4/22 (18%) | **26** |
-| GPT-5.5 | 24/24 (100%) | 2/22 (9%) | **17** |
-
-> Note: binary recall/specificity is *necessary but not sufficient*. See README — a judge-based
-> precision pass reveals that high recall often hides false-positive flags. Project-context (RAG)
-> dramatically improves specificity on this set.
+# Review track — a methodology finding, not a leaderboard
+
+We ran five models on 24 buggy + 22 "post-fix" Unity gameplay methods, expecting to rank them by
+recall and specificity. We are **not** publishing a balance ranking — here's why.
+
+| Model | Recall (real bug flagged) | Flag rate on "post-fix" methods |
+|---|---|---|
+| GPT-5.5 | 100% | 91% |
+| Qwen3-Coder (open, no context) | 92% | 77% |
+| Gemini 2.5 Pro | 67% | 82% |
+| Claude Sonnet 4.5 | 46% | 82% |
+| Qwen3-Coder + project context (RAG) | 67% | 18% |
+
+A naive harmonic-mean "balance" would rank GPT-5.5 last. **That is wrong.** We hand-read every flag
+GPT-5.5 raised on the "post-fix" methods: they are overwhelmingly **real, reasonable review issues**
+(unchecked return values, missing null guards, non-atomic reward grants, missing idempotency,
+`Resources.Load` null, `OnDestroy` on scene-unload) — not false positives. GPT-5.5 even returned
+`CLEAN` on the two methods with no residual issue, so it isn't flagging blindly.
+
+The ground truth is the problem: a "post-fix" method is just the version where *one* bug was fixed,
+not a method with *zero* issues — real shipping game code always has residual concerns. So
+"specificity" punishes the most thorough reviewer, and a balance score built on it is upside-down.
+
+**Takeaway:** binary recall/specificity is the wrong instrument for code-review quality. The right
+one is a per-flag verdict (real / nitpick / false-positive); see the repo README for the full
+methodology. Treat this track as a case study in why bug-detection benchmarks are hard.