Add combined Perch Lint Bench results

README.md

@@ -74,6 +74,18 @@ _Sorted by F1 at 0.5. Bold marks the best quality metric in each column; fewer F
 
 Inference cost is for scoring the method sides in that table. Jev uses API-reported billable input tokens at [TypeSafe's published $0.042 per million input tokens](https://typesafe.ai/blog/introducing-system-one-models-and-jev); output tokens are free. CLM is an estimated share of 11 minutes 49 seconds of active Runpod GPU time at $0.59/hour, allocated by input tokens; model startup and idle time are excluded. JevK5 and jev-lite ran on a local Mac Studio, so no metered GPU cost was recorded for them. Costs are not ranked.
 
+## Perch Lint Bench
+
+[Perch Lint Bench](https://huggingface.co/datasets/perchscan/perch-lint-bench-public) tests whether a model can apply a concise user-written semantic lint rule to a method. The same rule is scored independently against the method before and after a fix. The benchmark combines bug and security source pairs into one score; the [private partition](https://huggingface.co/datasets/perchscan/perch-lint-bench-private) is disjoint by source repository.
+
+| Model | Partition | Pairs | AP | Precision at 0.5 | Recall at 0.5 | F1 at 0.5 | TP | FP | FN | Pair order |
+| --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) | Public | 214 | 95.0% | 94.0% | 73.8% | 82.7% | 158 | 10 | 56 | 96.3% |
+| [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) | Private | 71 | 94.3% | 90.0% | 76.1% | 82.4% | 54 | 6 | 17 | 94.4% |
+| [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) | Combined | 285 | 94.8% | 93.0% | 74.4% | 82.7% | 212 | 16 | 73 | 95.8% |
+
+The rule audit reviewed all 91 original Jev misses and 30 detected controls without showing the model score to reviewers. Fourteen pairs lacked enough support in the supplied code and were excluded; 15 rules were clarified. The benchmark includes the [runner](https://huggingface.co/datasets/perchscan/perch-lint-bench-public/blob/main/benchmark.py) and scored all 570 method sides at the fixed 0.5 cutoff. This supplied-rule task is separate from the broad bug and security discovery results above.
+
 ## Files
 
 - [`benchmark.py`](https://huggingface.co/datasets/perchscan/benchmark-results/blob/main/benchmark.py): run a model on the public pairs and calculate the metrics in these tables.