Describe Perch method-pair benchmark and unify model tables

README.md

@@ -5,11 +5,13 @@ pretty_name: Perch Benchmark Results
 
 # Perch benchmark results
 
-Bug and security results are separate. The primary tables compare models given the same saved Perch method context and one broad question per method (`perch-broad-v1`). Model names link to their model cards; Jev links to TypeSafe's description because it has no public model card.
+This benchmark tests whether a model can find bugs or security issues in individual code methods. Each record contains a method before and after an open-source change, plus the input adapted from a [Perch scan](https://github.com/lakeday-org/perch): method source, imports, callers, callees, and static call-graph edges where available. Models score each version independently, without seeing the patch or the other version. The labels also include benign reference methods, so a `before` method is not automatically positive.
 
-**Columns:** AP (average precision) summarizes how well the model ranks issue-bearing methods above non-issue methods across score cutoffs. At the 0.5 cutoff, a method is flagged when its issue score is at least 50%. Precision is the share of flagged methods that have the issue; recall is the share of issue-bearing methods flagged; F1 is their harmonic mean. TP is a correctly flagged method, FP is a method falsely flagged, and FN is an issue-bearing method missed. Pair order is the share of eligible before/after pairs where the before method scores higher than the fixed after method; ties do not count as correct.
+The pairs come from open-source projects in Rust, JavaScript, TypeScript, Python, Go, Java, C, C++, C#, PHP, and other languages. Bug and security detection are scored separately. Each model is judged on whether it flags issue-bearing methods, avoids alerts on methods without the issue, and ranks an affected before method above its fixed after method.
 
-The public pairs and their per-method predictions are available here. Private pair-level data and predictions are not published. The public and private sets are disjoint by source repository across both domains.
+## Reading the results
+
+AP (average precision) measures ranking across score cutoffs. For the other metrics, a score of at least 0.5 counts as an alert. Precision is the share of alerts that are correct; recall is the share of issue-bearing methods found; F1 is their harmonic mean. TP is a correct alert, FP a false alert, and FN a missed issue. Pair order is the share of eligible before/after pairs where the before method scores higher than the fixed after method; ties do not count as correct.
 
 ## Bugs
 
@@ -21,6 +23,7 @@ The public pairs and their per-method predictions are available here. Private pa
 | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
 | [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) | 47.6% | 49.9% | 22.0% | 30.6% | 410 | 412 | 1451 | 42.5% |
 | [JevK5](https://huggingface.co/alibiserikbay/JevK5) | 44.1% | 47.5% | 7.6% | 13.1% | 141 | 156 | 1720 | 32.1% |
+| [CLM-v0.1-8B](https://huggingface.co/Contrastive-LM/CLM-v0.1-8B)† | 44.0% | 44.9% | 13.9% | 21.2% | 258 | 317 | 1603 | 46.9% |
 | [jev-lite](https://huggingface.co/vagmi/jev-lite) | 43.7% | 44.8% | 2.3% | 4.4% | 43 | 53 | 1818 | 22.4% |
 
 ### Private — 498 pairs
@@ -30,6 +33,7 @@ The public pairs and their per-method predictions are available here. Private pa
 | [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) | 48.6% | 50.7% | 24.1% | 32.6% | 102 | 99 | 322 | 43.4% |
 | [JevK5](https://huggingface.co/alibiserikbay/JevK5) | 45.8% | 46.4% | 9.2% | 15.4% | 39 | 45 | 385 | 32.7% |
 | [jev-lite](https://huggingface.co/vagmi/jev-lite) | 45.2% | 60.0% | 3.5% | 6.7% | 15 | 10 | 409 | 22.7% |
+| [CLM-v0.1-8B](https://huggingface.co/Contrastive-LM/CLM-v0.1-8B)† | 45.2% | 44.6% | 17.5% | 25.1% | 74 | 92 | 350 | 43.4% |
 
 ## Security
 
@@ -41,6 +45,8 @@ The public pairs and their per-method predictions are available here. Private pa
 | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
 | [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) | 49.2% | 54.4% | 10.2% | 17.2% | 117 | 98 | 1025 | 40.5% |
 | [JevK5](https://huggingface.co/alibiserikbay/JevK5) | 44.8% | 46.4% | 12.4% | 19.6% | 142 | 164 | 1000 | 30.6% |
+| [CLM-v0.1-8B](https://huggingface.co/Contrastive-LM/CLM-v0.1-8B)† | 44.0% | 43.6% | 64.9% | 52.1% | 741 | 959 | 401 | 41.5% |
+| [SecureBERT2.0](https://huggingface.co/cisco-ai/SecureBERT2.0-code-vuln-detection)† | 43.6% | 43.5% | 99.8% | 60.6% | 1140 | 1479 | 2 | 50.2% |
 | [jev-lite](https://huggingface.co/vagmi/jev-lite) | 43.3% | 36.0% | 3.2% | 5.8% | 36 | 64 | 1106 | 22.9% |
 
 ### Private — 439 pairs
@@ -50,25 +56,16 @@ The public pairs and their per-method predictions are available here. Private pa
 | [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) | 48.9% | 52.6% | 2.5% | 4.8% | 10 | 9 | 388 | 36.2% |
 | [JevK5](https://huggingface.co/alibiserikbay/JevK5) | 47.4% | 45.0% | 4.5% | 8.2% | 18 | 22 | 380 | 37.9% |
 | [jev-lite](https://huggingface.co/vagmi/jev-lite) | 47.3% | 43.8% | 1.8% | 3.4% | 7 | 9 | 391 | 29.4% |
+| [CLM-v0.1-8B](https://huggingface.co/Contrastive-LM/CLM-v0.1-8B)† | 47.1% | 46.9% | 67.8% | 55.4% | 270 | 306 | 128 | 52.8% |
+| [SecureBERT2.0](https://huggingface.co/cisco-ai/SecureBERT2.0-code-vuln-detection)† | 43.0% | 45.4% | 99.7% | 62.4% | 397 | 478 | 1 | 55.8% |
 
-## Different input protocols
-
-These models received their native decision inputs, so their scores are reported separately from the prompt-matched tables above.
-
-| Domain | Track | Model | Protocol | AP | F1 at 0.5 | TP | FP | FN |
-| --- | --- | --- | --- | ---: | ---: | ---: | ---: | ---: |
-| Bug | Public | [CLM-v0.1-8B](https://huggingface.co/Contrastive-LM/CLM-v0.1-8B) | `clm-native-choice-6144-v1` | 44.0% | 21.2% | 258 | 317 | 1603 |
-| Bug | Private | [CLM-v0.1-8B](https://huggingface.co/Contrastive-LM/CLM-v0.1-8B) | `clm-native-choice-6144-v1` | 45.2% | 25.1% | 74 | 92 | 350 |
-| Security | Public | [CLM-v0.1-8B](https://huggingface.co/Contrastive-LM/CLM-v0.1-8B) | `clm-native-choice-6144-v1` | 44.0% | 52.1% | 741 | 959 | 401 |
-| Security | Public | [SecureBERT2.0](https://huggingface.co/cisco-ai/SecureBERT2.0-code-vuln-detection) | `perch-security-state-4096-v1` | 43.6% | 60.6% | 1140 | 1479 | 2 |
-| Security | Private | [CLM-v0.1-8B](https://huggingface.co/Contrastive-LM/CLM-v0.1-8B) | `clm-native-choice-6144-v1` | 47.1% | 55.4% | 270 | 306 | 128 |
-| Security | Private | [SecureBERT2.0](https://huggingface.co/cisco-ai/SecureBERT2.0-code-vuln-detection) | `perch-security-state-4096-v1` | 43.0% | 62.4% | 397 | 478 | 1 |
+† CLM-v0.1-8B uses its choice head with a 6,144-token encoder; SecureBERT2.0 uses its detection head with a 4,096-token input cap. All models are scored on the same method sides and labels within each track.
 
 At 0.5, SecureBERT2.0 flags nearly every security method. Its high F1 therefore does not indicate useful alerting precision.
 
 ## Protocol and versions
 
-Each before and after method was scored independently without showing the repair. All reported runs cover every side of their split. `perch-broad-v1` uses one broad bug or security question, not the full multi-question Perch scan. CLM uses its native choice head with a 6,144-token encoder; SecureBERT2.0 uses a 4,096-token input cap. Different protocols are not a prompt-matched ranking. No cutoff was selected on these benchmark labels.
+Jev, JevK5, and jev-lite received one broad bug or security question per method. This is not the full multi-question Perch scan. No cutoff was selected on these benchmark labels.
 
 | Model | Frozen checkpoint |
 | --- | --- |