TriageMate Java scan labels

Java static-analysis alerts from Vul4J entries and GitHub advisory fix commits, with a binary TP/FP label.

vul4j-labeled.jsonl is the earlier 66-row full-repository scan. vul4j-scan-labeled.jsonl is the current 14-row changed-file scan. ghsa-scan-labeled.jsonl is a 174-row set from GitHub advisory fix commits, used as a second real-world check.

Each row carries the entry ID, rule, CWE, file, line, code window, and the Semgrep severity score used as the no-model baseline.

Label rule

A finding is a true positive when its reported line overlaps a line changed by the fix, including two lines of context, or falls inside a method the fix changes. The labels are a location proxy, not proof that the finding is the exact CVE.

Current result

The changed-file set has 14 findings across 10 entries: 5 location-proxy true positives and 9 false positives. It covers every Vul4J entry with a usable patch commit: 126 of 127 checked out and scanned, with test sources excluded. The Semgrep severity baseline reaches ROC-AUC 0.5778 and precision at 90% recall 0.3571. LiquidAI d1-3B reaches ROC-AUC 0.8000 and precision at 90% recall 0.5556, a gain of 0.2222 AUC. Kev-4B reaches ROC-AUC 0.6667 and precision at 90% recall 0.4545, a gain of 0.0889 AUC.

The d1 gain survives leave-one-entry-out (never below 0.1250) and the row bootstrap (P(delta <= 0) is 0.094). Kev-4B flips negative when VUL4J-65, VUL4J-83-S or VUL4J-107-S is dropped, and its Vul4J row bootstrap puts P(delta <= 0) at 0.35.

With five positives the ranking claim is still small-sample. Labels are based on overlap with fix hunks or methods and are not human confirmation of vulnerability identity. The earlier 46-row set mixed full-repository and changed-file scans, so 36 of its rows were findings in files the fix never touched; its numbers are superseded.

Second set

ghsa-scan-labeled.jsonl is 174 findings from 68 fix commits in Maven packages, sampled from the OSV advisory export across 12 CWE families and 314 repositories. The scan and label code is the same as Vul4J: changed Java files only, test sources excluded, positive when the finding overlaps a fix hunk or a changed method.

Scorer	n	ROC-AUC	Precision at 90% recall
Semgrep severity heuristic	174	0.5886	0.4885
LiquidAI d1-3B	174	0.6941	0.4847
LiquidAI d1-3B + criteria	174	0.7028	0.5132
Kev-4B	174	0.7521	0.6063

d1 gains +0.1054 ROC-AUC but loses 0.0038 precision at 90% recall on the stored labels, so the gate fails there and d1 is not promoted. The heuristic has only two scores: 90% recall needs 77 positives, only 59 sit at 0.75, so the cut moves to 0.5 and accepts every row. Precision there is 85/174 = 0.4885, the heuristic's overall precision. Ties are accepted whole, so row order cannot change it. An entry-level bootstrap puts the d1 AUC gain at -0.0952 to +0.3232 with P(delta <= 0) of 0.1855.

Adding the cookbook's true/false criteria fixes that gate failure: AUC 0.7028 and P@90R 0.5132, both above the heuristic. The gain is rule-level calibration: weighted within-rule AUC is 0.474, still below 0.5.

The verdict is label-dependent. Of the 85 stored positives, 44 sit on a changed hunk line and 41 sit only inside a touched method. Under a hunk-only rule (44 TP), a CWE-consistent rule (53 TP), or both (28 TP), d1 passes the same gate, with AUC from 0.667 to 0.837. Only the strictest definition keeps d1's entry-bootstrap gain above zero.

Kev-4B passes this gate on the stored labels, but its Vul4J interval includes zero, so no scorer replaces the heuristic default.

On the stored labels a metadata lookup matches or beats both models. Out-of-fold logistic regression on rule name alone reaches 0.767 AUC, and rule plus file name reaches 0.786 entry-grouped and 0.799 repo-grouped, against d1's 0.694 and Kev's 0.752. Under the hunk-only label the same baseline falls to 0.461 entry-grouped and 0.587 repo-grouped, below d1's 0.667, so the loose label is what metadata explains. A swap-window control finds no code-reading advantage either: replacing each row's code window with another row's moves within-rule AUC from 0.474 to 0.590 after a global swap and 0.444 after a within-rule swap, and every AUC difference interval includes zero.