Deterministic code-intelligence.
Gamma builds a high-dimensional capability fingerprint per file and binds distributed capability signals across files into a single coordinated-pattern verdict — backdoor and exfiltration shapes that no single file reveals. Precision-favoured by design. No model inference, no GPU, no network at scan time.
Numbers from the 2026-05-08 blind test (24 risk fragments in 5 groups + 50 innocent decoys) and the 8-repository real-world cohort. Reproducible from the pinned research corpus.
Zero false alarms on 50 decoy files and 8 public repositories. Per-file recall is intentionally low (0.208): one flagged file is enough to triage the surrounding module. A recall-tuned profile trades precision for coverage — stated, not hidden.
Internal rule names, scoring weights, the capability-algebra mechanics and formula identifiers stay sealed. Known gaps (availability-class sabotage, distributed-ransomware binding) are disclosed under evaluation, not advertised.