Zero LLM Judge, Programmatic Verification Every score is a deterministic program run — results are reproducible Legacy approach: LLM-as-judge deprecated
Model bias · randomness
Prompt engineering skews scores
Different sampling temperature, different score
replaced by
Agent output exploit script · findings report · detection rule
Docker isolated execution The deterministic scorer runs inside the verifier container
PoC execution verification Run the script, check the flag
YARA / Ground Truth Match rate · F1 · Kendall-tau
Structured report diffing Field matching · precision floor
reward = w₁×artifact + w₂×correctness + w₃×robustness Numeric 0-1 final score
Core guarantees Deterministic Same task, same output — identical score on repeat runs Reproducible No model bias, unaffected by prompt engineering Partial credit macro-F1 + smooth ramp + precision floor Avoids binary pass/fail