Update!
I have been actively developing since submission (verification + benchmarking hardened)
I have continued improving 4n6 Nexus against a self-assessment of where I wanted to be. The recent work landed with preserving the F1=1.00 detection invariant (these are integrity/verification overlays, never changes to the scored detection path):
- Adversarial verification tier.
The verifier now re-derives each finding's asserted IP/PID anchors against its own tool-output record (catching a real but misattributed anchor a corpus-wide check would miss), and runs an iterative verify/remand loop that re-adjudicates contested findings across bounded rounds and escalates to human review at a remand cap. This closes the "single-engine verification" gap.
Hallucination/abstention benchmark
Added blind separate ground truth that the detectors were not authored against (scored through the real engine, with a detector-logic anti-overfit hash), a defenses OFF/ON ablation, byte-flip tamper tests proving the receipt/audit-chain integrity path catches corruption, and named FP/FN accounting separate from F1.
One point that I wanted to be honest about was the OFF/ON ablation reports the injection defense's neutralization efficacy, not a live hallucination-rate delta, the deterministic engine can't be manipulated by injected prose, so that rate is 0 in both arms (a live delta would require an LLM-in-the-loop analyst path). I would rather disclose that than paper over it.
Verification is deliberately a deterministic tier (CI-safe, reproducible), not a live rival-model debate.
Log in or sign up for Devpost to join the conversation.