Benchmark
Never hides a real bug.
On a real Slither run across contracts with known vulnerabilities, Sift surfaced every annotated bug — 0 hidden — while cutting 91.9% of the noise: 334 raw findings down to 27 worth reading.
Known bugs hidden0 / 10
every known bug surfaced0 dismissed
The one error a security tool can’t make is hiding a real bug. When Sift isn’t certain, it surfaces the finding for review — it never silently dismisses a possible vulnerability.
Noise removed91.9%
27 surfaced334 raw
Most of the cut is trivial informational noise removed by a severity filter — no AI needed. The model does the hard part: judging the security-relevant remainder.
| Raw Slither findings | 334 |
| Informational / optimization — auto-filtered | −303 |
| Security-relevant — sent to the AI | 31 |
| ↪ false positives cut by the AI | −4 |
| Real findings surfaced | 27 |
Measured on raw Slither output, scored against known audit results. The triage model holds its accuracy on unseen 2023 audit contests it was never trained on — it generalizes, it doesn’t memorize. We’ll keep these numbers honest as the model improves.