ModelRig / security / built-in detectors
Built-in detectors — the measured false-positive rate
Everyone ships a guardrail; almost nobody publishes its measured false-positive rate. These are ModelRig's own built-in detectors, probed on public, license-clean synthetic corpora, with the false-positive rate published as a verifiable receipt — not a benchmark someone else ran, and not your traffic.
Public corpora only
Measured on public corpora only. Every figure on this page comes from a probe of ModelRig's own detectors against a named, license-clean synthetic corpus. No customer's traffic is ever measured here — the false-positive rate on your own captures is a private receipt you author and publish only if you choose.
Detectors, corpora, and receipts
| Detector | Kind | Public corpus | License | Measured FP |
|---|---|---|---|---|
| patterns@v1 | deterministic | injection-synth@v1 | CC0-1.0 | pending — published as a probe receipt |
| patterns@v1 | deterministic | benign-synth@v1 | CC0-1.0 | pending — published as a probe receipt |
| safe-harbor@v1 | deterministic | injection-synth@v1 | CC0-1.0 | pending — published as a probe receipt |
What a measured rate is, and is not
A measured false-positive rate on a stated corpus and detector version — a signal with a known error rate, not a guarantee, not prevention, and not a claim any detector catches everything. Rerun any figure yourself with modelrig screen probe, or verify its receipt.