false-positives

AI-detector auditing toolkit: measures how often a detector flags genuine human writing, how stable its verdict is across seeds, and whether that verdict survives meaning-preserving edits. Detector-in-the-loop measurement harness. Claude Code skill + Python CLI. MIT.

Source code used for the results reported in the TOIS2021 journal paper.

Occupancy models that account for misclassifications.

Primer for blast radius analysis and differential testing of dead code detectors and static analysis tools across pinned real-world projects

Trust-first, precision-biased Discord automod for tight-knit communities: independent-signal corroboration, cross-user burst raid detection, shadow rollout, auditable trust. An explicit recall/precision trade-off — not a zero-FP guarantee.

Living freight-fraud investigation simulator: autonomous freight operations, behavioral signals, false positives and risk investigations — with simulation ground truth hidden from the analyst.

Most AI security tools try to find more. SecHelix tries to prove itself wrong: every candidate finding goes to an independent verifier whose only job is to disprove it. Open-source AppSec agent for Claude Code, Codex and Copilot.

A false-positive corpus for secret scanners: 109 formats of ordinary log and build output containing zero credentials, plus a recall half of 42 formats with 81 planted synthetic credentials and an answer key. Ships as a GitHub Action that gates the build.

Low-false-positive, context-aware secret scanner with a pre-commit guard

Supplementary code for: Trivial rational contamination in PSLQ-based PCF searches. Pre-screening protocol + AEAL governance log.

Interactive Bloom filter — add keys and watch k hash functions flip bits, then check membership: 'definitely not present' vs 'possibly present'. Trigger a real false positive and watch the estimated error rate. No false negatives, ever. Runs in your browser.

One real vulnerability, one scary-looking false positive — can your scanner tell which is which? 10 AppSec cases, 3 of them decoys. A false positive costs exactly what a miss costs.

A target that looks vulnerable and is not. Sixteen routes shaped like classic findings, none of which do the thing they appear to do, so every finding reported against it is a false positive by construction. The precision control for scanner benchmarking.

Versioned, Ed25519-signed catalog of known-public credentials (Stripe test keys, AWS canary keys, RFC and jwt.io examples) so leakferret marks documented placeholders as FIXTURE instead of false-alarming. Source URL required per entry. Data licensed CC-BY-SA-4.0.

A rejection filter for 9 um ink detection: five conjunctive criteria calibrated on known ink and two geometric witnesses. Code MIT, measurements CC BY-NC 4.0.

Interactive GNU Guile lab for Bloom filter saturation, false positives, verification load, scalable layers, and rotating generations.

Does ancestry calibration of polygenic scores control false positives in gene-by-environment tests? A simulation study on real genomes.

Cross-validate Prowler Azure findings with 5 independent methods to cut false positives; Excel report + per-finding PNG audit evidence with Azure Portal deep-links.

An evaluation rig for command-guarding functions. Publish the harness, not the rules.

Battle-tested OVAL patterns to eliminate false positives/negatives in host configuration & security baseline checks