Reproducible evidence
Claims should survive contact with the test harness.
Mythiril 1 publishes the controlled corpus, release gates, and failure conditions behind its current results. These are engineering snapshots, not universal claims.
100%Sensitive-region recall104 controlled instances
99.9825%Target glyph coverageLocal OCR detector
0Leaked target pixels12px hardened margin
140 / 140Multi-DPI checksAll release gates passed
Release discipline
What the current evidence proves
The same input and recipe replay to byte-identical deterministic PNG output.
The controlled redaction corpus passes the hardened 12px boundary policy with zero planted leaks.
Local OCR supplies reviewable candidates when cloud credentials are missing or rejected.
SAM execution remains provider-dependent and is never presented as ready when credentials are absent.
Precision hardening V2Adaptive pixelation, hostile segmentation, cross-runtime conformance, and bounded before-and-after results.Benchmark market reportMeasured pipeline, quality, determinism, and market-position evidence with bounded claims.Local detection hardeningCredential-independent OCR detection results across the controlled seven-page corpus.Redaction precision hardeningBoundary stress tests, planted corruption checks, and the proven 12px safety margin.