GhostDrift Math Research Institute Releases Evaluation OS for Verifiable AI-Era Trust
The release reframes AI-era evaluation from a single score to a check of whether conclusions hold when the measurement ruler changes, turning corporate claims into publicly verifiable trust.
Reporting from 1 source: ASCII.jp.
GhostDrift Mathematical Research Institute released Evaluation OS on GitHub, a system that mechanically verifies whether an evaluation conclusion survives changes to the measurement criteria. In a case study, 60 evaluation conditions applied to On the Links, a representative organization of the Hiroshima AI Assurance Council, all produced the same conclusion, with the minimum value 70.0 clearing the preset strict threshold of 68.0. The institute disclosed the rules, evidence, limitations, and code for third-party reproduction.
GhostDrift Math Research Institute is a strategic partner of On the Links, so the case study is not an independent third-party audit. The institute published the evaluation rules, evidence, limitations, and Python implementation code so third parties can reproduce the calculation, and it fixed 32 investigation conditions before the evaluation, covering employee reviews, product reviews, and complaint candidates.
The release is the second technical output of the Hiroshima AI Assurance Council's HAAP protocol, following a pharmaceutical cold chain proof of concept. The institute also formalized in Lean 4 a selection principle that prioritizes whether a claim can be verified over company size or brand recognition, and a provenance principle that treats unresolved origins as unconfirmed regardless of later adoption.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.