A Negative Result on State-Chaining over MITRE ATLAS

Abstract

Method note reporting the pre-registered retrospective test of AI-RISKPATH: whether a mechanical chaining engine, driven only by precondition/effect labels attached to the 208 techniques of MITRE ATLAS v2026.09, can reconstruct the 73 attack chains ATLAS documents. The case studies were split by a public randomness beacon (drand round 6479280, committed 24 hours before it existed), the vocabulary was built on one half and frozen with published salted commitments, and the metrics were measured once on the half never seen. Chain recall on the held-out half is 33.3 percent (12/36, 95 percent CI 20.2-49.7) and lift is 1.253 (95 percent CI 1.153-1.364), below the pre-registered continuation threshold: the rule stops the project. The cause is structural. After calibration, 108 of the 208 techniques carry no precondition anywhere in the ATLAS text, 42 percent of transition weight, so almost any technique remains admissible at almost any point. The note reports the protocol, the negative result, a contamination statement including two identifier disclosures found by audit and declared, and the measurement that explains the failure. The three earlier versions of this record are the pre-registrations it relies on. The labelled vocabulary is not published; the deposited commitments allow any claim about it to be checked without disclosing it.