Research, with its limits stated
Benchmarks, specifications and papers from Compex. Each entry states its status, version and limitations. Results are evidence, not customer traction.
Benchmark · · Draft for external review
Authorized to Use, Forbidden to Influence: Measuring Purpose Selectivity in Financial AI
In a benchmark of two model and task pairs, chosen because they responded to the signal, a prohibited synthetic signal changed decisions in 96–100% of cases when the model could see it. Technically excluding it brought that back to about the models’ natural variation of 0–4%.
Specification · · Draft for external review
Purpose-Bound AI Execution Profile (PBAE)
What an organization approves and what its AI actually does can drift apart unnoticed. In a controlled test of 26 attack families on a research runtime, PBAE cut silent compromise from 69.2% to 7.7%.