ACE Research
A Retrospective Benchmark of Enterprise Trust Risks in Language-Model Endpoints
A bounded study of nine historical endpoint evaluations, with explicit scoring, denominator, reliability, and attribution limits.
The original ACE benchmark methodology, scoring framework, model results, and representative test cases.
This is the frozen record of the June 2026 benchmark round. Its figures are historical and are not recalculated when new evaluations are published.
Short technical analyses of verified failure patterns, enterprise impact, and remediation hypotheses. Each note records its evidence boundary and limitations.
No public notes released yetClassifying repeatable enterprise trust failures across data, regulation, agents, and evidence.
Testing whether audit findings can become targeted training and measurable regression controls.
Tracking behavior changes across endpoint, provider, policy, and harness updates.
Separating model behavior from routing, tool execution, provider controls, and evaluation infrastructure.
Private prompts, holdout cases, and non-public evidence remain outside public research artifacts.
Material changes to data, method, or interpretation receive a dated version and correction record.
ACE publishes evaluation limits and discloses when benchmark authors and maintainers share an affiliation.
Public evidence