Research
Open problems in agent governance
The work below is unfinished. It is published because the gaps in a security product are more useful to a serious reader than another description of the parts that already work.
Open problems
Independent trust anchors
Evidence checkpoints anchored outside the administrative domain that produced the evidence, so completeness can be argued against a party the writer does not control. Currently Planned — nothing has been demonstrated.
A second implementation of the record format
The strongest remaining criticism of the evidence model is that the verifier shares an author with the system. A verifier written independently from the documented format would replace an assertion with a check.
Formal statements of authority
Expressing delegation and bounded authority in a form that can be reasoned about mechanically, rather than reviewed by reading a policy file.
Instruments that can be shown to bite
Positive controls and must-fail vectors make a conformance run meaningful. Generalising that discipline — proving a test suite would notice a deleted assertion — is unsolved here.
Cross-domain evidence
Actions that span organisations need records that remain checkable when neither party trusts the other's infrastructure.
Undeterminable outcomes at scale
Recording an outcome as unknown is correct and operationally awkward. Reconciliation strategies that preserve the honesty of the unknown state are an open area.
No papers, benchmarks or datasets are published yet. When they are, they will appear here with their methods and their failure cases, not only their results.