The same investigative contract
Humans and agents can exchange an investigation
Investigators need to know which question was answered, about whom and when, using what evidence and criteria. That need is the same whether the assessment came from a person, an agent, or a mixed team.
AI security teams are the first adoption audience. Core's scope remains all cyber investigations: an agent can be the system under investigation, an investigator, or both in different contexts.
Keep distinct claims distinct
- Observation: a tool-call log records an outbound request.
- Assessment: the available evidence is insufficient to establish what content left the system.
- Decision: an investigator selects further collection and records the applicable policy context.
- Action: a collection operation was requested, attempted, or completed; those are different states.
An agent's explanation is not automatically evidence of an external event. Core requires an inspectable rationale and method, not disclosure of private model reasoning. Confidence values need their scale and interpretation.
Disagreement survives the handoff
Preserve competing assessments and the evidence each used. An explicit decision selects an assessment for a transition. Actor type, timestamp, or a larger confidence number does not silently select the winner. An unresolved conflict can remain unknown.
When new evidence changes an assessment, add a revision and review affected decisions. Keep the historical inputs intact. This supports examination of a decision without promising that rerunning a model will reproduce its output.
Authority stays explicit
Attribution identifies who or what made an entry. It does not establish permission, delegation, or credibility. Represent policy references and authority assertions separately, then enforce access and execution in the surrounding system.
Importing an investigation must not execute its recorded actions, fetch arbitrary locators, or grant access. Vendor and business layers remain responsible for their actual controls.
Try both sides of the model
The Investigation Lab includes an AI-system investigation and a conventional identity investigation involving an agent. Both are synthetic and use the same candidate binding.
Core contract and examples · ATLAS, CASE/UCO, and other standards · What validation can establish