How should an AI system prove that a candidate change is behaviorally acceptable?
baseline/candidate experiments · deterministic invariants · trace comparison · judge reliability · uncertainty
ProofDiffBehavioral change assurance↗RESEARCH DIRECTION
Research questions that can fail a test.
The research direction sits where probabilistic model capability meets software constraints: evaluation, economics, authorization, governance, and operational reliability.
How should an agent release surface newly gained effective authority before deployment?
semantic permission plans · boundary scenarios · approval-path diff · severity-aware release gates
PermitDiffPre-release permission change assurance↗When is an inference-routing optimization economically real rather than anecdotal?
workload characterization · constraint envelope · observed provider profiles · baseline evidence · Pareto analysis
InferenceLedgerInference decision economics↗How should authorization remain cryptographically and transactionally bound to what an agent executes?
canonical intent · deterministic policy · bounded credentials · execution revalidation · replay control · audit integrity
AgentGuardRuntime action authorization + execution integrity↗How should an enterprise prevent generated text from becoming an unsupported external commitment?
evidence registry · provenance · freshness · conflict detection · review policy · source-change impact
TrustFlowExternal enterprise claim integrity↗How can AI participate in business operations without becoming the authority for destructive actions?
resolution reasoning · deterministic policy · typed actions · risk gates · idempotency · reconciliation
ResolveOpsGoverned business-resolution workflows↗METHOD
Claims must be able to lose.
Narrow the claim
Define exactly what the system should prove — and what it should not claim.
Make it falsifiable
Build a gate capable of rejecting the preferred result.
Preserve the boundary
Keep model judgment separate from security, policy, money, and release authority.
Publish the limitation
Scope metrics, uncertainty, failure modes, and unsupported generalizations beside the result.