Execution provenance
Reconstruct who did what, when, with which evidence, memory, tools, and permissions.
Independent AI systems research initiative
Assurance infrastructure for long-horizon multi-agent AI systems.
OME Veritas develops framework-neutral methods for execution provenance, causal failure attribution, and verified recovery.
Research prototype · Kazakhstan
Research focus
From opaque traces to verifiable explanations and recovery.
The research connects evidence, claims, tool calls, memory operations, inter-agent messages, permissions, and external state changes in one typed trace.
Causal attribution
Separate the first causal fault from downstream symptoms and propagated errors.
Controlled fault injection
Test corrupted evidence, tool errors, stale memory, communication failures, and permission violations.
Verified recovery
Evaluate retry, rollback, compensation, quarantine, and human escalation with post-recovery checks.
Method
A controlled and measurable evaluation process.
- Define deterministic scenariosSpecify fault location, expected safe behavior, and ground truth.
- Capture typed tracesRecord evidence, claims, tools, memory, messages, permissions, and state changes.
- Compare baselinesEvaluate logs-only methods against provenance-aware approaches.
- Measure outcomesTrack root-cause accuracy, containment, recovery success, false positives, latency, and compute cost.
- Audit labelsUse deterministic checks and manual review to reduce self-evaluation bias.
Current phase
Research prototype and benchmark design.
The work builds on an existing read-only multi-agent prototype and early components for permission analysis, circuit breaking, recovery, and trace-to-evaluation conversion. Results will be published only after reproducible evaluation.
Founder
Naiman Yelubayev
Founder & Principal Researcher
OME Veritas is operated by Individual Entrepreneur Yelubayev Naiman Muratovich, Kazakhstan.