Sergei Parfenov outlines a framework for managing agentic AI reliability by replacing scalar trust scores with typed provenance vectors. The approach addresses silent failure modes where a single trust score—such as a 0.0 to 1.0 rating—fails to account for independent degradation across different performance axes. By propagating a multi-dimensional record alongside model outputs, developers can enable downstream consumers to apply independent policies, ensuring that decisions reflect specific data requirements rather than a collapsed, potentially misleading average.

In the proposed model, provenance acts as a typed record defining how a value was produced, including the model version, input freshness, and tool execution status. Consumers evaluate this provenance using a policy that defines required floors for each axis, such as freshness or capability. For instance, a summarization task might tolerate a weaker model but require high data freshness, whereas a price calculation step might prioritize verification and calculation accuracy over model strength. When a provenance vector fails to meet these per-consumer requirements, the system does not simply halt; it triggers specific recovery actions linked to the failed axis, such as re-fetching stale data or escalating to human oversight.

The structure utilizes a min-aggregation method where output provenance is determined by the weakest performance axis across all input streams. This prevents degradation from being washed out by successful operations in other parts of the chain. Practical implementation suggests starting with axes that map directly to unique recovery actions—such as freshness for re-fetching or capability for re-running on primary models—to avoid excessive complexity. This paradigm, categorized alongside data lineage and taint analysis, emphasizes that trust is not an inherent property of data but a verdict computed by the end consumer based on its specific operational context.

Sergei Parfenov outlines a framework for managing agentic AI reliability by replacing scalar trust scores with typed provenance vectors. The approach addresses silent failure modes where a single trust score—such as a 0.0 to 1.0 rating—fails to account for independent degradation across different performance axes. By propagating a multi-dimensional record alongside model outputs, developers can enable downstream consumers to apply independent policies, ensuring that decisions reflect specific data requirements rather than a collapsed, potentially misleading average.

In the proposed model, provenance acts as a typed record defining how a value was produced, including the model version, input freshness, and tool execution status. Consumers evaluate this provenance using a policy that defines required floors for each axis, such as freshness or capability. For instance, a summarization task might tolerate a weaker model but require high data freshness, whereas a price calculation step might prioritize verification and calculation accuracy over model strength. When a provenance vector fails to meet these per-consumer requirements, the system does not simply halt; it triggers specific recovery actions linked to the failed axis, such as re-fetching stale data or escalating to human oversight.

The structure utilizes a min-aggregation method where output provenance is determined by the weakest performance axis across all input streams. This prevents degradation from being washed out by successful operations in other parts of the chain. Practical implementation suggests starting with axes that map directly to unique recovery actions—such as freshness for re-fetching or capability for re-running on primary models—to avoid excessive complexity. This paradigm, categorized alongside data lineage and taint analysis, emphasizes that trust is not an inherent property of data but a verdict computed by the end consumer based on its specific operational context.