02 / From evidence to attribution
From source evidence to an attributed decision.
Read-only evidence is processed asynchronously, resolved through deterministic and probabilistic stages, and materialized for bounded lookup. The customer request path never depends on live graph traversal.
Question
How does read-only source evidence become a decision-time attribution record?
Mechanism
How the mechanism works
Evidence collection, normalization, relationship resolution, inference, allocation, and materialization run outside the synchronous customer request path. Decision time reads only the bounded materialized index and available policy context.
Absolute invariant: Venturi failure must never block customer AI traffic. Shape is the default advisory path. Gate is an explicit, customer-controlled, per-workload opt-in path.
Annotated visual
Inspect the path and its unresolved states
- 01 Observe source evidence Evidence and attribution plane
- Input
- Read-only source records
- Output
- Governed source observations
- System owner
- Customer data plane
- Timing
- Asynchronous
- Failure behavior
- Source is marked degraded; request path is unaffected
- Truth domain
- Observed . Directly present in a source record.
- Claim state
- Architecture commitment . A required system behavior that still needs deployment evidence.
- 02 Normalize governed records Evidence and attribution plane
- Input
- Source observations
- Output
- Versioned governed records
- System owner
- Customer data plane
- Timing
- Asynchronous
- Failure behavior
- Invalid records are rejected or quarantined
- Truth domain
- Observed . Directly present in a source record.
- Claim state
- Mechanism . Describes how the product is designed to work.
- 03 Resolve deterministic relationships Evidence and attribution plane
- Input
- Eligible exact joins and constrained rules
- Output
- Resolved edges or eligible residuals
- System owner
- Venturi attribution engine
- Timing
- Asynchronous
- Failure behavior
- Unresolved relationships continue without forced completion
- Truth domain
- Inferred . Derived from eligible evidence with uncertainty preserved.
- Claim state
- Mechanism . Describes how the product is designed to work.
- 04 Infer eligible unresolved relationships Evidence and attribution plane
- Input
- Eligible residual candidate set
- Output
- Calibrated candidates, alternatives, or abstention
- System owner
- Venturi attribution engine
- Timing
- Asynchronous
- Failure behavior
- Ambiguous or unknown remains visible
- Truth domain
- Inferred . Derived from eligible evidence with uncertainty preserved.
- Claim state
- Mechanism . Describes how the product is designed to work.
- 05 Allocate shared cost separately Evidence and attribution plane
- Input
- Shared-cost evidence
- Output
- Fractional allocation result
- System owner
- Allocation stage
- Timing
- Asynchronous
- Failure behavior
- Ownership attribution is not overwritten
- Truth domain
- Allocated . A governed split of shared cost, separate from ownership.
- Claim state
- Mechanism . Describes how the product is designed to work.
- 06 Materialize decision record and index Evidence and attribution plane
- Input
- Graph state, evidence, policy and model version
- Output
- Inspectable record and bounded lookup
- System owner
- Customer data plane
- Timing
- Asynchronous
- Failure behavior
- Prior valid index can remain available with staleness visible
- Truth domain
- Current-state . The latest known state at the time of inspection.
- Claim state
- Architecture commitment . A required system behavior that still needs deployment evidence.
- 07 Read the index at decision time Customer request path
- Input
- Request context and materialized index
- Output
- Advisory or customer-controlled decision context
- System owner
- Decision-time interceptor
- Timing
- Synchronous
- Failure behavior
- Request proceeds unmodified
- Truth domain
- Event-time . The state that applied when the governed event occurred.
- Claim state
- Architecture commitment . A required system behavior that still needs deployment evidence.
In summary: steps one through six execute asynchronously in the evidence and attribution plane. Step seven executes in the customer request path by reading a materialized index. No step in evidence collection or graph resolution can block customer traffic.
Concrete example
Apply the mechanism to one synthetic workload
Support-routing workload
- Observe: provider billing, deployment metadata, service identity, and organization hierarchy.
- Resolve: exact invocation-to-service and service-to-project relationships.
- Infer: one eligible project-to-organization relationship remains and produces a strongly inferred result.
- Materialize: record the alternatives, 0.91 operational confidence, event-time basis, and decision-support action class.
- Serve: read the index and attach advisory context. The request proceeds whether or not context is available.
Evidence and implementation boundary
What the current claim supports
The architecture is defined. Production validation is pending. The forbidden synchronous dependencies are live graph traversal, connector calls, billing queries, evidence processing, and Venturi control-plane reachability.
Technical labels used in the implementation
AI Inference Detection and Invocation Discovery is abbreviated IDD. Deterministic resolution is Stage A. Calibrated attribution inference is Stage B. Fractional cost allocation is Stage C. Operational confidence at materialization is abbreviated coper. Plain-language labels remain primary.
Known limitations
Unknown and unresolved states remain valid outcomes
- 01A source can be healthy while a specific pathway remains only partially capturable.
- 02Eligible residual inference can abstain when candidates conflict or evidence is insufficient.
- 03A materialized index can become stale, so time basis and source health must travel with the lookup.
Decision enabled
Identify the stage that produced the record, the boundary that applies, and the source gap blocking operational use.
The sequence supports evaluation planning by making source activation, failure behavior, and intended decision explicit before any production workflow depends on the output.