Skip to content

04 / The attribution engine

Deterministic first. Inference for the unresolved.

Venturi resolves exact and rule-supported relationships before calibrated inference is considered. The attribution engine preserves alternatives, conflicts, and unknown states, while financial allocation remains separate.

Question

How does the attribution engine resolve eligible relationships without hiding uncertainty?

Mechanism

How the mechanism works

The Venturi attribution engine runs three governed stages. Exact and constrained evidence is evaluated first. Only the eligible residual candidate set can enter calibrated inference. Shared cost allocation is calculated separately.

  1. Stage A

    Deterministic resolution

    Exact joins and constrained rules resolve supported relationships. A missing match remains eligible for later evaluation only when the schema and evidence permit it.

  2. Stage B

    Calibrated attribution inference

    Eligible unresolved relationships are ranked with calibrated uncertainty. Alternatives remain visible, and the engine can abstain.

  3. Stage C

    Fractional cost allocation

    Shared cost can be allocated across governed recipients. The allocation result cannot replace or rewrite ownership.

Illustrative synthetic example

Operational confidence

0.84

Output state

Strongly inferred

At or above the 0.80 review threshold

Explanation
One synthetic residual candidate leads after calibration; alternatives remain in the record.
Time basis
Event-time, synthetic evaluation window
Source health
Enabled synthetic sources healthy
Action eligibility
Decision support

Operational confidence is produced at materialization and capped at 0.95. A score alone never determines product use. Attribution coverage, source health, feasibility, time basis, output state, and intended decision travel with it.

Annotated visual

Inspect the path and its unresolved states

Three-stage attribution pipeline
  1. AExact and constrained evidenceresolved edges or eligible residuals
  2. BEligible residual inferencecandidate, alternatives, or abstention
  3. CSeparate allocationfractional cost result, ownership unchanged
Deterministically resolvedStrongly inferredBoundedAmbiguousUnknownNot identifiable
Highlighted: the output state of the example record below.

In summary: Stage A resolves exact and constrained relationships. Stage B receives only eligible unresolved relationships and can return a leading candidate, alternatives, or abstention. Stage C allocates shared cost separately. The resulting output state is one of six canonical states.

Concrete example

Apply the mechanism to one synthetic workload

Illustrative synthetic example

Decision record

Inspectable residual-inference record

Mechanism . Describes how the product is designed to work.
Workload
evaluation-runner-workload
Invocation
invocation.synthetic.0318
Service
quality-evals
Project or code owner
AI Platform
Identity
svc-evaluation-runner
Organization
Applied AI Systems
Budget responsibility
Model Quality Programs

Operational confidence

Operational confidence

0.68

Output state

Ambiguous

Below the 0.80 review threshold

Explanation
Two eligible owner paths remain close and the deployment source is degraded.
Time basis
Event-time, synthetic deployment window
Source health
Deployment metadata degraded
Action eligibility
Human review required

Evidence sources

  • Provider billing record
  • Service identity
  • Partial deployment metadata

Alternative candidates

  • Search Experience, overlapping deployment window
  • Research Systems, shared queue evidence

Truth domains

Observed . Directly present in a source record. Inferred . Derived from eligible evidence with uncertainty preserved. Event-time . The state that applied when the governed event occurred.
Output state
Ambiguous
Time basis
Event-time, synthetic deployment window
Source health
Deployment metadata degraded
Action class
Human review required
Correction history
No correction event in this synthetic record

Evidence and implementation boundary

What the current claim supports

Architecture is defined and production validation is pending. Calibrated inference applies only to eligible unresolved relationships. A score alone never determines product use.

The decision record retains the winning candidate, alternatives, conflicts, source health, time basis, output state, operational confidence, and action class. Ambiguity is not converted into a false resolution.

Calibration and model-governance detail

The uncapped posterior attribution mass produced by Stage B is abbreviated cpost. Eligible post-hoc calibration may use Platt scaling, isotonic calibration, or beta calibration by edge type and evidence regime. Model promotion requires versioned evaluation, calibration checks, and rollback criteria.

These calibration-model names describe technical implementation choices. They are not customer outcome claims.

Known limitations

Unknown and unresolved states remain valid outcomes

  • 01Probabilistic inference cannot repair an uncapturable pathway or invent a relation type outside the schema.
  • 02Source degradation can reduce the eligible evidence set and move a result to ambiguous or unknown.
  • 03Fractional cost allocation may split shared cost but cannot overwrite the ownership attribution result.

Decision enabled

Determine whether a relationship is resolved, bounded, ambiguous, unknown, or not identifiable, and which action class is allowed.

The output supports a bounded review decision: observe, use for decision support, send to finance review, require a human, or keep the customer-controlled Gate path ineligible.