AI transparency¶
This page consolidates transparency for the attribution pipeline: what the system decides, how confidence is represented, what happens when the model is unavailable, and how a human can review or override an attribution result.
Release state
The transparency semantics below are documented and customer-facing. The model-card index is published at Model-card index, and a card is not represented as published until the evidence is available there.
Decision flow¶
- A metadata-only invocation record reaches the attribution processor through the gateway, proxy, ingestion API, or connector path.
- Deterministic checks attach tenant, source-system, policy, and timing context.
- RAIL evaluates candidate ownership and spend paths, using a trained RAIL model only for the ambiguous Stage B residual.
- The synchronous interceptor performs a fast index lookup and emits the event; it does not run trained inference.
- The asynchronous processor may score the attribution edge with the trained Stage B adapter inside its 20 ms call budget and 100 ms processor objective.
- If that adapter is unavailable, slow, stale, or demoted, the processor uses a deterministic heuristic-baseline result instead.
- The output carries an evidence trail, confidence value, result source, and output state.
The request path remains fail-open and never waits for trained attribution.
Confidence semantics¶
Operational confidence (coper) is deliberately capped at 0.95. Venturi
does not display certainty. Confidence is used to decide whether an edge is
chargeback-ready, review-needed, or analysis-only; it is not a claim that the
system has perfect visibility into every pathway.
See Confidence & evidence for output states and evidence-card fields.
How to act on a result¶
The canonical vocabulary is the six
output states plus the three confidence bands
(chargeback-ready coper >= 0.80, provisional 0.50 <= coper < 0.80,
estimated coper < 0.50). In
practice they reduce to these customer actions:
| You see | Which results | Customer action |
|---|---|---|
| Chargeback-ready | deterministically_resolved or strongly_inferred at confidence ≥ 0.80 |
Eligible for chargeback workflows. |
| Provisional (needs review) | strongly_inferred or ambiguous below the 0.80 floor |
Review the evidence card and accept, correct, or defer; never auto-billed. |
| Bounded | bounded |
Use for planning and coverage work, not chargeback. |
| Unknown / not identifiable | unknown, not_identifiable |
Connect the source that would close the gap, or leave out of chargeback; never zeroed or hidden. |
| Heuristic baseline marker | Any result answered by the deterministic fallback | Treat as visible but degraded; filter or exclude until re-observed. |
Human controls¶
Reviewers can accept, correct, or reject attribution evidence according to their role and tenant policy. Review actions are written to the audit trail. Corrected edges keep the original evidence and the reviewer action so downstream reports can distinguish model output, heuristic fallback, and human review.
Data use and opt-out boundary¶
Venturi processes metadata needed for attribution: identity, service, project, cost, token counts, timing, policy context, and source-system evidence. Prompt and completion bodies are not requested for connector setup and are not needed for these transparency controls.
If a pathway cannot be observed without content capture, Venturi treats that as a known limitation rather than silently expanding capture. See Known limitations.