Glossary¶

Canonical terms used across this site, the product, and Venturi’s governed documentation. When a page and this glossary disagree, treat the glossary as the intended vocabulary and tell us.

Attribution¶

InvocationEvent: The metadata-only ingestion record for 1 AI invocation. It contains identifiers, model, token, cost, timing, and correlation fields, never prompt or completion bodies. See Ingestion API.
AttributionRecord: The durable result produced by RAIL for an invocation. It carries the resolved graph edges, output state, stage origin, operational confidence (coper), and evidence metadata used by dashboards, exports, and the API.
Attribution graph: The connected representation of AI usage and organizational responsibility. Venturi resolves usage across six layers: model invocation, service, code ownership, identity, organizational hierarchy, and budget.
RAIL (Reconciliation Attribution Intelligence Layer): Venturi’s platform runtime and customer-facing attribution term. It spans Stage A deterministic entity resolution, Stage B inference and training, Stage C fractional allocation, evidence, uncertainty, anomaly interfaces, mapping, and feedback-loop learning. RAIL’s research and training machinery supplies verified Stage B models; it is part of RAIL, not a separate product. See RAIL pipeline.
Attribution coverage: The share of capturable AI usage that resolves through the required attribution chain at or above the confidence floor. Venturi reports coverage against an honest denominator; it never claims complete visibility.
Pathway: A distinct route by which AI usage reaches your enterprise (managed provider APIs, cloud-hosted model services, self-hosted models, AI embedded in SaaS, automation platforms). Each pathway has its own capture mechanism.
Unknown / Shared / Unresolved: Explicit attribution outcomes. Unknown means the evidence does not identify an owner; Shared means multiple owners legitimately split responsibility; Unresolved means conflicting evidence is awaiting reconciliation. None of these are hidden, zeroed, or silently guessed.
Multi-truth model: Usage, cost, and ownership are recorded as distinct truths. A single event can be certain in one dimension and provisional in another.
Ownership vs allocation: Two distinct relationships in the attribution graph: ownership is a direct single-owner edge, allocation is a fractional split across several owners. Keeping them separate stops shared responsibility from collapsing into a false single-owner answer.

Confidence¶

Confidence score (coper): The operational confidence attached to every attribution result, capped at 0.95 as a deliberate policy ceiling. It is never called accuracy or certainty. The descriptive API field attribution_confidence maps to this same value; it is not a second confidence object. See Confidence and evidence.
Confidence bands: chargeback_ready (confidence at or above 0.80), provisional (0.50 to 0.79, labeled and never auto-billed), and estimated (below 0.50, excluded from chargeback).
Evidence card: The per-result explanation of which signals produced an attribution and at what confidence, available wherever a result is shown.
Posterior confidence (cpost): The uncapped model-internal confidence preserved for training, audit, and analysis. It is never shown on operational surfaces, which display only the capped coper.

Enforcement posture¶

Decision-time interceptor: The fail-open component on the AI request path that observes usage at decision time. It adds bounded latency and can never block customer AI traffic by failing: on any fault or timeout it passes traffic through and records null attribution.
Shape mode: The default advisory posture. Venturi recommends; your systems and people decide.
Active (Gate): The explicit, scoped, opt-in enforcement posture. Gate issues a signed deny recommendation to a customer-controlled enforcement point; the decision-time interceptor never denies traffic inline.
Fail-open / fail-closed: The interceptor data path fails open (customer AI traffic is never blocked by Venturi unavailability). Everything else fails closed: authentication, authorization, tenant isolation, admin actions, exports, and billing.
Policy evaluation record: The audit event written when a governance policy evaluates an invocation or attributed record, making advisory and enforcement behavior reviewable after the fact.
Manual override record: The audit event recording a human override of an attribution or policy outcome, preserving the intervention as evidence rather than mutable state.

Deployment¶

Control plane: The vendor-managed surface used for licensing, signed updates, configuration metadata, and governed orchestration. Deliberately separate from the data plane so raw telemetry never has to leave your infrastructure.
Data plane: The dedicated single-tenant runtime containing ingestion, processing, graph state, materialization, and serving. It runs in the customer’s cloud in self-hosted mode or in the contracted residency lane in SaaS mode.
Customer trust boundary: The boundary inside which customer telemetry, event processing, graph state, and materialized indexes remain unless explicitly exported.
Read-only integrations: Customer system connectors default to read behavior (get, list, describe) and never mutate your systems; a platform gate blocks any connector module that requests write access from shipping.
Workload Identity Federation (WIF): Keyless workload authentication in which a trusted external identity exchanges a signed assertion for short-lived cloud credentials. Venturi uses it by default for Google Cloud and Azure onboarding.
External ID: The customer-specific value Venturi must present when assuming an AWS cross-account role. It mitigates the confused-deputy problem; a role ARN alone is insufficient.

Cost and value¶

Chargeback-eligible: A cost line whose attribution confidence is at or above the 0.80 floor. Lines below the floor are labeled provisional or estimated and are never auto-billed.
Eco-efficiency: Capability delivered per unit of environmental cost. It is distinct from the Model Energy Rating and is null when any governed capability or energy input is unknown.
Identified savings vs realized savings: An identified saving is an opportunity the platform has found; a realized saving is one you have captured, verified against a frozen baseline. The two are never conflated.
TRAC: Total Risk-Adjusted Cost: billed cost plus a confidence risk premium plus carbon liability: spend adjusted for the uncertainty and externality context of the usage behind it.
Cost derivation method: The per-record field describing how a cost value was derived: direct billing evidence, calibrated estimation, or another explicit method, so measured spend is never conflated with inferred spend.
FOCUS export: Cost export pinned to the FinOps Open Cost and Usage Specification v1.2, with Venturi extension columns. See Reporting and exports.

Value metrics¶

Gross AI value: The total measured value attributed to AI before any cost is subtracted: measured financial value, plus measured productivity value, plus included inferred-productivity value. The top of the cost/value waterfall.
Net AI value: Gross AI value minus AI spend and the explicit quality and risk adjustments: the honest bottom line of the cost/value waterfall.
ROI: Net AI value divided by AI spend, computed at query time and undefined when spend is zero; never served as a precomputed ratio.
Cost per successful outcome: AI spend divided by the count of accepted outcomes, computed at query time and undefined when there are no accepted outcomes, so cost is judged against value actually delivered.
Quality-adjusted savings: Optimization savings scaled by the realized quality adjustment factor, so a savings figure is never claimed at full value when output quality regressed.

Data handling¶

Metadata-only capture: Venturi’s default: invocation metadata is captured, prompt and completion content is not. There is no content field in the event schema.
k-anonymity floor (k = 5): Aggregated views never expose groups smaller than five people. Individual- level views do not exist; every UI, API, export, filter, drill-down, and share rolls a smaller cohort into its parent.
WORM (write once, read many): Audit-storage retention that prevents records from being changed in place. Governance mode is the onboarding default; Compliance mode adds root-resistant retention for the configured 5-year window.
Crypto-shred: Erasure by destroying the per-subject or per-tenant encryption keys, backed by a deletion certificate.
Break-glass access: The only support access path: time-boxed, MFA-gated, audit-logged, and visible to you. There is no standing support access.

Product surfaces¶

Projection: A decision surface derived from the attribution graph (for example the chargeback view or the coverage map), not a standalone module.
Adoption intelligence: Attributed usage signals about organizational uptake of AI, reported cohort-only. It is not HR analytics and is never used for individual performance scoring.
Decision surface state: The maturity label on a product surface: observe-only (informational), decision-support (human-guided decisions with caveats), or operationally trusted (evidence quality supports tighter operational reliance).
Detected-only: A pathway or source observed through discovery evidence without enough request-level evidence for attribution. It remains visible in the discovery ledger and is excluded from chargeback.
Fan-out warning: A dashboard warning that one upstream action appears to be triggering a disproportionately large volume of downstream AI calls, so multiplied cost is investigated rather than misread as normal demand.
Model Energy Rating: The star rating shown on model energy surfaces, derived from measured watt-hours. See Energy methodology.
System of record: A destination Venturi is building toward, not a status it claims today.