Skip to content

Confidence & evidence

Venturi attaches a calibrated measure of certainty to every attribution and shows the evidence behind it. This is what makes the numbers safe to act on: you can tell at a glance whether a chargeback line is directly resolved, strongly inferred, or genuinely uncertain, and decide accordingly.

This page explains the confidence score, the output states, the evidence card, and how Venturi reports a result it cannot resolve.

Operational confidence: coper

Every attributed edge carries one customer-facing confidence number, coper (operational confidence), on a scale from 0 to 1.

coper is capped at 0.95

coper is capped at 0.95: a deliberate, conservative policy ceiling. Venturi never reports an attribution as fully certain, even when its internal evidence is a direct, exact match. This is a transparency decision, not a statement about model error: it keeps every score auditable and prevents overconfident chargebacks. There is one customer-facing confidence object, and 0.95 is its ceiling.

You will see coper on dashboards, in exports, and in API responses, both as a number and as a band.

The chargeback floor: 0.80

For an attribution to be chargeback-eligible (provider spend charged back to an internal team), it must clear the canonical confidence floor and the independent evidence and financial gates:

Chargeback floor: coper >= 0.80

Operational confidence of 0.80 or higher is necessary for chargeback. It does not establish financial accuracy or a statistical error guarantee. The record must also have export_eligibility of eligible, a permitted output state, current evidence, no open dispute and no degraded result. Reconciliation and finance acceptance govern publication of the exact export. The same confidence floor applies to underlying spend in verified-savings calculations; savings need their own baseline and acceptance evidence. These checks do not determine Venturi’s annual license fee.

Attributions below 0.80 remain fully visible and useful for analysis and investigation; they are not promoted into a chargeback line until they clear the floor. Stage C allocation-derived splits carry a confidence ceiling of 0.70 where proportional allocation signals are present and 0.50 where none are, so shared-cost estimates are never silently charged back as if they were direct ownership.

Confidence bands

To make the score easy to act on, coper is summarized into bands:

Band Range What it means
Chargeback-ready coper >= 0.80 Cleared the confidence floor; independent export and financial gates still apply.
Provisional 0.50 <= coper < 0.80 A usable estimate: good for analysis, not yet for chargeback.
Estimated coper < 0.50 A weak signal. Surface it, investigate it, do not bill on it.

Stage C fractionally allocated edges can never reach Chargeback-ready; their highest attainable band is Provisional, because they represent shared or indirect cost rather than direct ownership.

Output states

Confidence answers how sure; the output state answers how the answer was reached. Each edge has one of six output states. They never blur together:

Output state Meaning Comes from
deterministically_resolved Resolved from facts with certainty. Stage A: exact keys and rules.
strongly_inferred Inferred with high confidence by the model. Stage B: trained attribution model.
bounded Uncertain enough that Venturi reports a constrained range, not a single owner. Stage B.
ambiguous Multiple candidates remain plausible and no single answer is justified. Stage A conflict or Stage B.
unknown The evidence does not support an attribution at all. Any stage.
not_identifiable The signal cannot be tied to a known node in your environment. Any stage.

These states let attribution degrade gracefully (from precise, to provisional, to a transparent unknown) without ever collapsing distinct kinds of truth into one falsely precise number. The last two states are how Venturi tells you the truth instead of guessing.

Evidence cards

Every customer-visible result carries the metadata you need to interpret and trust it: an evidence card. Wherever you can see an attribution, you can open its card and see exactly why Venturi reached that conclusion:

Field What it tells you
Stage origin Which stage produced the result: stage_a, stage_b, or stage_c.
Output state One of the six states above.
Confidence The coper score and its band.
Evidence basis The signals that supported the attribution: e.g. direct account binding, HRIS membership, deployment metadata, telemetry correlation.
Model version When inference was involved, the trained model version and the hash of the model artifact that produced it, linked to its tracked record.
Result source For Stage B inference, whether a trained model or a heuristic baseline produced the answer. The queryable degradation_state marker preserves fallback status. Read it with stage origin and evidence basis to distinguish deterministic Stage A results.
Degradation state Whether the result was produced normally or by the heuristic-baseline fallback.
Demotion reason If a previously chargeback-ready edge has been capped below the floor, the reason it was demoted: calibration drift or model age.
Freshness The timestamp the underlying data was current as of.

Because every result carries its basis, a disputed chargeback is never a black box. You can trace any number back to the evidence, including conflicts, where Venturi preserves all the disagreeing sources rather than hiding them.

Unknown is reported, not guessed

The most important thing the engine does is refuse to guess when the evidence does not justify an answer.

Unknown is a first-class answer

When attribution genuinely cannot be resolved, Venturi marks the result unknown or not_identifiable and surfaces it explicitly. Unattributed and shared spend is visible and exportable: never coalesced into an aggregate, never silently assigned to the nearest team.

This is why the dashboard shows you an unattributed-spend figure rather than hiding it. When you investigate one of these, the evidence card tells you why it is unknown: stale org data, a shared API key, conflicting CI/CD metadata, weak telemetry, or an unmapped identity, so you know whether the fix is a data-quality issue, a coverage gap, or a genuine limit of what is knowable.

Confidence you can trust over time

A confidence score is only worth as much as it stays true. An edge that was chargeback-ready last quarter is not automatically chargeback-ready today: the model behind it can drift, and the data behind it can age. Venturi watches for both and moves an edge out of chargeback-grade the moment it can no longer stand behind the number; it never silently keeps billing on decayed confidence.

A continuous calibration and freshness check

For every serving model, Venturi continuously measures whether its confidence still matches reality, comparing the scores it emits against accumulating ground-truth outcomes, and watches how long that model has been in service. When either signal says the edge is no longer trustworthy at chargeback grade, the edge is demoted:

What Venturi detects What happens to the edge
Calibration has drifted: the model’s confidence no longer tracks real outcomes over a sustained window. The edge is capped to at most Provisional (coper held below the 0.80 floor) until the model requalifies.
The model has aged past its serving limit. The edge is capped below the floor and flagged for refresh, so a stale model stops feeding chargeback lines.

A stale edge cannot stay chargeback-ready

These checks only ever demote confidence; they never raise it. Restoring an edge to Chargeback-ready always requires a fresh, human-approved model, never an automatic promotion. When an edge is demoted, its evidence card records why (calibration drift or model age) so a previously chargeback-grade line that drops is explained, not mysterious.

The practical effect: attributions that no longer meet the chargeback gates leave the internal chargeback base. If the engine can no longer stand behind an attribution’s confidence at chargeback grade, that attribution leaves the chargeback base on its own (visibly, with a reason you can read) rather than quietly continuing to bill at a confidence it no longer earns.

Knowing a model stood behind the result

On the synchronous hot path, the decision-time interceptor performs a fast index lookup and is built to fail open. In the asynchronous processor, if the trained model cannot answer within its 20 ms Stage B timeout, a fast heuristic baseline answers instead so attribution processing continues. That is the right behavior for availability, but a fallback answer must never be mistaken for a model-backed one, even when the heuristic is confident.

Every inference result says whether a model produced it

Each Stage B result carries an explicit, queryable result source marker (trained model versus heuristic baseline) alongside a degraded flag, and this marker travels with the result all the way to the API, unchanged even when the heuristic’s coper is 0.80 or higher. A confident fallback is still labeled a fallback. In the API, the marker is the degradation_state field: degraded_serving or disabled means the heuristic baseline answered, model_recalled means the prior validated model served the result, and none means the result was produced normally.

This is the signal Venturi itself uses to keep degraded results out of the figures you act on. Any result whose degradation_state is not none has an export_eligibility other than eligible, whatever its coper, so it is excluded from every chargeback export and from the internal chargeback base, and fallback attributions are also excluded from the savings counterfactual and from adoption cohort rollups. Deterministic Stage A results do not require a model call; they remain eligible when they meet the confidence, reconciliation, and workflow gates. For Stage B results, filter on the same marker in the dashboard and API to distinguish trained inference from a degraded fallback.

How to act on confidence

You see… Do this
deterministically_resolved, Chargeback-ready Verify export eligibility, disputes, reconciliation and finance acceptance before billing.
strongly_inferred, Chargeback-ready Review the evidence card and the same export gates; an open dispute excludes the record.
bounded or Provisional Use for analysis and budgeting; investigate before charging back.
Estimated (coper < 0.50) Surface and investigate it; do not bill on it.
Allocated split Treat as a fair fractional estimate of shared cost, not direct ownership.
A demoted edge (Chargeback-ready → Provisional) Read the demotion reason on the evidence card (calibration drift or model age) and do not bill on it while demoted. It stays demoted until a refreshed model requalifies and a human approves that model for serving.
Result source = heuristic baseline / degraded Treat as a fallback estimate; exclude it from chargeback and savings figures until the same edge is re-observed under trained-model attribution.
unknown / not_identifiable Investigate the cause shown on the evidence card; improve the underlying data source.

Where to go next

Give feedback on this page

Sign in to send feedback. Opens in a new tab.

Copy page details 
Docs feedback: Confidence & evidence
Page: https://docs.venturi.systems/engine/confidence-and-evidence/