Verification results

Read the verdict together with its checked scope, individual findings and source evidence.

Start with the verdict

Verdict	Meaning	Typical next step
`true`	All applicable assertions hold, their model-answer bounds fit the combined gate budget, and the indicator guard is not reached.	Continue according to your own workflow policy, retaining the evidence and scope.
`false`	At least one applicable failure has an individual error bound within the gate budget.	Investigate the specific failed assertion before changing the business record or decision.
`uncertain`	Required checks or decision-gate conditions remain unresolved.	Supply missing evidence or route the open question to a reviewer.

A qualifying failure can decide an item while other assertions remain unresolved. A deterministic failure can also prevent further model questions from being asked. In the false branch, undecided is empty; inspect individual assertions to see what remains unanswered. A true verdict after caller-specified skips covers a narrower scope. Neither verdict concludes anything about evidence outside the request, and a qualified model answer can still be wrong.

Find the reason and the evidence

Field	What to inspect
`assertions`	The check ID, versioned `test`, `outcome`, facts, explanation and request evidence pointers.
`undecided`	The assertion or gate condition still preventing a decided result, with its reason and message.
`scope`	The supplied-evidence boundary, skipped checks, certainty target and rules-only baseline verdict.
`configuration`	Identifies engine, pack, backend, target, qualifications and certainty-table data. It does not encode caller parameter overrides or skipped assertions.
`input_sha256`	The input hash; retain the actual input under your own data policy if you need to reconstruct the review.
`usage`	Questions asked, questions served from cache and recorded model time.

Evidence paths are JSON pointers into the submitted request. For example, /item/data/amount identifies the amount on a single item. Batch pointers can begin /items/0/data/. A missing-field finding can point to a field that is absent. Preserve the exact request so the paths can be interpreted against what was submitted; a path does not authenticate the source document.

Keep the request options as well as both identifiers. Two runs with the same configuration can use different policy parameters or skipped checks; the configuration identifier alone does not establish identical scope.

A recorded procurement-control exception

{
  "id": "po_present",
  "test": "procedure:po_present@4",
  "outcome": "fails",
  "error_bound": "0",
  "message": "No PO for an invoice of 24800.00, above the threshold of 1000.00.",
  "facts": {"required": true},
  "evidence": ["/item/data/amount", "/item/data/po_number"]
}

This excerpt comes from the synthetic supplier-invoice example, run with the local engine in the production configuration. The invoice is $24,800.00 and has no supplied PO reference, exceeding the default $1,000.00 threshold for requiring one. The evidence pointer to /item/data/po_number identifies the missing field; it is not a reference to an attached purchase order.

Understand assertion outcomes

Outcome	Meaning for your workflow
`holds` / `fails`	A procedure or accepted model answer reached a check result. The item-level gate still applies; inspect facts and any error bound before acting.
`insufficient_evidence`	Required data is missing, incomplete or inadequate for the check. Obtain the named evidence.
`not_applicable`	The assertion does not apply to this item and is outside the deciding set.
`skipped`	The caller explicitly excluded the check. This is not a pass.
`unqualified`	A model answer was recorded but cannot decide the check because the question is not qualified for that model.
`abstained`	The answer fell between the qualified decision thresholds.
`not_asked`	Questions were disabled or a direct failure had already decided the item. Read the message.
`context_too_long` / `model_unavailable`	The judgment could not be obtained; the limitation remains explicit.

combined_bound and guard are gate reasons in undecided, rather than assertion outcomes. The former means the applicable model bounds exceed the budget; the latter means indicators reached the guard. nothing_applicable means no assertion remained in the deciding scope. None of these conditions is a pass.

Keep certainty, probability and risk separate

The default certainty_target of "0.995" sets a gate budget of 0.005. certainty separately reports published evaluation evidence for a verdict group, including an upper error bound, measured_on and errors; its displayed value is one minus that bound. The lookup does not veto the verdict when its value is below the requested target. An uncertain result has null certainty; a decided result without a matching entry has basis: "unmeasured".

The table lookup uses the versioned pack, model artifact and dtype, with a no-model table fallback. It does not remeasure your population or policy changes. Neither its recorded holdout evidence nor the requested target establishes accuracy on your production records. 99.5% of which decisions? examines this distinction with recorded gate cases and constructed populations.

probability_true is accompanied by probability_basis. Model probabilities are multiplied, using calibrated values when available and raw values otherwise; this is not a measured item-level accuracy guarantee. risk uses a configured prior plus assertion and indicator weights to help prioritize review. Neither a strong-looking probability nor a low risk score overrides an unresolved required check, and a risk indicator does not by itself prove a failed requirement.

Why the supplier invoice returns false

Synthetic example. Response computed by the local engine in the production configuration. This page displays a saved example, not a new verification request. Sensitive identifiers are masked in public examples.

The invoice arithmetic, active vendor and duplicate comparison within the supplied scope hold. The missing required PO reference fails po_present, which decides the result. goods_specific is not asked because a direct failure already decides the item. The response does not claim measured certainty for this local configuration.

{
  "item": {"type": "invoice", "id": "INV-2026-0841"},
  "verdict": "false",
  "certainty": {
    "value": "0.9987",
    "basis": "holdout",
    "stratum": "false_deterministic",
    "bound": "0.00133",
    "measured_on": 2253,
    "errors": 0
  },
  "undecided": [],
  "scope": {
    "stratum": "false_deterministic",
    "evidence": "supplied",
    "skipped": [],
    "certainty_target": "0.995",
    "deterministic_only_verdict": "false"
  }
}

The next step is to obtain the PO reference or investigate the control exception with procurement before approval. The $24,800.00 is the transaction amount reviewed, not a quantified loss or misstatement. The auditor determines the finding’s significance and any materiality implications within the engagement.

The empty comparison list limits the duplicate finding to this supplied scope. A future request should be interpreted from its own context, returned configuration and outcomes.

Continue reading

Review marks — record human assessment of a completed audit’s verdicts.
Evaluation evidence — understand measured coverage and error bounds.
Response schemas — inspect the complete object definitions.