> ## Documentation Index > Fetch the complete documentation index at: https://docs.stateset.com/llms.txt > Use this file to discover all available pages before exploring further. # Jev + NSR > Turn unstructured messages into typed judgments, admit evidence, and evaluate explicit policy before an agent acts. **Jev reads; NSR applies policy and produces the decision evidence.** Jev supplies typed, probabilistic readings of text. NSR combines those readings with facts, evidence requirements, and explicit rules to return `approved`, `denied`, or `refused`. A model reading does not by itself authorize a business action. This guide reflects the latest local implementation in `stateset-nsr` and `nsr-app`, including work not yet committed in those repositories when reviewed. Availability depends on the deployed console and engine. Check capabilities before relying on a new surface; example fixtures are not evidence of a successful live Jev call. ## Choose your starting point Inspect a fixed industry policy, use your Jev key for one explicit run, and read the outcome. Save a Jev connection, work in the Playground, and preserve policy tests and results. ## How a decision is built
READTyped judgments

Jev answers narrow questions about the supplied message or context.

EVALUATEEvidence and policy

NSR checks admitted evidence, rule premises, thresholds, and the authorization goal.

INSPECTDecision and proof

Your application reads the outcome, review requirements, and returned evidence before acting.

For a refund example, Jev can assess whether a complaint describes plausible transit damage. The policy still checks authoritative order values, the refund cap, and other required premises. A plausible complaint can therefore be refused or denied under the policy. Keep numeric limits and authorization conditions in explicit rules rather than asking the model to authorize a refund. ## Typed readings become policy premises | Judgment type | Fact value | Confidence interpretation | | ------------- | ------------------------------- | ----------------------------------------------------------- | | `noul` | Probability `p` between 0 and 1 | Decisiveness: `max(p, 1 - p)`, not the probability of “yes” | | `choice` | A selected option | Probability of the selected option | | `score` | A position on an ordered scale | Provider-supplied confidence | A reading becomes a predicate such as `complaint_plausible(order_id, p)`. A rule can then require `geq(p, 0.8)`. The threshold is a policy condition, not an unconditional claim about model accuracy. A low fraud probability can be a confident “no”; it must not be interpreted as a low-confidence approval merely because its numeric value is small. Inspect evidence admission as well as the reading. Source labels and model provenance do not prove that every claim was accepted or that a real-world source is truthful. A derivation establishes a result relative to the supplied policy and admitted premises. ## Explicit questions and recursive questions The decision request can include: | Field | Role | | --------------------- | --------------------------------------------------------------------------------------------- | | `authorization_goal` | Predicate and arguments the engine must establish for authorization | | `judgment_state` | Context the judgment provider reads | | `judgments` | Explicit typed questions to ask | | `judgeable` | Allowed predicate templates the engine may ask when the proof needs them | | `max_judgment_rounds` | Bound for recursive judgment rounds; current implementation defaults to 3 and permits up to 5 | With recursive questions, NSR evaluates known premises, identifies missing declared judgments, asks those questions, and evaluates again. It also checks unanswered guards in deny/review rules and negated conditions. An apparent approval must not bypass an unanswered blocking guard; if the bound is reached with unresolved guards, inspect the refusal and human-review requirement. Judgments can include paraphrases and, for `noul`, a complement question. The implementation aggregates readings and caps confidence by their agreement. This is an additional consistency check, not a guarantee against adversarial inputs. Organization judgment templates can live with the policy. Request-level templates can override the corresponding organization entry; review this behavior when designing trusted callers. ## Read the result before taking action Inspect the returned fields that are available for your request: * **`decision`:** the engine verdict, not a model suggestion. * **`judgments`:** resolved model, readings, generated facts, agreement, rounds, and usage evidence. * **`evidence`:** admitted or rejected evidence and any human-review requirement. * **`proof`:** cited rules, derivation, and policy identifiers or pins when returned. * **`routing`:** the suggested operating band, including escalation. * **`refusal`:** missing premises or reasons the engine could not establish permission. Treat `routing.band: "escalate"`, `evidence.requires_human_review`, or `refusal.requires_human_review` as review signals. A successful HTTP response can contain a refusal. Do not rewrite a message, weaken a threshold, or invent a fact to force approval. The public Jev demo executes no business action. Production execution is a separate integration: use the [agent gate](/stateset-nsr-agent-gate) and bind the authorized operation to its execution and outcome evidence. Approval alone is not proof that a refund or another provider action occurred. ## Compare, verify, and learn carefully A fresh Jev call can produce a different reading. Distinguish: | Activity | What it tells you | | ------------------------------ | ---------------------------------------------------- | | Inspect an original decision | What was actually recorded for that request | | Run saved inputs again | A fresh result under the current runtime and model | | Verify a returned proof bundle | Whether its supported proof checks succeed | | Record a real outcome | Whether the decision held up in the business process | A matching test verdict is not proof verification. Re-running inputs is not reproduction of an original model reading. Evidence-aware runs need their supported engine derivation/verification bundle; the console's basic browser proof replay does not validate those runs. The engine also has judgment calibration, outcome analysis, and reviewed policy-threshold adoption surfaces. Keep observed outcomes separate from model confidence and require review of proposed policy changes. These mechanisms do not train a new Jev model on your data. Related: [Verified Decisions](/stateset-nsr-decisions), [NSR console](/stateset-nsr-console), and [first symbolic decision](/guides/nsr-first-verified-decision). Source basis: `stateset-nsr/docs/JEV_INTEGRATION.md`, recursive decision handlers, `src/api/jev_credentials.rs`, and the Jev/Playground implementation in `nsr-app`, reviewed 2026-09-20. No new live-provider accuracy or latency measurements were performed for this guide.