Insights / AI Operations · · Updated · 7 min read

AI explains, verified data decides

Design AI explanations around authorised source records, explicit business rules and semantic checks. Schema-valid output alone does not establish factual accuracy.

AI explains, verified data decides is a useful design principle when a product turns operational records into a readable answer. Retrieve the permitted evidence, apply the relevant rules outside the narrative, and verify that the generated explanation preserves both. Treat the output as a candidate until those checks pass.

This guide proposes a review method for digital-product teams. It does not establish that every Oryvelon company has implemented these controls, passed an audit or launched the features described. The examples are hypothetical. Group context is available on the companies page.

The problem in one hypothetical example

Suppose a merchant asks how many black phone cases remain. A generated number is not an inventory observation. The application should first check the caller’s store access, identify the correct product variant and location, and retrieve the intended inventory quantity with its observation time.

Even an API result needs interpretation: available, reserved and physically on-hand stock can mean different things. If the source is unavailable or ambiguous, show that limitation instead of filling the gap. See inventory truth for related questions.

What verified data decides means

A trusted system is a candidate source, not a blanket guarantee of correctness. Define the record, its authority for the question, its version or observation time, and the identity allowed to use it. Reconcile discrepancies before representing a value as settled.

  • Source fact: a recorded quantity, transaction or policy text, with provenance.
  • Rule result: a calculation or permission outcome derived under an identified rule version.
  • Interpretation: an explanation, hypothesis or suggested next step that must remain distinguishable from those records.

A deterministic rule can still be wrong or outdated. Give it an owner and tests, and avoid treating a stored eligibility result as legal advice or an authoritative decision by an external institution.

What AI explains means

A language model can draft a summary from selected records, organise supporting details or make an explanation easier to read. Constrain the task: what evidence may be used, what is unknown and which conclusions are prohibited without further checking.

For a hypothetical operational briefing, keep the exact figures visible next to generated text. Require references to the source records where practical. A source citation helps trace a claim; it does not prove that the source supports the interpretation or that the information is current.

Controls to evaluate before delivery

Use several checks with different responsibilities:

  • Access: authorise retrieval and tool calls in application code; do not delegate the permission boundary to a prompt.
  • Structure: validate required fields, types and permitted values. Handle refusals and incomplete responses explicitly.
  • Meaning: compare record identifiers, quantities, dates and rule outcomes with the evidence supplied. Reject unsupported additions and contradictions.
  • Behaviour: evaluate missing sources, conflicting inputs, adversarial text and failure paths before a release.

The OpenAI Structured Outputs documentation distinguishes schema adherence from correctness: structured responses can still contain mistakes, and refusal or incomplete-response handling is necessary. A valid JSON object is not factual verification. See structured-output validation for the related design topic.

Why a better model does not replace the evidence path

Evaluate models on the actual task rather than assuming that a newer or larger model makes every failure less likely. Model capability does not supply current private records or establish the caller’s permission to view them.

Keep retrieval, calculations and authorisation testable independently of narrative generation. If a model changes, check whether it preserves source values, admits unavailable information and follows the permitted action boundary. A lower measured error rate still does not make an unsupported individual answer reliable.

Apply the distinction to consumer products too

In a hypothetical consumer application, entertainment text and account facts have different roles. Generated reflection should not invent a purchase, credit balance, subscription entitlement or professional conclusion. Retrieve account status from its authorised source and enforce entitlement in code.

Describe the intended limits clearly and test whether outputs respect them. A label alone does not establish safe behaviour. This is design guidance, not a claim about ZodiVela’s current billing, content checks or availability.

A proposed evidence-to-explanation flow

  • Resolve: authenticate the caller and authorise the requested scope.
  • Retrieve: obtain relevant evidence, freshness and provenance.
  • Decide: apply versioned business rules or calculations where required.
  • Generate: request an explanation using only the permitted material.
  • Check: validate structure, source consistency and output limits.
  • Deliver or degrade: show accepted output or a clearly labelled limited result.
  • Observe: record useful errors and outcomes with appropriate data minimisation.

A shared AI gateway may organise provider calls, but it does not automatically authorise retrieval or verify truth. Review tenant isolation across caches, files and derived answers as well as the database.

Common failure modes to test

Retrieved records and documents can contain misleading content or instructions. Keep external material separate from trusted application instructions. The OWASP prompt-injection guidance explains that retrieval does not eliminate this risk; restrict tool permissions and assess effects beyond the generated text.

  • Accepting a schema-valid but unsupported number.
  • Converting an absent field into zero or a confident conclusion.
  • Treating a retrieved instruction as permission to run a tool.
  • Reusing an answer after its source or access rights changed.
  • Retrying indefinitely or silently expanding the permitted action.

What the interface should make clear

Show which values came from records, when they were observed, and which text is generated interpretation. If an important source failed, say what is missing and offer a useful next step. Do not present a partial answer as a complete check.

Provide a way to inspect relevant evidence or request review without revealing another user’s data. Avoid promising that an AI-assisted answer is error-free. These are proposed interface requirements, not verified descriptions of every Oryvelon product.

Start product design with evidence and failure behaviour

Before drafting a prompt, list the questions the feature answers and the facts each answer needs. Identify missing records, stale rules and permission boundaries. Define what can be delivered safely if one input is unavailable.

Keep consequential actions separate from suggestions. A proposed refund, price change or access grant should require its own permission checks and review appropriate to the action. For related planning questions, see human escalation and degraded modes.

A proposed release checklist

  • List sources, owners, rule versions and acceptable freshness.
  • Test permitted and forbidden access using representative identities.
  • Check accurate values, unsupported conclusions and missing-data behaviour.
  • Handle refusal, timeout, incomplete output and validation rejection.
  • Limit retries, tool permissions and resource use.
  • Label generated content and preserve useful source references.
  • Define review responsibility and the correction path.

The OpenAI evaluation guide describes testing model outputs against specified criteria. Use representative and difficult cases, keep test data appropriately protected, and record known limitations. Passing a finite evaluation set is evidence about those tests, not proof of every future output.

What to measure after release

Track source availability, stale-data incidents, invalid outputs, semantic mismatches, escalations and corrections. Review latency and cost alongside answer quality rather than replacing quality with a faster response metric.

Record prompt and model versions so later observations can be investigated. Do not assume a change caused an improvement just because it came first. These are suggested measurements, not reported Oryvelon results or first-hand deployment lessons. See prompt versioning and evaluation.

Worked example: a hypothetical refund briefing

Imagine an authorised store report shows more refunds for a product over a defined period. Before generating a briefing, verify the product, reporting window, denominator and baseline. A refund count without comparable order volume can be misleading.

The explanation may identify the observed change and propose checking the listing or variant records. If notes mention sizing, distinguish a reported reason from a confirmed cause. Do not invent a percentage, assert causation or execute refunds.

If generation fails but the underlying report is valid, a labelled factual summary may remain useful. If the source itself failed, do not present its prior values as current. This example does not claim that MerchNivo has implemented or passed these checks.

Summary

Build the answer around authorised evidence and explicit rule outcomes. Treat generated explanation as a candidate, then check its structure and its meaning. Preserve uncertainty and source freshness, and keep action permissions outside the narrative. Use the checklist to identify what needs testing; the checklist itself is not proof of a deployed control.

For Oryvelon’s broader company-builder context, see Build.

Questions and answers

Does schema-valid AI output prove factual accuracy?

No. A schema constrains structure and permitted values. Compare the content with authorised evidence and handle unsupported claims, refusals and incomplete responses separately.

What should happen when source data is missing?

Identify the missing source and limit the answer. Do not replace an absent record with an invented value or present a partial check as complete.

Does this guide verify controls in Oryvelon products?

No. It proposes engineering and interface checks with hypothetical examples. It is not an audit, deployment record or guarantee about every company’s AI features.

NextShared infrastructure, separate data →