> ## Documentation Index
> Fetch the complete documentation index at: https://docs.narrativebanking.com/llms.txt
> Use this file to discover all available pages before exploring further.

# AI safety

> AI safety principles and operating boundaries for Narrative SDK assistant surfaces.

<Info>
  Assistant outputs are intended for decision support. They do not replace legal, tax, or regulated professional advice.
</Info>

## Core principles

<Tip>
  Use non-alarming, actionable language focused on user next steps.
</Tip>

<Check>
  Ground responses in authorized data and active product context.
</Check>

<Warning>
  When data is missing or conflicting, state uncertainty clearly and avoid fabrication.
</Warning>

<Info>
  Respect tenant and user access boundaries for every generated response.
</Info>

<Danger>
  Do not provide regulated legal, tax, or financial advice; provide guidance and escalation paths instead.
</Danger>

## v1.5.8 safety updates

`v1.5.8` strengthens graceful degradation, spending explanations, and data-grounded coaching.

* Coach gives clearer limitation messages when a tool, model call, context assembly step, or data source is unavailable.
* Responses should avoid fabrication and offer a next-best action when complete data is not available.
* Spending answers can use meaningful transaction categories where available, improving the evidence users see for financial explanations.
* Sector context, behavioural signals, and research into external support opportunities can provide richer grounding while preserving the boundary between guidance and regulated advice.
* Eligible conversation analytics can support quality review through topic, intent, sentiment, and confidence signals.

## Safety controls in practice

<Columns cols={2}>
  <Card title="Prompt + routing guardrails" icon="https://mintcdn.com/narrative-b13c445c/qi6sqKB_OLTe8ty2/SVG/prompt-guardrails.svg?fit=max&auto=format&n=qi6sqKB_OLTe8ty2&q=85&s=1b80d39c7fb42726d26b571a39bbd617" width="64" height="64" data-path="SVG/prompt-guardrails.svg">
    Detect high-risk or out-of-scope requests and route to restricted handling paths.
  </Card>

  <Card title="Runtime disclaimers" icon="https://mintcdn.com/narrative-b13c445c/qi6sqKB_OLTe8ty2/SVG/runtime-disclaimers.svg?fit=max&auto=format&n=qi6sqKB_OLTe8ty2&q=85&s=528dc3cfb668922d4248e7b47e88464c" width="64" height="64" data-path="SVG/runtime-disclaimers.svg">
    Show explicit user-facing disclaimers for known uncertainty or processing states.
  </Card>

  <Card title="Refusal behaviour" icon="https://mintcdn.com/narrative-b13c445c/qi6sqKB_OLTe8ty2/SVG/refusal-behavior.svg?fit=max&auto=format&n=qi6sqKB_OLTe8ty2&q=85&s=7f2e9e92249aa3eb1c27b298f31fd56f" width="64" height="64" data-path="SVG/refusal-behavior.svg">
    Refuse disallowed requests with clear rationale and safe alternatives.
  </Card>

  <Card title="Validation suites" icon="https://mintcdn.com/narrative-b13c445c/qi6sqKB_OLTe8ty2/SVG/validation-suites.svg?fit=max&auto=format&n=qi6sqKB_OLTe8ty2&q=85&s=cd7d6a89b3789817982e6c08d58a997e" width="64" height="64" data-path="SVG/validation-suites.svg">
    Test tone, scope, and trigger behaviour to catch regressions before release.
  </Card>
</Columns>

## Escalation model

<Steps>
  <Step title="State limitation">
    Explain confidence gaps or missing context explicitly.
  </Step>

  <Step title="Constrain recommendation">
    Avoid prescriptive advice in regulated or high-risk scenarios.
  </Step>

  <Step title="Escalate to human channel">
    Route users to approved human support workflows when needed.
  </Step>
</Steps>