# Future AI / Human-in-the-Loop Extension

Status: Draft for Spec 350  
Scope: Documentation only

## Current Rule

Spec 350 does not implement AI.

No runtime AI call, no AI suggestion rendering, and no AI-driven execution path are part of v1.

## Why This Document Exists

The shared case/action contract is a plausible future attachment point for AI-assisted guidance. Documenting that boundary now avoids later feature-local AI drift.

## Reserved Extension Shape

If a later spec enables AI suggestions, the `ResolutionCase` envelope may gain a strictly optional field like:

```php
[
    'ai_suggestion' => [
        'enabled' => false,
        'provider' => null,
        'model' => null,
        'confidence' => null,
        'summary' => null,
        'requires_human_approval' => true,
        'policy_gate' => null,
        'budget_gate' => null,
        'context_refs' => [],
    ],
]
```

That field must stay absent or disabled in Spec 350 runtime work.

## Mandatory Future Gates

Any future AI-enabled follow-up must require:

1. AI policy gate
2. AI context boundary
3. AI budget/cost gate
4. audit trail
5. human approval gate
6. capability gate
7. existing domain action or `OperationRun` execution path

## Forbidden Future Shortcuts

- direct AI execution without human approval
- direct AI writes around existing policy/capability checks
- storing AI-generated resolution truth as canonical truth without a separate spec
- bypassing existing audit requirements

## Current Implementation Guidance

Spec 350 runtime work should leave clear seams for later extension but should not:

- add AI fields to rendered UI
- add AI fields to persisted rows
- add AI-specific copy keys
- add AI-specific jobs or service calls