AI: Remediation Actions¶

Safe mutations an AI can apply.

Overview¶

When validation fails, Invariant returns remediation actions — structured suggestions the AI can apply to fix the query.

Safe mutations¶

Remediation actions are designed to be:

Bounded — Limited scope of change
Reversible — Can be undone
Auditable — Logged for review

Available actions¶

Action	What it does
`DEFINE_NUMERATOR_DENOMINATOR`	Add indicator definition for recomputation
`USE_NONE_AGGREGATION`	Change aggregation to display-only
`APPLY_CROSSWALK`	Add crosswalk for version mapping
`AGGREGATE_TO_STABLE_LEVEL`	Roll up to boundaries that didn't change
`ACKNOWLEDGE_DIFFERENCE`	Accept comparison with disclosure
`APPLY_SUPPRESSION`	Apply suppression policy

Action schemas¶

Each action has a defined schema:

@dataclass(frozen=True)
class RemediationAction:
    action: str
    description: str
    parameters: dict[str, Any]
    requires_confirmation: bool

Applying remediations¶

# AI receives validation result with remediations
result = kernel.validate_query(query)

for issue in result.issues:
    for remediation in issue.remediations:
        if remediation.action == "USE_NONE_AGGREGATION":
            # Apply the fix
            query = apply_none_aggregation(query, remediation.parameters)

# Re-validate
result = kernel.validate_query(query)

Confirmation requirements¶

Some actions require human confirmation:

Action	Requires confirmation
`USE_NONE_AGGREGATION`	No
`ACKNOWLEDGE_DIFFERENCE`	Yes
`APPLY_SUPPRESSION`	Depends on policy

Auditability¶

All applied remediations are logged:

AuditEvent(
    type="REMEDIATION_APPLIED",
    action=remediation.action,
    applied_by="ai_agent",
    query_id=query.id,
    timestamp=now()
)

This ensures humans can review AI-applied fixes.