Put Deterministic Authorization Beneath Probabilistic Intake
The AI may gather context conversationally.
Capture
- AI gathers observations.
- User confirms relevant facts.
- A reviewed Compact establishes possible eligibility.
- A deterministic tree controls permitted action.
- The user performs or declines.
- The system records result and reroutes.
The AI may gather context conversationally. Concrete physical authorization must be bounded by deterministic logic.
Why
Generative AI is useful for language and clarification but should not improvise the final physical action boundary.
Conversational intake may be probabilistic. Concrete authorization is bounded and deterministic.
Why-Not
- decision trees;
- route transitions;
- eligibility checks;
- failure states;
- exact question ordering;
- testing thresholds.
Commit
Decision: The AI may gather context conversationally. Concrete physical authorization must be bounded by deterministic logic.
Conversational intake may be probabilistic. Concrete authorization is bounded and deterministic.
Public boundary: This ADR publishes the judgment and product boundary while withholding exact prompts, routes, thresholds, schemas, tests, and other reproducible implementation details where applicable.
Confidence: Medium-high for public architecture; implementation details remain private and subject to launch evidence.
Timestamp
2026-07-17