Choosing between rules and an LLM
Stable complete rules should remain deterministic. Expected disposition: Proceed.
Task boundary
Choosing between rules and an LLM is assessed as a specific task and workflow, not as a general endorsement of AI for the surrounding job. Stable complete rules should remain deterministic.
Assumptions
- Inputs, rules, and expected outputs can be completely specified.
- The same valid input should always produce the same result.
- Ordinary unit and boundary tests can determine correctness.
Decision
- Disposition: Proceed
- Pattern: Conventional software
- Inherent risk: low
Rules that drive the decision
- A deterministic solution satisfies the task with less variability and simpler verification.
- The methodology prefers the least complex sufficient pattern, so an LLM adds no fit advantage.
Why alternatives were rejected
- A model adds variability without adding needed judgment.
- An agent adds cost and failure modes.
Controls
- Use typed inputs, explicit error states, versioned rules, and deterministic tests.
- Define how missing, malformed, and boundary values fail before implementation.
Acceptance threshold
These are example starting thresholds for a bounded pilot of this workflow, not universal benchmarks. For choosing between rules and an LLM, the accountable owner should make them stricter when the task, consequence, or policy requires it.
- All normal, boundary, malformed, and historical regression cases return the exact expected result.
- Repeated runs with the same input are identical and no output depends on probabilistic interpretation.
Reassess when
- Inputs become ambiguous natural language that rules cannot completely express.
- The task begins requiring synthesis across incomplete or changing evidence.