Decisions · Layer 1
Should loyalty use machine learning, an LLM, or rules
Use rules for obligations that must be deterministic, machine learning for ranking or risk signals, and an LLM for language tasks that can be reviewed. Keep ledger changes behind deterministic controls.
Use rules for obligations that must be deterministic, machine learning for ranking or risk signals, and an LLM for language tasks that can be reviewed. Keep ledger changes behind deterministic controls.
## The decision boundary Rules map known conditions to known outcomes. Machine learning estimates or ranks from patterns. An LLM generates or transforms language.
## The options ### Rules engine Choose this when eligibility, earn, expiry, or redemption must be auditable. The cost is additional operating complexity. ### Machine learning Choose this when a pattern or risk signal needs ranking. The cost is more measurement and exception handling. ### LLM workflow Choose this when the task is language generation or assisted service. The cost is a stronger funding and review obligation.
## What it costs you The currency is liability, margin, engineering time, operating capacity, or member trust. In a hypothetical illustration, 2 points per dollar on 400 dollars of spend creates 800 points. At a hypothetical 1 cent value, that is 8 dollars of liability. Ten comparable cases therefore represent 80 dollars of stated value.
## How to decide 1. Name the behaviour or obligation the rule must control. 2. Identify the independent evidence that proves the event. 3. Price the member, operating, and liability cost of each option. 4. Choose the least complex control that protects the least reversible outcome. 5. Set the evidence that would cause a later review.
## What breaks The failure mode is a rule that measures its own success. Keep source evidence, eligibility, issuance, adjustment, and fulfilment separately observable. Revisit the recommendation when the balance has no meaningful value or the event cannot be independently verified.