LLM 不做决策

Prediction Layer vs Decision Layer

An LLM completion is a probability distribution over tokens. Calling that a 'decision' is a category error — and a liability error when users treat band estimates as official scores.

Band9AI splits the stack: the model proposes structured feedback; validators enforce format, range, and policy; only then does the UI present an estimate clearly labeled as practice, not examiner truth.

Scoring Workflow (Conceptual)

{
  "workflow": [
    {"step": "generate", "role": "AI proposes criterion-level feedback"},
    {"step": "validate", "role": "Check format, band range, and policy compliance"},
    {"step": "filter", "role": "Block official-exam impersonation language"},
    {"step": "present", "role": "Show practice estimate with clear disclaimer"}
  ],
  "principle": "AI suggests — the system decides what learners see"
}

Why This Matters for IELTS Prep

Students don't need another chatbot that sounds authoritative. They need predictions bounded by public band descriptors, with visible uncertainty when the model is extrapolating.

When designed correctly, LLMs become engines for authoritative insights — not suggestions dressed as decisions.

Decision Ownership in High-Stakes Prep

Who decides what a learner sees? If the answer is 'whatever the model sampled,' you have abdicated product responsibility. Band9AI keeps decision ownership in code and policy: schema checks, range clamps, and disclaimers that refuse official-exam impersonation.

That separation — prediction versus decision — is the same principle Quality IT Resourcing highlighted in its feature on trustworthy enterprise AI, and it is how we keep practice estimates useful without pretending they are examiner marks.