没有护栏的 LLM 不是产品

Three Guardrail Classes

Input guardrails: protect privacy, reject off-topic abuse, and limit oversized uploads.

Inference guardrails: structured output modes, controlled randomness per task, and rubric-grounded context.

Output guardrails: format validation, band-range checks, mandatory disclaimers, and blocklists for official-exam impersonation language.

Output Policy (Conceptual)

{
  "output_policy": {
    "must_include": ["band_estimate", "criteria_breakdown", "actionable_fixes"],
    "must_never_say": ["guaranteed band", "official examiner", "IDP endorsed"],
    "on_low_confidence": "ask_learner_to_retry_with_clearer_input",
    "label": "practice_estimate_not_official_score"
  }
}

Reliability Is the Product Surface

Users don't experience your base model — they experience the guardrail stack. A powerful model with weak guardrails loses to a focused model with rigorous policies in high-stakes domains like test prep.

This is why Band9AI is built as reliability infrastructure first, chat second.

What Breaks Without Guardrails

Without output contracts, a scoring model can invent band numbers outside the IELTS scale, invent criterion names, or imply official endorsement. Learners treat fluent text as truth — so a missing disclaimer is not a UX nit; it is a trust failure.

Band9AI therefore treats guardrails as release gates: if structured validation fails, the UI does not present a polished wrong answer. It retries, degrades gracefully, or asks for clearer input. That behavior is the product users feel.

Related architecture notes: evaluation discipline over prompting, and why LLMs predict rather than decide — both reinforce the same stack: generate, validate, then present.

Shipping Guardrails With Regression

Policies drift when prompts change or providers update models. We keep a golden set of learner responses with expected band ranges and forbidden phrases. Any change to prompts, models, or validators must pass those cases before release.

This is how guardrails stay real in production: not a slide in a pitch deck, but automated checks that block bad generations from reaching candidates preparing for high-stakes exams.