KORP AI — anti-hallucination guardrail snippets (2026-06-01)

output-guardrail.js — Layer 5. Regex-first screen that BLOCKs forbidden promises (full refund / 100% guarantee / free-shipping-no-min) and vague price guesses before the reply reaches the customer. No LLM call.
no-invent-price.md — Layer 2. System-prompt fragment + deterministic injection pattern so the LLM phrases verified numbers but never invents a price/stock/policy.
confidence-handoff.py — Layer 4. Make "I'm not sure" a valid answer: low retrieval similarity or out-of-scope → hand off to a human instead of guessing.
citation-enforcer.py — Layer 3. Every factual claim must carry a [src:id#n] tag; un-grounded claims are suppressed (Thai-aware, one-claim-per-line contract).
audit-logger.py — Layer 7. Append-only, hash-chained conversation log with PII redaction (phone/email/Thai ID) for PDPA ม.30 record-of-processing.
hallucination-eval.py — measure it before you ship: trap-question harness that scores hallucination / over-promise rate. Run weekly.