QuoteGuard

Problem

AI support copilots produce confident customer replies while the facts underneath keep changing. A price can be retired, a feature can remain beta-only, and a policy promise can have no approved source. A citation from a model is not authority. Support leads need a fast pre-send decision surface that explains why a claim is unsafe, who owns the fact, and what can be approved.

Solution

QuoteGuard is a pre-send authority gate for AI support replies. It extracts atomic price, capability, policy, and unknown claims, evaluates them against a versioned fact ledger, renders a closed-world corrected draft from allowed facts only, and requires a human decision before the text can leave the desk. It records a tamper-evident receipt and can recheck an approved decision when the underlying authority changes.

Product features

• Deterministic BLOCKED, ALLOWED, and UNKNOWN claim dispositions.
• Fact ledger with source, owner, status, effective dates, beta handling, and version updates.
• Closed-world corrected drafts that never restate blocked values.
• Human approve and reject workflow with signed reviewer identity.
• Durable SQLite WAL history, shareable JSON/Markdown receipts, and receipt verification.
• Fact-owner mutations with authorization and before/after audit events.
• Would-send dry run and authority drift detection with STALE_AFTER_APPROVAL and REVIEW_BEFORE_SEND.
• Headless /v1 API and optional webhook delivery status for support integrations.

GPT-5.6 and Codex

GPT-5.6 is a bounded structured claim-extraction adapter. It can propose claims from unfamiliar wording, but deterministic TypeScript merges ledger-backed matches and owns every final disposition, corrected customer text, approval, and receipt signature. The UI identifies LIVE versus OFFLINE operation and falls back explicitly when live extraction is unavailable.

Codex built the versioned authority ledger, paraphrase-aware matcher, state machine, SQLite store, reviewer and fact-owner authorization, signed receipts, drift monitor, operator desk, headless API, webhooks, tests, and demo. The core design decision was to keep model output away from authority and approval.

Judge proof

Run QUOTEGUARD_DEMO_MODE=1 npm start. Click Verify against ledger on the prefilled reply. Inspect retired pricing as BLOCKED, beta SSO as BLOCKED, and an unsupported promise as UNKNOWN. Approve the corrected draft, open the receipt, tamper it to show VERIFY FAIL, then retire a supporting fact and recheck the approved decision to surface stale authority.

Built with

Node.js 20, TypeScript, SQLite WAL, zod, OpenAI GPT-5.6, Codex, and a headless HTTP API.

Built With

• codex
• gpt-5.6
• node.js
• openai
• sqlite
• typescript
• zod