Built for Guardian — Vietnam health & beauty retail.
Claude reads. Rules decide. A person approves.
A supplier's pack — claims, ingredients, certificates. In Vietnamese.
By hand. Every page.
Every new product, every week.
This is where onboarding dies.
Neither would we. So we split the work.
One supplier file becomes a reviewed decision — with the evidence, in under a minute.
The trust boundary is proven in code — not promised. A test proves no path but a human's can approve.
One real flagged case (D1) traces end to end — the agent plans the run, calls its tools, drafts an action, routes it to a person, and verifies the result. All visible in the case view.
plan → vision_extract → cross_check_extract → cross_check_compare → rulepack_gate → council_defender / council_prosecutor / council_chairman → council_merge → persist_and_queue (verify)
Runs only on ambiguous flags (capped per case). Two Sonnet advocates argue; an Opus chairman drafts a ruling with confidence. Split or low confidence → auto-escalate with both arguments attached. Merging hard-asserts the gate stays FLAG — a suggestion is never a state change.
Supplier packs are confidential, so demo dossiers are authored synthetic PDFs (organizer-blessed) — which means we know the right answer for every one. Each page carries a "TÀI LIỆU MÔ PHỎNG" footer; generators are seeded and deterministic, checked by 117 validation assertions.
The demo uses a labeled synthetic price snapshot; every observation records its source and provenance. The production path — Apify actors run out-of-band, never from inside the VPC — is already wired in the collector CLI.
Impact figures (70–80% / 90%) are Guardian's targets, not measured results — we automate the mechanism behind them and say exactly that.
| Input | Value | Basis |
|---|---|---|
| New SKUs onboarded per month | 100 | assumption |
| Manual first-pass per dossier | 45 min | assumption |
| Manual onboarding effort at that volume | 75 h / month | derived |
| Onboarding effort reduction | −70–80% | Guardian target |
| Effort freed if the target holds | ≈ 52–60 h / month | derived |
| Price monitoring reduction (Top 200 SKUs) | −90% | Guardian target |
Mine officer decisions and propose new lexicon rules — each with a citation and evidence. A human rule committee approves before anything applies.
The gate can learn — but only a person changes the law.