Methodology
A short, honest note on how supplier assessments are produced and where the guardrails sit.
The 7-layer compliance stack
Every supplier profile is decomposed into seven layers: Company, Facility, Raw Materials, Process, Products, Packaging, and Logistics. Layers are populated from three source types — human-verified certificates, AI-inferred assessments, and supplier-declared inputs — and each row is labelled with its origin. Verified certificates always outrank AI inferences; AI-only signals cannot mark a supplier as non-compliant.
Trust score
The trust score is a weighted sum of per-layer statuses, floored by verified evidence. AI-inferred raw materials never deduct points — uncertainty is not the same as non-compliance. Where a certificate has expired but no material flags exist, the layer degrades to Partialrather than Non-compliant.
AI usage
AI is used to (a) extract fields from certificate documents, (b) infer company-level compliance from public web signals, and (c) answer knowledge-base questions with retrieved citations. All AI-derived rows are visibly tagged and cannot promote a supplier to Certified; only a human-verified certificate can.
Rate limits (cost containment)
The AI endpoints (knowledge chat, certificate classifier) and the enquiry form have modest daily caps to contain provider cost and reduce form spam:
- Certificate classifier — 10 classifications per user per day (sign-in required).
- Knowledge chat — 20 messages per IP per day for anonymous visitors; 100 messages per user per day when signed in.
- Enquiries — 5 submissions per user per day.
These limits are best-effort cost containment, not abuse defence. They are enforced by a shared daily counter in the database and can be circumvented by determined attackers (e.g. by rotating IPs). Real abuse mitigation requires infrastructure-level rate limiting, which is a separate roadmap item.
Corrections
Every supplier can claim their profile and correct any layer. Non-owners can request a correction via the enquiry form; verified certificates take precedence over AI-inferred assessments in every case.