I built an AI 8D reviewer that isn't allowed to invent evidence (try it with a sample report)

A developer built 8D Gate, a free AI tool that reviews automotive and industrial 8D quality reports before they are sent to customers, returning an advisory ACCEPT, ACCEPT WITH CONDITIONS or REJECT decision with per-section scores and a fix list. The tool uses a fixed weighted rubric, hard caps the model cannot argue past, and labels every finding as FACT, CLAIM, ASSUMPTION, MISSING or INCONSISTENCY, refusing to treat a filename or 'see attachment' as evidence. It runs on Next.js and Vercel with Anthropic for review, accepts PDF, DOCX and Excel input, requires no signup, and is English-only and advisory-only at version 0.1.

If you work in automotive or industrial manufacturing, you know the 8D report : when a supplier ships a defect, the customer demands eight disciplines of problem solving team, problem, containment, root cause, corrective action, verification, prevention, closure . If you don't, here's the short version: it's the document that decides whether a supplier gets trusted again. Most 8Ds don't fail on the hard engineering. They fail on completeness: root cause = "operator carelessness", corrective action = "retrain + 100% inspection", effectiveness = "verified" with no data. The customer bounces it, and the loop costs days. So we built 8D Gate : a free, skeptical pre-send check. Paste the text or upload PDF / DOCX / Excel, and get an advisory ACCEPT / ACCEPT WITH CONDITIONS / REJECT with a score per section and a fix list. 👉 Try it: https://8d-gate.vercel.app/?utm source=devto&utm medium=community&utm campaign=feedback24h https://8d-gate.vercel.app/?utm source=devto&utm medium=community&utm campaign=feedback24h 1. A fixed, weighted rubric. Root cause 25%, evidence quality 20%, corrective action 15%, problem definition / containment / verification 10% each, prevention and consistency 5% each. Nicer wording can't raise the score. 2. Hard caps the model can't argue past. 3. Evidence labels. Each finding is FACT, CLAIM, ASSUMPTION, MISSING or INCONSISTENCY. A filename, "see attachment" or "available on request" is never a FACT, because the text isn't there. When something can't be judged, the output says "Insufficient evidence to conclude." We'd take that over a confident wrong answer. 4. Golden tests. A set of synthetic 8Ds with known expected decisions REJECT for unsupported root cause, symptom-only CA and contradictions; CONDITIONAL for weak D6 and a missing escape analysis; ACCEPT for solid ones . A run only passes if the decision matches exactly. Stack: Next.js on Vercel, Anthropic for the review, pdf-parse / mammoth / xlsx for extraction, pdfkit for the downloadable report. No signup. The 8D body isn't stored in our database it's sent to Anthropic for processing . English only, V0.1, advisory only. No 8D of your own? Copy the fictional report below company names are made up , paste it into https://8d-gate.vercel.app/review?utm source=devto-sample&utm medium=community&utm campaign=feedback24h https://8d-gate.vercel.app/review?utm source=devto-sample&utm medium=community&utm campaign=feedback24h and see what it does with "operator carelessness" as a root cause. Our expected result is a REJECT for an unsupported root cause; tell me if it says otherwise. 8D Report — Brake Hose Clip Dimensional Escape | Field | Value | |-------|-------| | 8D No. | ACME-8D-2026-0142 | | Customer | Acme Auto — Plant Bratislava | | Supplier | TierCo Fasteners s.r.o. | | Part No. | TC-BH-4412 / Clip, brake hose retainer | | Complaint No. | AA-Q-88921 | | Date opened | 2026-03-12 | | Date of this revision | 2026-03-28 | | Severity | Customer line stop risk sorted at customer | --- D1 — Team | Role | Name | Function | |------|------|----------| | Champion | M. Horvath | Quality Manager | | Team lead | P. Novak | Process Engineer | | Member | A. Kovac | Production Supervisor, Line 3 | | Member | L. Benes | Incoming / Final Inspection | | Customer contact | J. Meier | Acme Auto SQE | --- D2 — Problem description On 2026-03-11, Acme Auto Plant Bratislava reported that brake hose retainer clips TC-BH-4412 from delivery note DN-78441 shipped 2026-03-04, qty 4,800 pcs failed to snap onto the hose during vehicle assembly. Operators observed clips that “felt loose” and did not retain the hose after hand fit. - Symptom: clip opening width oversize; hose retention force insufficient customer feel-check . - How detected: assembly station B12, Acme Auto Bratislava. - Quantity affected at customer: 37 pcs confirmed nonconforming from one tote; full tote qty 500 pcs quarantined by customer. - Supplier stock: 12,400 pcs on hand same tooling family placed on hold 2026-03-12. - Spec: opening width 12.40 ± 0.15 mm; retention force ≥ 35 N per drawing TC-BH-4412 Rev C. Problem statement: Clips from lot L260302A exhibited opening width above upper spec, causing failed hose retention at customer assembly. --- D3 — Containment 1. Customer sorted remaining tote 500 pcs : 37 NOK, 463 OK customer report dated 2026-03-12 . 2. Supplier hold of FG stock 12,400 pcs effective 2026-03-12 14:30 CET owner: L. Benes . 3. 100% dimensional check of FG stock started 2026-03-13 using digital caliper opening width . Results to date: 412 / 12,400 checked; 9 NOK found and scrapped. Sort ongoing. 4. No further shipments of TC-BH-4412 until hold release approved by Quality Manager. 5. In-transit: none per logistics ASN closed . Break point: last conforming shipment prior to DN-78441 was DN-78102 2026-02-18 . --- D4 — Root cause Occurrence root cause: Operator carelessness during setup of the progressive die. The operator did not pay enough attention when adjusting the forming station, which led to an oversized opening width. Escape root cause: Final inspection did not notice the problem. Method: Team discussion on 2026-03-14. Ishikawa sketch prepared attachment referenced as “Fishbone v1.pptx” — not attached to this PDF export . Why this is the root cause: Experienced operators normally produce good parts; therefore the issue must be related to attention during setup. No capability study, setup sheet comparison, or tool wear measurement is included in this revision. --- D5 — Corrective actions | | Action | Owner | Due | Status | |---|--------|-------|-----|--------| | CA1 | Retrain all Line 3 operators on “careful setup” and awareness | A. Kovac | 2026-03-20 | Done 2026-03-19 | | CA2 | Remind inspectors to check opening width carefully | L. Benes | 2026-03-20 | Done 2026-03-19 | | CA3 | Post a quality alert on the line board | P. Novak | 2026-03-18 | Done | Interim action: continue 100% sort of FG until CA1/CA2 complete. --- D6 — Implementation & validation - CA1 completed: training attendance sheet signed by 11 operators 2026-03-19 . - CA2 completed: verbal reminder in morning meeting. - Effectiveness: “No further customer complaints received since training. Therefore corrective actions are effective.” - No before/after opening-width distribution, no retention-force retest data, and no process capability after training are provided in this report. --- D7 — Prevent recurrence - Add “be careful during setup” to existing work instruction WI-LINE3-04 revision pending . - Share lessons learned in monthly quality meeting scheduled April 2026 . - Similar parts: “will be reviewed” — list not yet generated. --- D8 — Congratulate / close Team is thanked for fast containment response. 8D proposed for closure pending customer acceptance of this report. Residual open items: FG sort not finished; WI revision not released; similar-part review not started. --- Attachments listed not embedded - Customer complaint AA-Q-88921 PDF - Sort log excerpt 412 pcs - Training attendance 2026-03-19 - Fishbone v1.pptx referenced, not attached A comment with what landed and what missed : was the verdict fair, was anything confusing, would you trust it on a real report? Especially valuable if you're an SQE or quality manager. Tell me where it's too lenient.