Configuration

Return Policy Files

W-Scrut reads your return policy at evaluation time — not at submission time. This means you can update policies instantly without touching any code. The AI reads your Markdown policy and applies it to every claim for that SKU.

File location

Policies are stored as Markdown files in your GitLab repository. The file path must match the SKU exactly:

your-gitlab-repo/
└── policies/
    ├── SHOE-11A.md        ← policy for SKU "SHOE-11A"
    ├── BAG-22B.md
    └── WATCH-33C.md

Policy file format

# SHOE-11A — Classic Runner Return Policy

## Covered Defects (→ APPROVE)
- Sole separation or splitting within 90 days of purchase
- Stitching failure or upper detachment
- Defective hardware (broken buckles, zips)
- Product arrived damaged or different from what was ordered

## Not Covered (→ REJECT)
- Normal wear, scuffs, or surface abrasion
- Damage caused by misuse, water, or chemicals
- Claims submitted after 90 days without a valid exception
- No visible defect in submitted photo

## Escalate to Human Reviewer
- High-value orders (>$300)
- Customer is a repeat claimant (>2 claims in 6 months)
- Photo quality too poor to assess
- Damage is ambiguous (could be defect or misuse)

## Additional Notes
Replacement shoes are sent within 3–5 business days.
Store credit option available for customers outside the return window.

How the AI uses it

The Gemini Vision model receives:

  1. The customer's claim text (their description of the defect)
  2. Their evidence photo(s)
  3. Your full policy Markdown for that SKU

It then determines whether the visible defect is covered under your policy and outputs: APPROVE, REJECT, or ESCALATE — with a confidence score and reasoning.

Fallback policy

If no policy file exists for a SKU (or your GitLab connection is unavailable), W-Scrut falls back to the W-Scrut Standard Policy v2.1 — a reasonable general-purpose policy that approves manufacturing defects and rejects wear-and-tear claims.

Tips for better AI decisions

  • Be specific — "sole separation" is better than "shoe damage"
  • Include time windows ("within 90 days") — the AI checks claim timestamps
  • Add an Escalate section — reduces borderline auto-decisions
  • Use consistent SKU format — exact match to what your store sends
API Reference →Dashboard Guide