00:00 / Open
Run the public demo.
Open `/demo`, keep the Baby/kids resale sample selected, and click Run public demo scan.
Buyer evaluation guide
Test the workflow before sending a file.
This is the fastest way to decide whether Catalog Recall Monitor is meaningfully different from asking a chatbot once. Use public demo rows, break the matching logic on purpose, check the private-data blocker, and inspect the report artifact.
The test
01:00 / Inspect
Check exact versus fuzzy.
Confirm the exact UPC row becomes Suppress and fuzzy bed rail / BISSELL rows become Manual Review.
02:00 / Break
Remove hard evidence.
Delete the Fisher & Paykel UPC and rerun. The row should stop behaving like a hard suppress case.
03:00 / Block
Paste unsafe data.
Paste a fake CSV with `customer_email` or `wholesale_cost`. The demo should block it before matching.
04:00 / Source
Open a CPSC link.
Click an official source link from a match and confirm the report points back to the underlying recall record.
05:00 / Decide
Use the pass/fail criteria.
If the workflow produces clearer decisions than a one-off prompt, send five redacted product rows.
What should pass
The product earns trust only if these checks work.
These are intentionally practical. A buyer should not need to understand the internals to see whether the workflow is disciplined.
Pass criteria
should happen- Exact UPC or model evidence can support a Suppress recommendation.
- Brand/title/category overlap routes to Manual Review, not automatic suppress.
- Every visible match includes a specific reason and official CPSC source link.
- Unsafe sample data is blocked before scan or download.
- The sample report reads like an operating queue, not vague AI prose.
Fail criteria
should not happen- Fuzzy evidence is treated as automatic suppress.
- The output claims a product is safe.
- The page accepts customer, order, payment, cost, margin, or credential fields.
- Matches appear without a source link or recall ID.
- The buyer cannot explain what to do next after reading the report.
What to send next
Only send a tiny redacted sample if the public test passes.
The first real buyer action is deliberately small: five product-only rows from one recall-sensitive category. The goal is to test whether the report changes work, not to ingest a full database.
1Keep
Product facts only.
Product title, brand, category, SKU, model number, UPC/barcode, product URL, and vendor/source are enough when present.
2Remove
Private/business-sensitive data.
No customer names, emails, addresses, order IDs, payment data, wholesale costs, margins, credentials, or contracts.
3Judge
Did the report change a decision?
Continue only if the first source-backed report helps a real operator suppress, review, or clear work more confidently.
Low-friction next step
Run the demo. Then decide if five redacted rows are worth sending.
If the public test does not make the workflow obvious, do not send data. If it does, use the browser-only sample prep page to clean a tiny product sample.
Legal boundaryThis provides source-backed decision support. It does not certify products as safe, provide legal advice, or replace human safety/compliance review.