Pitfalls

1. A product ID is not a printing. Normal, Foil and Reverse Holofoil can have different prices. Select the actual printing and condition from valuations.

2. Uncertain identity is not a valuation. confident:false must lead to confirmation; price enrichment is skipped. Agreement in candidate lists does not itself prove agreement on the winner.

3. Raw is not graded. Grade and grader describe a slab. Identification can find its underlying card, but enrichment remains raw/ungraded pricing.

4. Lowest listing is not guaranteed value. Components and empty packaging appear on real product pages. Filters help but are not an authenticity check.

5. Asks are not sales. Read marketBasis; listed median is asking-price data.

6. Unsupported games and absent variants need abstention. Supply hints.game; never silently accept a product from a different game.

7. There is no measured general accuracy claim. Regression fixtures and hand-written examples are not a labeled, representative evaluation corpus.

8. Network failure is not zero market activity. Inspect status and sources.

9. Card numbers vary by game. Pass the number exactly as printed; reprints may keep their original number even when set codes differ.

10. One image may contain many products. Inspect each item with agent vision or OCR first, one blob per item.

11. Rules text is not identity. A Nami card can mention Luffy in its effects. Read the printed title; never substitute a character from ability text.

12. Vision can be wrong too. Do not guess unreadable collector numbers, assume a rare parallel from glare, or let an API score override visible contradictions.