102 inventory photos to 22 live eBay listings: why right answers for wrong reasons fail validation
Ok_Appearance_7559 · reddit · 2026-09-17
A developer ran a second batch of 102 raw photos covering 22 resale items through their photo-to-eBay-listing pipeline; all 22 published, 17 precisely identified, 5 kept generic due to insufficient evidence, only 4 needing outside research.
Key findings:
- Marketplace behavior itself needs verification: one price mutated after submission, one category required a safety disclaimer, one shipping trigger forced manual review
- The most interesting failure mode: correct answers reached via unsupported reasoning. One item with two identical units got the right quantity of 2 through an insufficient evidence path — counted as a failure. A regression now fails validation for correct-but-unsupported outputs
- Models cooperate with your premise: two passes can agree on the same bad assumption; suggested identifications get supported rather than refuted
- Fix: move authority away from the model — unsupported info stays blank or gets kicked back; confidence only prioritizes review; reversibility, financial exposure and account risk determine permissions
Error rate: across 31 real items, 1 system-caused material error (3.23%), 2/31 (6.45%) including the marketplace price mutation. Takeaway: identifying an item once isn't hard anymore — reliable, well-reasoned operation at scale is.
More from coding & agent
- Dev builds model router with Jev to auto-route requests to the best-fit LLM — emeka_boris · 2026-09-17
- GPT Work Can Now Read and Reply to Email — Inboxes Go Robotic — johnseach · 2026-09-17
- Redditor wants an AI agent to phone 15 venues and return structured quotes — Evening_Hawk_7470 · 2026-09-17
- 7 scary-good sub-1,000-star GitHub AI agent projects for browser automation — JafarNajafov · 2026-09-17
- Showly turns AI coding agent output into deployable, versioned, shareable sites — LearnWithBishal · 2026-09-17
- Open-source ComfyUI workflow stitches 4 MiniMax H3 clips into seamless 29s video — ssx0731 · 2026-09-17