What the check does
You open a lot and run the check. Our server sends the lot photos to the model, receives the breakdown and shows it in the extension:
- Areas of the car: front, rear, left and right side, roof, underbody, glass, interior, wheels, engine bay. For each: damage severity (none, light, moderate, heavy, or “can’t tell”) and a short note on what is visible.
- Photo references: which photo shows each damage, so you can check it with your own eyes.
- The signs that move the price most: structural damage, deployed airbags, signs of flood or fire, missing parts.
- A check against the declared damage: whether what is visible matches the damage the auction listed.
- A confidence level and a short summary for the buyer.
Which model
The check is performed by Claude, a model made by Anthropic, through Anthropic’s commercial API. LotRadar does not have its own model: we use an existing one and take responsibility for what we send it, how we validate its answer and how we present it to you.
How reliability is ensured
- A strict response format. The model must answer in a defined structure. An answer that fails validation is thrown away entirely rather than shown in part.
- Evidence for every finding. A damage without a reference to the photo that shows it is not accepted. Areas not visible in any photo are not described.
- A second pass on doubtful cases. If the first pass is unsure or contradicts the declared damage, the lot goes through a second verification pass.
- “Can’t tell” instead of a guess. “Not visible in these photos” and “no damage” are different answers, and the check keeps them apart. With too few or poor-quality photos it lowers its confidence and says what is missing.
- Confidence you can see. You see not just the conclusion but how far you can rely on it.
Limits
- Hidden damage is not visible in photos to anyone, human or model. Damage behind panels, frame geometry, electrics, engine and transmission cannot be judged from photos.
- Photo quality matters. Glare, wet paint, darkness and awkward angles hide dents. The auction shows the photos it took, not the ones you need.
- The model can be wrong. That is why every finding is tied to a photo: you can verify it in seconds.
- It is not an expert assessment and not a substitute for inspection. The check saves time when screening lots. Buying decisions are yours.
What is sent to Anthropic
- The URLs of the lot photos on the auction site. The model fetches the images from those URLs; we do not copy them to our servers for this.
- Lot details: year, make, model and the damage declared by the auction.
The request contains no personal data: no email, no name, no account identifier, none of your bids or calculations.
Anthropic processes the request as a processor to return a response. According to Anthropic’s commercial terms, data sent through the API is not used to train its models by default; how long Anthropic retains requests is governed by its own policies. We store the result of the check with the lot data in your workspace. More: Privacy Policy.
Allowances
One check is one run on one lot. The free trial includes 15 checks; after that it depends on your plan: Start 20, Pro 150, Team 600 a month. If you run out, lot calculations keep working, and you can buy more as a pack: +50 for $15, valid for 12 months. More: Pricing.
If the check gets it wrong
Write to the Support section in the extension: the lot number is attached automatically. Reviewing mistakes is how we improve what we ask the model and how we validate its answers.