Testing antique AI before building the app
The first Cavio milestone is a research workflow and a free guide. The iPhone app and online scanner are not available yet. Before building them, we need to distinguish a plausible identification from evidence that can support a buying decision.
A photo can suggest a form, material, or decorative style. It may not establish the maker, date, authenticity, restoration history, or whether two visually similar pieces are truly comparable. A confident paragraph does not solve that problem.
Our evaluation approach starts with documented museum and auction sources. Keep the source record and labels separate from the anonymous image presented to the model. Group images of the same object together so a crop cannot leak from development into the holdout set. Review the cited pages manually rather than treating a search citation as proof.
Price needs a separate check. An asking price is not a completed sale. For each comparison, check maker attribution, dimensions, material, condition, quantity, currency, and whether the amount includes buyer fees. If the system retrieves the exact historical target sale, record that as retrieval; it does not demonstrate that the model can estimate an unseen object.
The practical guide follows the same habit: photograph the front, back, underside, marks, damage, and scale; write down visible details; compare genuinely similar sold items; then subtract fees, shipping, risk, and a target profit before setting a buy ceiling.
There is no accuracy or revenue claim in this update. The next decision depends on reviewed evidence and real interest in the website. We would rather have a useful “not enough evidence” result than an unsupported number.
The current free Cavio guide: https://cavio.ferryman.app/