Transparent product evidence
Radish has preregistered a US grocery receipt-recognition benchmark covering 300–500 consented receipts from at least 10 retailers. The external one-shot study has not run, so no public field-accuracy percentage is available.
Last updated: August 16, 2026.
Current status
No external field result exists today. The study cannot read the hidden receipt set until every frozen development gate passes on the exact source revision and the external corpus completes consent, rights, privacy, and dual-truth review.
The current development baseline does not authorize holdout access. Publishing the protocol now makes later changes visible and prevents thresholds, metrics, or slices from being selected after seeing the field result.
Frozen scope
| Dimension | Frozen requirement |
|---|---|
| Population | 300–500 consented US grocery receipts from at least 10 retailers |
| Language and currency | English-language receipts and USD totals |
| Product surface | The production iPhone app on a physical device |
| Measured moment | The exact pre-edit draft shown on the final Review screen |
| Truth | Independent dual review with exact asset and final-truth SHA-256 bindings |
| Visibility | Unseen locked holdout, consumed once only after development gates pass |
Preregistered metrics
| Metric | What it measures |
|---|---|
| Merchant exact match | Whether the reviewed merchant matches the verified receipt merchant after one frozen normalization rule |
| Local date exact match | Whether the proposed receipt date is exactly correct |
| Printed total exact match | Whether the proposed printed total matches to the cent |
| Line-item precision | How many proposed grocery lines match verified receipt items |
| Line-item recall | How many verified receipt items appear in the proposed draft |
| Line-price error | Mean absolute cent error for matched item occurrences |
| Printed-total reconciliation | Whether the pre-edit draft is arithmetically balanced to the printed total |
| Hard-failure rate | Receipts that do not reach a saveable Review draft |
| Corrections per receipt | Median visible field or line corrections before save, for saved receipts |
| Scan-to-save time | Median seconds from capture start to durable save, for saved receipts |
Required coverage
The same metrics are reported for preregistered slices with minimum receipt counts. They cover short, medium, long, and very-long receipts; multi-page receipts; clean, crumpled, faded, glare, shadow, cropped, and low-resolution images; and receipts containing discounts, weighted items, returns, tax, deposits, or fees.
A valid external corpus must meet every slice floor before scoring. Missing a difficult slice is a contract failure, not permission to publish a more favorable aggregate.
One-shot method
The benchmark adds no authentication bypass. Physical-device evidence must use the production app and its real protected receipt-processing path.
Publication contract
Every published rate must include its numerator, denominator, measured value, and Wilson 95% confidence interval. The report also carries the exact receipt, retailer, page, and item counts, the preregistered slice counts, the mean line-price error, and the saved-receipt medians.
Only aggregate and preregistered slice results can be published. The public report type cannot encode case IDs, receipt text, retailer identities, predictions, truth, timing records, correction records, image paths, or receipt bytes.
A favorable number is not enough. If source identity, development readiness, consent, corpus coverage, private-file integrity, producer binding, or one-shot state is invalid, no field result is eligible for publication or marketing.
Interpretation limits
No. The external one-shot field study has not run, so Radish does not publish or market a field-accuracy percentage. This page publishes the frozen method and current status.
It will measure exact merchant, date, and printed-total matches; line-item precision and recall; line-price error; total reconciliation; hard failures; corrections; and scan-to-save time at the pre-edit final Review stage.
The frozen plan requires 300–500 consented US grocery receipts from at least 10 retailers, in English and USD, with documented rights, privacy review, and independent dual review of the final truth.
Single use prevents the team from repeatedly inspecting the same hidden receipts and tuning the product or scoring rules to improve that known result.
No. Raw receipts, merchant-level identities, truth records, predictions, timings, and correction records stay outside Git, CI, and the public report. Only aggregate and preregistered slice results can be published.
No. Synthetic tests help find regressions and must pass before the hidden field study can run, but they are not a representative estimate of real-world grocery receipt recognition.
Evaluate the workflow yourself
Free to download. Review and save one receipt before choosing a monthly or yearly plan for continued use. Requires iOS 18 or later. Optimized for English-language US grocery receipts and USD totals. Manual entry remains available.