Evidence record · October 11, 2026
Here is what
we tested.
68 local checks passed. A separate production run used the real model on a fictional product family. Neither establishes real-world accuracy or superiority to another product.
What does a passing check mean?
| Test group | Passed | What it covers |
|---|---|---|
| Catalog and evidence | 21 | Identity, source quotes, decisions, simple metafield scope, preserved values, export and comparison. |
| File readers | 10 | A valid text-based PDF, file limits, encoding, formulas and actual compressed-workbook expansion, including forged size claims. |
| Payment boundaries | 15 | Mocked order verification, expired or interrupted checkout, partial refunds, disputes and competing updates. |
| Export delivery | 9 | Real export and comparison routes with mocked payment and database responses: paid access, file integrity, revoked access and recheck limits. |
| Private access | 8 | Token handling, expired access, request limits and restricted editing states. |
| Published interface | 5 | The sample CSV, documented workflow and private-operation access requirements. |
These are reproducible development tests. The payment checks use mocked Stripe and database responses; they are not purchases, refunds or disputes involving real money. A passing file test is not a security certification.
Did the real model complete a repair?
On October 11, 2026, a separate run on specstead.com submitted the public fictional example: two SKUs and two TXT source documents. The configured gpt-6.1-sol model ran extraction and an independent review. The resulting report contained four proposed corrections in total, including one unresolved source conflict, and no held fields.
The same run verified private-link recovery, denial of an unpaid download, rejection of stale review choices, checkout reuse and recovery after checkout expiry. It created a live, unpaid $299 USD card checkout and expired it. No card was charged. This verifies one synthetic workflow through production; it does not establish an accuracy rate, a completed paid order or successful Shopify import.
Which catalog mistakes are covered?
| Case | Tested boundary |
|---|---|
| Exact item identity | A sibling’s quote, a partial SKU match, or a passage naming several catalog SKUs cannot justify a correction. |
| Shared product fields | A field shared by several variants changes only when every included SKU has matching evidence. |
| Conflicting sources | Different supported values stay visible as alternatives. The source date alone does not choose a winner. |
| Missing or unsafe information | Empty values, invented replacements, formula-like text and instruction-like passages do not become corrections. |
| Weight and images | Weight arithmetic is checked. An image replacement must be a literal supplied HTTP(S) URL; the test does not inspect the image. |
| Preserved catalog values | Accepted changes leave identifiers, options, prices and inventory as they appeared in the uploaded snapshot. |
| After-import comparison | Shuffled rows are matched by SKU. A missed intended change and a change to the wrong SKU are both detected. |
We also maintain ten fictional product families: 45 SKU records, with 18 proposed corrections and eight required abstentions in the broader design set. That set includes product-versus-package dimensions, document versions, unit conversion, sibling identity, unknown facts, hostile text, image associations and an after-import error.
The design set is broader than the first Shopify adapter. Those 18 target corrections are not 18 measured successful repairs. General metafield conversions and automatic document precedence are not supported. The tests separately verify that a 25-SKU family fits and a 26-SKU file is rejected.
How did other products perform?
We have not completed a head-to-head processing test. We inspected official product pages and attempted one free upload. There is no competitor accuracy score.
| Product | Documented capability | Our test status |
|---|---|---|
| Microcity Labs | Supplier-file conversion, source tracing and a free two-page preview. One-time conversion starts at £15. | Preview opened. Our local browser upload setting prevented sending the fictional file. No result was produced; this was not a vendor failure. |
| WISEPIM | Source fact-checking, an audit trail, variant handling and rules that block or quarantine exports. | Official features and the free onboarding route inspected. No account created and no processing test completed. |
| Matrixify | Shopify product and variant imports, exports, metafields, image mappings and dry runs. | Official format and processing documentation inspected. No Shopify app test completed. |
Source tracing and export controls already exist elsewhere. Our proposition is a finite product-family repair that you can inspect. Whether that is a better fit for your job depends on the supported fields and your sources.
Where does the evidence stop?
These checks do not establish that a supplier document is authentic, current or complete. They do not certify a whole storefront, promise fewer returns, or measure sales uplift. A full snapshot also contains the prices and stock values from the time you uploaded it: importing it later can restore stale values. The handoff explains that boundary.
Read the file and evidence guide, or check a product family before deciding whether to buy an export.
Sources and date
Official pages reviewed October 11, 2026: Microcity pricing; WISEPIM capabilities and pricing; Matrixify pricing and dry-run limits; Shopify CSV behavior. Local test record: Specstead 1.0.0, October 11, 2026.