Every figure with its date and the document it was measured with. If one drops, it shows as dropped. And where we lose, too.
The canary exercises the product in production every 30 minutes. Last round: 2026-09-24T01:50:33+00:00 · 30/30 quality checks · green
Core Web Vitals p75 · n=1 · LCP 0.00 s ✓ ≤ 2.0 s · INP 0 ms ✓ ≤ 150 ms · CLS 0.000 ✓ ≤ 0.08
50/50 · 20/20
how it's measured 50 trap questions on 5 documents we know word for word (contract, policy, payslip, minutes, quote): it must say “it isn't there” without giving any figure. And 20 answerable ones: the data and the right page.
7/7
how it's measured A contract and two invoices through the 7 AI tools that return amounts, dates or deadlines: every figure must be in the document or follow from it by a calculation we redo.
227/227
how it's measured OpenPEPPOL's official unit tests (BIS Billing 3.0), each with what the official validator must say.
0 · BR-CO-16 ✓
how it's measured The rules published by the European Commission, run as they are: the official example passes and, with the total tampered, the right rule fires.
3/3
how it's measured The official examples from the Spanish Tax Agency's hash specification (record, chained record and cancellation).
161 · 182 · ✓
how it's measured The official example of Gipuzkoa's Foral Order 521/2020; and a consistent forgery —the same amount altered on paper and in the QR— that the check code exposes.
PAdES-BASELINE-LT
how it's measured Signed in production and validated with the European Commission's official validator (DSS).
1000 · 5.9 s · 37/37
how it's measured A ZIP with 1,000 invoices, 37 of them tampered: time in production and whether the report flags exactly the tampered ones.
0
how it's measured All 159 tools run in Chinese and compared with Spanish: interface texts left untranslated.
159/159 · 159/159
how it's measured All 159 tools, end to end, in Spanish and in Chinese, also listening for browser errors.
0.6 s
how it's measured Largest paint of the home page with the CPU 4× slower and slow 4G.
0
how it's measured Serious WCAG 2 (A and AA) violations according to axe, on the main pages.
/v1/video/join 77.2 s > 10 s · /v1/ai/translate-office 23.4 s > 10 s · /v1/ai/study 21.1 s > 10 s · /v1/pdf/translate 11.5 s > 10 s
Of 52 tools measured, 19 lose something (metadata, links, layers…). We tell you BEFORE processing, tool by tool, with what YOUR document would lose.
At each release, the request that coincides with the switch gets a 502 from the hosting provider. The web retries it after 2 seconds; API users will see it.
Peppol rules for CII invoices and country-specific rules (such as Germany's XRechnung) are not applied yet.
The check code is verified against Gipuzkoa's official example; for Bizkaia and Álava we have no verified official example (they share the technical standard).
If you don't upload yours, the signature proves the document wasn't touched, not who you are: the official validator marks it as “no chain of trust”.
7 of the AI models we use are 18 months old or more. They don't fail, but may answer worse and cost more; they're replaced when a new one wins on our bench, not because of their date.
Download the documents and what should come out, and run them through PAPYR or any other product.