Trust & test evidence

1. Automated test suite

Every commit runs the full suite. The tests are not smoke tests — they assert specific rule outcomes (for example, that a one-cent discrepancy in the totals chain is rejected, and that a legitimate one-cent VAT rounding difference is accepted).

One structural guarantee is enforced by test: generate → parse → validate round-trips cleanly. A document we produce, re-read from XML, must validate against the same rules — so the generator and the validator cannot silently drift apart.

2. The official XRechnung test suite

We run the public test corpus published by KoSIT (the German coordination office for IT standards) — itplr-kosit/xrechnung-testsuite@v2026-01-31, XRechnung 3.0.2 — through our engine on every parity run.

Both syntaxes are covered: 45 UBL and 41 CII (ZUGFeRD / Factur-X style) documents. Zero exceptions means no document in the official corpus crashed the parser — a baseline we hold to.

The three we flag — published openly

These are the reference documents our engine currently rejects. We publish them rather than hide them: they are open questions about whether our implementation of these two rules is stricter than the reference, and they are the first thing our parity run will settle.

3. KoSIT parity harness — what it is, and what it has not yet proven

The official KoSIT validator is the legal reference for XRechnung. Our harness pins exact versions of it (validator 1.6.2, XRechnung configuration 3.0.2) with byte-exact download verification, runs both validators over the same corpus, and emits a per-document delta report classified into two categories:

• False green — KoSIT rejects, we accept. This is the dangerous direction: it means we would have told you an invoice was fine when it was not. Our target for this number is zero, permanently.
• False alarm — KoSIT accepts, we reject. Noisy but safe; it costs you time, not a rejected invoice.

An honest limitation we will state on the report itself: the official corpus consists mostly of documents that should validate, so it tests the false-alarm direction well and the false-green direction only weakly. A mutation corpus is the next step.

4. Adversarial code review

Before launch the codebase was put through a multi-agent adversarial review across four dimensions — security, rule correctness, API/billing logic, and framework misuse — in which every claimed finding had to survive a second reviewer whose job was to refute it. Twenty-five findings survived that check and were triaged; all high-severity items were fixed before this page was written.

The most consequential fix is worth naming, because it is exactly the class of bug that makes compliance tools untrustworthy: our totals comparison originally used a ±0.011 tolerance on the EN 16931 summation rules, which meant a genuine one-cent-wrong invoice could pass. Those rules are exact equalities; the tolerance is now 0.005 (float-noise only), and the ±0.01 rounding allowance is applied solely to BR-CO-17, the one rule that involves multiplication. There is a regression test for both directions.

Other fixes in the same pass: invalid or revoked API keys now return 401 instead of silently falling back to the anonymous tier, duplicate subscriptions are blocked at checkout, and all API inputs are bounded and rejected if non-finite.

5. What we do not claim

We would rather lose a deal than win one on an overstatement.

• This is not legal advice, and passing Stampbench is not a legal guarantee that an authority or access point will accept your invoice. For certification-grade sign-off, run the official KoSIT validator too. Our job is to make that run pass.
• Rule coverage is a documented subset, not the complete rule set — exactly which rules are implemented is listed in the docs. It grows with every release, and the ruleset version is stamped into every response.
• No parity percentage yet, as stated above.
• No SOC 2 or ISO 27001 certification. We are early; we will say so plainly rather than imply otherwise.
• Generation is UBL today. Validation reads both UBL and CII; Factur-X PDF generation is on the roadmap, not shipped.