1. Financial independence
EvalQA never takes a success fee tied to whether an agent passes or fails. Not from the customer, not from a platform vendor, not from an implementation partner. Fees are fixed and published before an engagement starts. A finding that costs a customer a launch date costs us nothing, and a clean run earns us nothing extra.
2. Developmental separation
EvalQA personnel never author production agent prompts, semantic models or dbt models for an audit client. We cannot verify our own work and pretend it was someone else’s. Remediation is done by the customer’s team or by a Qualified Implementation Partner; we re-test it. Where a customer asks us for advice on how to fix a finding, we give it in the report as a suggested correction, and the boundary is that we do not commit the code.
3. Reporting autonomy
Findings are reported objectively, and a customer cannot contractually redact a confirmed Sev-1 hazard from the final assurance report. The customer owns the bundle and decides who sees it — it lives in their perimeter — but the bundle itself is complete. We would rather lose an engagement than sign a report with a hole in it.
4. Third-party integrity
The verification methodology is architected with reference to IEEE 1012-2024 IV&V principles — technical, managerial and financial independence from the developer of the system under test — and the calibration benchmark used to validate our own detector is sequestered from the engineers who build it. Formal third-party IV&V accreditation is pending, and until it is granted we say “architected with reference to”, not “compliant with”.
Corollaries
No public model leaderboards
Our former public comparison leaderboards (alt.qa) are indefinitely suspended and archived. A benchmark publisher that also audits the benchmarked has a commercial conflict of interest; and a static public query set becomes an LLM training target within months. Benchmark IP is now restricted to private, parameterised challenge suites with rotating holdout seeds.
No certification language
We do not issue badges, seals or the word “certified”. We produce evidence that supports the customer’s internal go/no-go gate and residual-risk acceptance. The customer’s accountable executive retains one hundred percent ownership of the deployment decision. See why there is no badge.
No “everything AI”
Independence also means scope discipline. On 12 September 2026 we removed every claim that EvalQA “evaluates everything in AI” and every statutory-compliance certification claim from our public footprint, and narrowed the company to warehouse SQL agents. That truth audit is documented here.
We recall ourselves
When our own method or scoring is found to be wrong, the assurance recall protocol puts the affected notary records on hold within four hours, notifies the customer within 24 business hours, and re-runs the battery at no charge.
What this charter does not prohibit
Refusing all advice would make the service needlessly frustrating (audit finding #36). We will explain a finding, suggest a correction in the report, run a retest after your fix, and talk to your engineers about the invariant. We will not write your production code, and we will not be paid differently depending on what we find.
Contractual text. Clauses 1–3 appear verbatim in the order form under “Independence”. Clause 4 appears with its accreditation status. Ask for the order form at [email protected] before any conversation about price; it is one page.