ZynoviqInvented cases. Clean rights. Private keys. Human answers.
Two questions matter for any data you train on or test with: could it expose someone's confidential information, and does the supplier actually have the right to license it? A third matters for benchmarks: can the answers have leaked? And a fourth for all of it: did a person really write this? We answer all four by design.
The five parts of the rights chain
| Part | What it settles |
|---|---|
| Contributor agreements | Every contributor assigns their rights in their contributions to Zynoviq, commits to confidentiality, and confirms the work is original: their own, not copied, not AI-generated. Member code of conduct |
| Source-material rights | Cases are invented from professional experience, never copied from client files or third-party material. Legal sources cited are public law and official publications |
| Customer contract | Your order form or SOW states who owns what you commission, what may be reused, and the correction obligations we take on. Contracting |
| Benchmark licence | Held-out packs are licensed, never sold: named users, internal testing only, no use in AI training, deletion at the end. Benchmark licences |
| Version and provenance records | Every record names its pool, version, creation and review history (by anonymous ID) and permitted use, so the chain of rights can be checked for any item |
Creation and licensing are priced separately, and exclusivity is priced separately again, because an exclusive case can never join the shared pool that keeps prices down for everyone else. How we price.
Every case is synthetic
Professionals write cases from experience, not from files. Facts, names, amounts, registration numbers and documents are invented. The patterns are real, the kind of situation that genuinely arises, but no case describes a real taxpayer, company, employee, patient or customer.
This matters twice over. It protects the people behind real cases. And it respects the confidentiality professionals owe their clients under their professional codes: no Chartered Accountant or Company Secretary is ever asked to reveal client work.
- Rules at the source. Contributors agree never to use real client material, and are trained on what counts as identifying detail.
- Automatic checks. Every case is scanned, by fixed rules, for things that look real: valid-looking registration numbers from real registers, real company names, real addresses. Anything suspicious goes back.
- Review. Reviewers are asked to flag any case that seems recognisable.
- Fixed formats. Synthetic registration numbers follow the format but are checked against the public GST registration search before use, so none matches a real taxpayer.
Only one service touches real data: the expert review queue, where qualified professionals decide your AI's flagged live cases, and only under a data processing agreement. That data is never used for anything else, never becomes a test case and never becomes training data. Expert review queue · Data processing.
What every delivery tells you
Each dataset includes a rights and provenance record: that every record is synthetic, who (by anonymous ID) created and checked it, when, which pool it belongs to, and under which licence you may use it. Formats, delivery and licences.
Keeping benchmark answers private
If an AI has seen the questions and answers before, it can pass a test without being any good, like a student who saw the paper in advance. Protecting the answer key protects the value of every test, including yours next quarter. Four layers do it.
1. How we deliver.
In most tests the answer key never leaves our systems: we send questions and mark the answers, or you send us your AI's answers to mark. How tests run.
2. How packs are built.
- A large hidden pool, rotated. Each test draws a subset; the next draws a different one.
- A different copy for each customer. The same legal points, with different amounts, dates, names and states, so any leaked copy shows its source.
- Trap facts. A few invented details that exist nowhere else. If an AI repeats them, it has been trained on our pack.
- Retired once shown. Cases explained to you in a report are replaced in your next test.
- Limited attempts. A set number of full runs per period, so answers cannot be worked out by trial and error.
3. Contracts.
Benchmark packs are licensed, never sold: internal testing only, no sharing, no use in AI training, including through outside AI vendors, named users, deletion at the end, and agreed remedies for breach. Licences.
4. Our own controls.
The answer key lives in a separate, encrypted store that only our marking service can read. Contributors see only their own small set of cases, sign confidentiality agreements, and every access to the key is logged. Training and test separation · Trust and security.
What this means for you
| You get | Because |
|---|---|
| A test your AI cannot have memorised | The pool is private and rotated |
| Fair comparison between quarters | Fresh cases each run, same standard |
| Honest regression testing | Your own releases cannot learn the test |
| Confidence in shared results | Packs are traceable if they leak |
| Your reports, every failed case with its correct answer, and any custom cases built only for you | These are yours to keep under your contract |
Human-certified means a human actually did it
The fastest way to ruin a human-certified dataset is for contributors to paste answers from an AI chatbot. It happens across the industry. It would make our answers no better than the AI we are testing, so we design against it at every step.
Tasks that reward judgment, not recall. Cases are written around realistic facts, exceptions and documents, where a generic AI answer is easy to spot as generic.
Time windows. Each task has a realistic time window. Answers returned implausibly fast, or after long idle gaps followed by a large paste, are flagged.
Writing inside the workbench. Answers are written in our workbench, which records normal working patterns (typing, editing, time spent) without recording anything outside it.
Hidden control cases. Every batch contains cases with known answers, including ones built around mistakes AI systems typically make. A contributor who falls into those traps repeatedly is investigated.
Same point, different words. The same legal point appears in different forms. Inconsistent answers from one person are a warning sign.
Follow-up questions. For high-value work, reviewers may ask a contributor to explain their reasoning, briefly, in writing or by a short call.
Independent solvers. Because two or more people solve each case independently, one person's pasted answer rarely survives comparison and review.
What happens when it is found
| Situation | Action |
|---|---|
| A single flagged answer | Reviewed; the contributor is asked to explain |
| A confirmed breach | Work rejected and unpaid; contributor removed |
| A pattern | All that contributor's recent work re-checked and replaced if needed |
Contributors agree to these rules when they join. We do not rely on software that claims to detect AI writing; such tools are unreliable. Task design, work patterns, control cases and human review are explainable and fair to contributors. When we say an answer is human-certified, that claim has to be true, or the test tells you nothing and the training data teaches your AI its own mistakes. The AI boundary.
Data rights: questions people ask
Are your cases based on real client files?+
No. Professionals write cases from experience, and every name, number and document is invented. Contributor rules, automatic scans and reviewer checks catch anything that looks real.
Who owns what contributors create?+
Zynoviq. Every contributor assigns their rights to us, which lets us grant you a clean licence or, where your contract says so, ownership of what you commission.
Can we own the data you make for us exclusively?+
Yes, by agreement. Exclusivity is priced separately because it forecloses reuse: a case we can never reuse costs more to make than one that joins the shared pool. Licence, exclusivity for a period, and outright ownership are the three options. Formats and licences.
Why do you not hand over the full benchmark answer key?+
Because an exam only works if the answers stay private. Answer keys that leak end up in AI training data, and then tests stop measuring real ability, for you and everyone else. Full-pack licences exist under strict terms as a separate, higher-priced option.
Do you use AI-detection software on contributors?+
No. Such tools are unreliable. We rely on task design, work patterns inside our workbench, hidden control cases and human review. We do not monitor contributors outside the workbench.