ZynoviqFrom a professional's judgment to a certified case.
One professional's opinion is only an opinion. Every case we deliver or test against has been written to an approved specification, solved independently by at least two qualified professionals, reviewed, adjudicated where they differ, tied to the law, and released as a versioned record. This page shows the eight steps.
The production workflow in eight steps
- Customer requirement
You tell us what your AI does, who uses it, which professional workflow it touches, what decisions or actions it takes, and what failure would be costly. You do not need to know which dataset or benchmark you need. Build an Evaluation Mission.
- Task specification
We write and you approve a specification before any case is authored: instructions to authors, accepted answers and alternatives, scoring rules, what counts as a critical error, the difficulty mix, evidence requirements (which authority each answer must cite), prohibited content (real personal data, for example) and acceptance criteria.
- Expert authoring
Professionals qualified in the specific capability write realistic, synthetic cases from a fixed template: facts, documents, question, expected answer fields. How we qualify professionals.
- Independent solving and review
At least two professionals solve each case without seeing each other's work. A different, higher-tier professional reviews the agreed answer, its reasoning and the rule behind it. Quality chain.
- Adjudication
Where solvers disagree, or a reviewer raises a problem, a third solver answers independently and then a senior adjudicator decides and writes the reasoning. Genuinely unsettled questions are marked ambiguous, not forced.
- Machine and manual QA
Fixed, rule-based checks run on every item: complete fields, valid formats, signs of real data, duplicates and near-duplicates across pools. The legal desk ties the answer to the governing section, rule, notification or circular with a valid-as-of date. Evidence and currency.
- Release approval
An item clears release only when it has passed every step above and the acceptance criteria in the specification. It receives a version number, a certification timestamp and a pool assignment: training or held-out, never both.
- Delivery and feedback
Datasets ship in your schema with provenance, quality and licence files; benchmarks run against your system in your test environment. Failures, customer feedback and rule changes feed back into new versions. Formats and licences · How tests run.
An authored item is not a released item. A case written by a qualified professional has cleared one step of eight. It becomes ground truth only when it has been solved independently, reviewed, adjudicated where needed, checked, tied to its authority and approved for release.
People set the answer. Software keeps it current.
- 01Qualified professionalsCapability-qualified, not just credentialed.
- 02Independent solving and reviewTwo solvers, one reviewer, no shared answers.
- 03Adjudication and legal evidenceDisagreement settled; authority cited.
- 04Certified caseVersioned, valid as of a date.
- 05Your training data or your testSeparate pools, never both.
AI never sets, changes or certifies professional ground truth. Software routes the work and runs fixed checks. The AI boundary.
Why it is built this way
| Design choice | Why |
|---|---|
| Capability-qualified professionals | A credential gets someone considered; demonstrated capability in the specific area is what makes their judgment count |
| A specification before authoring | What counts as correct, how it is marked and what you will accept are fixed before the first case is written |
| Independent solvers | One person's error or blind spot cannot become the answer |
| Disagreement is kept, not hidden | The tempting wrong answer is exactly what AI gets wrong; it is valuable evaluation data |
| Senior adjudication only when needed | Quality where it matters, without wasting expert time |
| Anchored in written rules | Agreement is not truth; the law, the standard or your written policy is the final check |
| Humans only | AI never sets, changes or certifies ground truth; people write, solve, review and certify every answer |
| Synthetic cases, test environments | No real customer data is needed, and production systems are never touched |
| Separate training and held-out pools | A benchmark only means something if the AI has never seen the answers |
The details
Where the steps lead
The same certified cases feed two products. Professional Data improves a model or agent; Benchmarks & Evaluation measures one. Training and test cases come from the same production system but never from the same items. Training and test separation. Our reports show what is wrong and why; they are not a certification or assurance of your AI. Trust.
Every failure makes the benchmark better
A failure found in an evaluation is explained with the correct answer and the rule, then kept as a regression case. Red-team scenarios that break a system join the same pool. Each round therefore leaves a better benchmark than the one before, and every system tested after that is measured against it. Customer-specific material is never reused without written permission; the benchmark compounds from synthetic cases and rights-cleared work. Benchmarks · Data rights.
How it works: questions people ask
How many professionals check each answer?+
At least two solve each case independently and a reviewer checks the result. If they disagree, a third solver and then a senior adjudicator decide, and the final answer is tied to the rule, standard or written policy behind it. Quality chain.
What happens when professionals disagree?+
The disagreement is investigated, not averaged away. A third solver answers independently, a senior adjudicator decides and writes the reasoning, and genuinely unsettled questions are marked ambiguous. The tempting wrong answer stays in the record, because it is often exactly what AI gets wrong.
Do you use AI to create or check answers?+
No. AI never sets, changes or certifies professional ground truth. People write, solve, review and certify every answer. Our software routes work and runs fixed, rule-based checks (formats, completeness, duplicates). Marking of your AI is rule-based and human today; AI-assisted marking of free text is planned, opt-in and always checked by people. The AI boundary.
Where does "correct" come from for configured business processes?+
Where the law or a standard settles the question, as in GST, answers are tied to it. For processes in systems such as SAP, Oracle or Salesforce, expected outcomes are set against your documented configuration and policy, plus professional practice, so that part of the ground truth is specific to your organisation. Expert ground truth.
How do answers stay current when the rules change?+
We track new notifications, circulars and policy changes, revise affected cases through the full quality chain, and release them as new versions with a new valid-as-of date. The old version stays in history. Evidence and currency.
Who owns the cases, and can they be reused?+
Contributors assign their rights to Zynoviq, so every licence we grant is clean. What you may do with a delivery is set in your contract or licence, and exclusivity is priced separately because it forecloses reuse. Data rights.