ZynoviqFix the gaps the test found.
An evaluation tells you your AI is wrong on, say, credit reversals or place of supply. Your team now needs material to fix it — and the reason the AI is wrong in the first place is that this material does not exist in a usable form. We build it.
Two kinds of fix, depending on your AI
| If your team can… | We supply | Example |
|---|---|---|
| Retrain or fine-tune the model (AI labs, agent builders) | Training data — worked reasoning, preference comparisons, hard cases and counter-examples | Reverse-charge cases with step-by-step expert reasoning Illustrative |
| Not retrain the AI (most SAP and packaged-AI users) | Reference packs — certified rules and worked examples the agent looks up while it works | A place-of-supply rulebook with worked examples Illustrative |
What goes into it
- Worked reasoning — how a professional gets from the facts to the answer, citing the rule.
- Hard cases and counter-examples — the traps that fooled your AI, and their near neighbours.
- Comparisons — two answers ranked by a qualified professional, with the reason one is better.
- The common wrong answer — recorded next to the right one, so your AI learns what to avoid.
How it is made
By the same qualified professionals, through the same quality chain, as our test cases: independent solvers, review, adjudication, and a legal evidence check with a "valid as of" date. The quality chain.
One firm rule: improvement data is always created fresh. We never turn test questions into training material — that would make the next test meaningless.
Data goes to builders. Benchmarks go to buyers.
We never sell improvement data for a system whose benchmark score we publish. A published score has to mean the same thing for every system it compares, and it cannot if we have also supplied the material that lifted one of them. Improvement data goes to the teams that build and fix AI; published benchmark results are for the people who buy it. Training and test separation.
How it is delivered
As structured files in the format your team specifies, in agreed batches, each with a provenance file, a quality report and a licence file. Large deliveries go straight into your own cloud storage. Formats, delivery and licences.
Who does the fixing
You do. We do not train, tune, host or deploy anything. The data is yours to use on your own infrastructure, under the licence agreed.
Measuring whether it worked — honestly
After your team makes changes, we can re-test against the same topics and report what moved. We call that a before-and-after measurement, because that is what it is. We made the data, so we will not present the result as proof from a party with no stake in it. If you need that kind of confirmation, get it from an organisation with no stake in the result. Regression.
Improvement data for AI — FAQ
What is improvement data?+
Expert-built material aimed at exactly what your AI got wrong: worked reasoning, preference comparisons, hard cases or reference packs, each certified and tied to the law.
What if we can't retrain our AI, such as an SAP-delivered agent?+
Use reference packs — certified rules and examples the agent can look up — and route risky case types to people. Reference packs.
Do you turn test questions into training data?+
Never. Improvement data is always created fresh, so your next test still measures real ability. Training and test separation.
Who applies the fixes?+
Your team does. We don't train, host or deploy any AI.
How do we know the fix worked?+
We can re-test the same topics and report what moved, described honestly as a before-and-after measurement.