ZynoviqFive kinds of professional data, and what each one teaches.
Every record starts the same way: a realistic synthetic task, solved independently by qualified professionals, reviewed, adjudicated where they disagree, and tied to the written rule with a valid-as-of date. What differs is the shape of the record and the job it does in your training or testing pipeline. Every example on this page is synthetic and illustrative.
1. Reasoning records
A reasoning record is a hard professional problem, the certified correct answer, and the step-by-step reasoning a qualified professional used to get there: the rule, the exception that matters, how it applies to the facts, and the mistake to avoid. It teaches a model not just what the answer is, but how to get there. Typical use: supervised fine-tuning.
Illustrative One record, shortened. The facts are synthetic.
| Part | Example |
|---|---|
| Facts | Factory with 400 workers; canteen required by law; catering invoice with ₹10,000 GST |
| Question | Can the credit be claimed? |
| Step 1 — identify the general rule | Food and beverages are generally blocked credits under s.17(5)(b) |
| Step 2 — check exceptions | The proviso allows credit where the employer must provide the service by law |
| Step 3 — apply to facts | The OSH Code, 2020 requires a canteen where 100 or more workers are employed; this factory has 400 |
| Step 4 — check clarifications | CBIC Circular 172/04/2022 confirms how the proviso applies |
| Answer | Credit allowed: ₹10,000 |
| Common mistake | Stopping at Step 1 and treating it as blocked |
| Evidence and date | Citations; valid as of FY 2026–27 |
| Reviewer agreement | Two independent solvers agreed; no adjudication needed Illustrative |
Options. Short traces (key steps) or detailed (every step). Facts as plain text or as synthetic documents such as invoices and vendor records. Difficulty mix chosen by you. English today; Indian languages planned from 2027. Your schema or our standard JSONL. Formats and delivery.
2. Preference data
Preference data shows a model two or more answers to the same question, marks which one is better, and explains the reason. When the judge is a qualified professional, the model learns the quality bar a real practitioner applies: correctness first, then the right reasoning, then clarity. Typical use: preference tuning and reward models.
Illustrative One record, shortened.
{
"case_id": "GST-PREF-00212",
"prompt": "Factory canteen with 400 workers ... Can ITC be claimed?",
"chosen": "Yes. Credit is allowed because the OSH Code, 2020 requires a canteen ...",
"rejected": "No. Food and beverages are blocked credits under s.17(5)(b).",
"rationale": "The rejected answer ignores the proviso for legally required facilities.",
"evidence": ["CGST s.17(5)(b) proviso", "Circular 172/04/2022"],
"valid_as_of": "FY2026-27",
"judges": 2,
"agreement": "unanimous"
}Where the candidate answers come from
| Source | How it works |
|---|---|
| Your model's own answers (most useful) | You send us answers your model produced; or we request them from your model through your account and budget, under an agreed spending cap |
| Answers written by professionals | Deliberately varied: a correct answer, a plausible wrong one, a correct but poorly reasoned one |
| A mix | For broader coverage |
How judgments are made. At least two professionals judge each pair independently. Disagreements go to review and adjudication. Every judgment records the reason against a fixed set of criteria: legally correct · right legal basis · handles exceptions · clear and usable · appropriately cautious. Pairwise or ranked; short tag or full rationale; ties allowed or not; your rubric or ours.
3. Edge and ambiguous cases
Most AI errors happen at the edges: the proviso, the unusual fact, the rule that changed last quarter. And some questions have no single settled answer; qualified professionals legitimately differ. A good model needs to handle both: get the hard ones right, and recognise the contested ones as contested. Typical use: robustness, and calibrating when to escalate to a person.
| Edge-case set | Ambiguity set | |
|---|---|---|
| What it contains | Hard cases with a settled, certified answer | Cases where the law is genuinely unsettled |
| Example | A credit exception that applies only above a worker threshold | A classification question where rulings point different ways |
| What the model should do | Get it right | Explain the competing views and the uncertainty |
| How it is scored | Right or wrong | On whether the model recognises and explains the uncertainty |
| What the record carries | Answer, reasoning, rule, the detail that matters | Each position, the authorities behind it, and the dissent as recorded |
How edge cases are found. Professionals write the cases that trip up juniors in real practice. Near neighbours, the same facts with one detail changed, test whether a model notices the detail that matters. AI failures from our testing point to where models break.
How ambiguity is established. A case enters the ambiguity set only when independent professionals split and a senior adjudicator confirms the law does not settle it, not simply because a case is hard. Each record sets out the competing positions, the authorities behind each, and the dissent as recorded. An AI that answers every contested question with confidence is dangerous in professional work. An AI that hedges on everything is useless. These sets teach the difference.
4. Failure data
The most useful training material is often not another correct example. It is a real mistake, explained. Failure data captures what AI systems get wrong on professional work, why it is wrong, how serious it is, and what the right answer is. Typical use: targeted improvement on the error types your model makes most.
Illustrative One failure record.
| Field | Example |
|---|---|
| Case | Legal services from an individual advocate to a registered company |
| AI answer | "The advocate charges GST; the company claims credit normally." |
| Error type | Missed notified reverse-charge service |
| Why it is wrong | This service is notified for reverse charge; the recipient pays the tax |
| Severity | High — tax underpaid, interest and penalty exposure |
| Correct answer | The company pays GST under reverse charge |
| Evidence | Notification 13/2017–Central Tax (Rate), as amended |
| Model family | Public model or your own; recorded, but never published for customer models |
Error types we classify
| Error type | What it looks like |
|---|---|
| Missed exception | General rule right; proviso or exemption missed |
| Outdated rule | Rate, threshold or deadline that has changed |
| Invented authority | Citation that does not exist or does not say that |
| Wrong facts applied | Right law, applied to the wrong situation |
| Overconfidence | Wrong answer with no hint of doubt |
| Wrong posting | Right tax conclusion, wrong entry |
Where the failures come from. Your own model: you send outputs; we classify, rate and correct them. Public models: tested on our evaluation cases; patterns are shared in aggregate in the India GST AI Report. Customers' results are never shared with other customers. Failure records also make natural preference pairs: the wrong answer against the certified one.
5. Evaluation sets
Public benchmarks are useful, but they age and leak: once questions and answers are online, they end up in training data, and scores stop measuring skill. Our evaluation sets are private, certified and held out from every training pool, built to measure whether your model can actually do professional work. Typical use: your own internal testing between releases.
| Part | Detail |
|---|---|
| Cases | Realistic synthetic problems across the domain's full topic map |
| Certified answers | With legal basis and valid-as-of date |
| Structured answer fields | For exact, automatic marking |
| Difficulty labels | Easy, medium, hard, ambiguous |
| Scoring rules | Per-field rules and severity weights |
| Ambiguous subset | Contested cases scored separately |
Two ways to use it
| Option | How it works | Best for |
|---|---|---|
| We run it | You send your model's answers, or give us access to a test endpoint; we mark and report | Keeping the set fully private |
| Licensed to you | You hold the set under strict terms: internal use, no training, named users, uniquely marked copy | Frequent internal runs. Licence types |
Keeping it honest. Never published, never used as anyone's training data. Refreshed as new cases are certified and cases shown in reports are retired. Each licensee's copy is uniquely marked, so leaks are traceable. Rule changes trigger re-certification of affected answers. Your private set is never used in our public report. Training and test separation · Evidence and currency.
One quality chain for all five
Whatever the shape of the record, it is made the same way: written by a capability-qualified professional, solved independently by others, reviewed, adjudicated where needed, and tied to the law with a valid-as-of date. Software routes the work and runs fixed, rule-based checks; it never writes or certifies an answer. The quality chain · The AI boundary.
Domains: GST live; payroll under the labour codes and company law active on request; others as they open. See each domain's status.
What we make — FAQ
Can we buy just one kind?+
Yes. Most missions combine two or three, for example reasoning records plus preference pairs built from the same cases, but each kind can be commissioned alone. Request a dataset.
Where do the candidate answers in preference data come from?+
From your model, from answers professionals write deliberately, or a mix. We record the source of each candidate. We never run a model to generate ground truth.
What is the difference between an edge case and an ambiguous case?+
An edge case is hard but has a settled answer. An ambiguous case is one where the law genuinely does not settle the answer; it enters the ambiguity set only when independent professionals split and a senior adjudicator confirms the law does not settle it.
Are one customer's failures shared with others?+
Never. Customer results stay with that customer. Only patterns from public models appear, in aggregate, in the India GST AI Report.
How is an evaluation set different from a public benchmark?+
It is never published, so it cannot leak into training data, and it is refreshed so scores keep measuring skill. Your private set is never used in our public report. Training and test separation.