Six layers. No single error ships.

Cases are written by calibrated professionals in the domain, from a fixed template: facts, documents, question, expected answer fields. Every author's identity and credential are checked against their profession's register, and every author has passed calibration in the capability they write for. How credentials are checked.

Our platform runs fixed checks, with no AI involved: are all fields complete, are formats valid (GSTINs, state codes, amounts), does anything look like real data (real names, real registration numbers), is the case too similar to an existing one in either pool? Failures go straight back to the author.

At least two professionals solve each case without seeing each other's answers: the treatment, the posting, the legal basis and their confidence. Agreement on the answer and the basis moves the case on; anything else triggers a third solver.

A different, higher-tier professional checks the agreed answer and its reasoning, and flags any ambiguity or missing exception.

Where solvers disagree, or a reviewer raises a problem, a senior professional decides and writes the reasoning. Genuinely contested cases are marked ambiguous and kept out of pass/fail scoring. The full routing rule is below.

The legal desk ties the final answer to the governing section, rule, notification or circular, with a valid-as-of date: no citation, no certification. Integrity checks confirm the work was done by the people credited, within normal time patterns, with no signs of AI-written answers. Evidence and currency · Keeping human-certified answers human.

A qualification shows someone studied the law. It does not show they apply it accurately, consistently and to our standard. So before anyone touches a customer's case, they prove it.

1. Calibration test.

   A timed set of 25 cases with certified answers, in the domain and at the level the professional will work at. Pass marks are set per domain and per role.

2. Probation.

   The first 30 pieces of work are all checked by a senior reviewer. Feedback is given on every correction.

3. Continuous measurement.

   After probation, every contribution keeps being measured: agreement with other solvers, agreement with adjudicators, reviewer corrections, and planted control cases with known answers.

Calibration decides which roles a professional can take (solver, reviewer, case writer, adjudicator), which topics they are assigned (calibration is by topic, not just by domain) and which tier they are paid at. Accuracy that falls below the standard triggers re-calibration and closer review; repeated problems remove a professional from the network and their recent work is re-checked. Before every new mission, the selected team completes a short calibration round on that mission's topics, so differences in interpretation are settled at the start. How we qualify professionals · Reliability profiles.

If one professional says something, it is an opinion. If two work independently and reach the same answer for the same legal reason, that is evidence. If they differ, that is a signal worth investigating, not something to average away.

1. Two solvers answer each case independently.
2. Same answer, same legal basis: a reviewer checks it, then on to the legal desk.
3. Different answers, or the same answer for different reasons: a third solver answers independently.
4. Still split: a senior adjudicator reviews all answers, decides, and writes the reasoning.
5. Genuinely contested law, where qualified professionals would reasonably differ: the case is marked ambiguous.

Adjudicators are senior professionals with 12 or more years of practice in the domain and a strong record of agreement with final certified answers. They never settle cases they solved themselves.

Every certified record carries a confidence score from 0 to 100. It tells you how strongly the answer is supported, not just what the answer is. AI teams use it to weight training data; test reports use it to separate firm failures from borderline ones.

The exact weighting is documented in our method note and is the same for every record.

Planned, subject to founder decision. Software may generate scenario variants (amounts, counts, dates) from a certified template by fixed rules, each spot-checked by a professional. People still write and certify every answer; a variant changes the numbers around a legal point that has already been certified, never the point itself.

We track first-pass agreement, adjudication rates, reviewer corrections and errors found after certification, and use them to improve case writing and calibration. We plan to publish headline quality figures every quarter from mid-2027.