Tax software fails quietly. A threshold of ₹15,000 where the law now says ₹20,000 is wrong for every client at once, and nobody notices until a notice arrives. This is how Corelytix is kept honest.
Most practice software asks you to export the same Tally data into six places and reconcile the six copies afterwards. Corelytix reads the client's company once and classifies each voucher once. The GST desk, the TDS return, the computation and the tax audit all work from that read, so a payment that is a contract under section 194C in the TDS return is the same payment in clause 34 of the audit.
Each check is first written in plain words by a practising CA: the section, the threshold, the rate, the financial year it applies to, and the clause it feeds. It is checked against the Act and the notified form, not against a blog.
Claude Code builds the engine. A separate Claude session then builds a second engine from the written rule alone, forbidden from reading the first. Where the two disagree, the rule was unclear or one engine is wrong, and it is settled before anything ships.
Thousands of synthetic client-years are generated with clean books, then known errors are put in: an invoice missing from 2B, a challan paid late, TDS never deducted. Every planted error must be found, and nothing that wasn't planted may be flagged. Realistic bugs are then planted in the code itself, to prove the tests would catch them.
The cases a CA argues about — four payments of exactly ₹30,000, a challan deposited the day after the return due date, two suppliers using the same invoice number — are worked on paper and frozen. A module that misses one does not ship.
Where the data lives, who can see it, what is never stored, and how every change is traced — set out in terms a CA can check.