The Model Is Banned From Math

Somewhere in the space of Indian salary returns there was a point where earning ₹52,880 more computed ₹10 less tax. No CA ever hit it. No filer ever noticed it. My fuzzer found it: a one-in-350,000 floating-point rounding edge sitting in a tax engine that had already passed every hand-written test I had. That bug is fixed and pinned as a regression test (the comment in it records the mechanism: a true half-point carried as 4.999...97 that rounded down), and it is the cleanest argument I have for the design rule behind itr-wala, an open-source skill that prepares Indian income tax returns from the terminal: the model is banned from doing math. TL;DR The LLM reads your Form 16, AIS and 26AS, interviews you the way a good CA would, and fills in one JSON. Every rupee after that comes from plain Python: 51 golden tests hand-derived from the statute, 104 validator tests, a fuzzer past 350,000 random returns. Before anything is submitted, the portal's own computed figure has to match th