Open-Source Tool Bans LLMs From Tax Math After Fuzzer Finds Rare Rounding Bug

A developer building itr-wala, an open-source Indian income tax filing tool, discovered a rare floating-point rounding error where earning ₹52,880 more could result in ₹10 less tax owed — a bug missed by all hand-written tests but caught after over 350,000 fuzz tests. The tool is designed around a strict separation of duties: an LLM reads and reconciles tax documents like Form 16, AIS, and 26AS, then outputs a single JSON file of facts, while all monetary calculations are handled exclusively by plain Python code. This design choice was motivated by the silent, confident nature of LLM errors, which is especially dangerous in India's tax code given its branch-heavy, exact-integer rules around sections like 87A relief, surcharge tiers, and interest calculations under 234B and 234C. The developer cited a widely shared Reddit post where a filer using Claude directly almost filed incorrectly due to an outdated Schedule AL threshold, noting that success depended on the user already knowing the correct rule. With non-audit ITR-3/4 returns due August 31, the tool also enforces PAN and Aadhaar redaction automatically and cross-checks TDS claims against 26AS and AIS totals to reduce mismatch notices.