Benchmarked against DeepMind's mathematics dataset
- 10,000 nested arithmetic expressions from Google DeepMind’s mathematics_dataset (arithmetic module, test split), evaluated by all five engines: 10,000 correct answers, zero errors.
- The rational-capable engines (FLINT, decimal, MPFI) returned the exact expected fraction on every single case —
-3/17means −3/17, not an approximation of it. - The floating-point engines (C++ double, NTL) agreed within their native precision, exact wherever the answer terminates — the expected, honest divergence between exact and floating arithmetic.
- The reference answers come from DeepMind, not from us: an external check on the whole chain, from parser to printed result.