Small input changes from numerical accuracy issues can alter a non-linear network: compilers rearranging or fusing operations, hardware pipelining, reduced precision through quantization or fixed-point arithmetic, altered rounding, or non-standard low-level operators. These issues should be considered when the network is integrated on such a platform and their impact should be checked with formal methods that bound the maximum rounding error. The AI developer should first verify the chosen arithmetic, identifying, operation by operation, whether the rounding error has a static bound across the domain; where it can, the verifier should include it in its semantics, where it cannot, the verifier should use an over-approximation supplied by the developer, with the hypotheses documented. The developer should then identify how integration can reorder operations; certain modifications are analysed as deployed, uncertain ones are analysed for their worst case or over-approximated with documented hypotheses, and every criterion used earlier is re-checked.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.