규칙 11.7

a host call computes what the GPU would, so it rounds as f32 does, which the default f64 algebra oracle is blind to by construction. The generated code computes each operation as the interpreter’s runtime helper does: a vector operation of a known width, on operands nothing can leave missing, is written one component at a time, each the scalar operation the helper applies to it (dot, length, distance, normalize and cross summed in the order their entries sum), and every other operation, one on an operand that may be missing (a helper’s discard, an element read past an array’s end) among them, calls the helper itself; src/core/cpu-codegen.test.ts holds the two to each other bit for bit in both precisions, so the tier is the oracle and not a fourth implementation (Rule 11.1). Building it with new Function at run time would need unsafe-eval, which a strict content security policy forbids, and would ship the IR and the generator; as module code it is minified with the application, and the application ships the op library only.

change 0009 in changes/ (“The CPU tier’s precision”); src/core/oracle.ts (the 'f32' mode, “a correctly-rounding f32 machine over the same IR”); Rule 11.1.