# Eval Engine
The eval engine executes validated bytecode over `u32` values and buffers. Its correctness is defined by deterministic execution, bounded resources, and exact parity with the CPU interpreter.
## Invariants
Deterministic: the same bytecode, buffers, parameters, and workgroup configuration produce identical output bytes on every run.
BoundedResources: operand stack, call stack, local storage, scratch buffers, loop counters, and output buffers never overflow, underflow, OOM, or access outside declared limits.
CPUParity: every valid bytecode program produces the same result as the CPU interpreter, byte-for-byte.
## Random Test Generation
Generate random valid bytecode only. The generator tracks stack height, variable scope, buffer declarations, loop bounds, branch targets, and type signatures so generated programs pass validation. Include arithmetic, bitwise, comparison, select, load, store, bounded loop, and branch combinations.
Adversarial programs target maximum stack depth, empty blocks, nested bounded loops, dead branches, out-of-bounds loads with defined zero results, out-of-bounds stores with defined no effect, and resource limits exactly at the declared maximum.
## Verification
Run the CPU interpreter and backend engine for each generated program. Check exact output bytes, run determinism with `N >= 10` for selected programs, and verify tracked resource high-water marks stay within declared limits. A valid program returning `Err` inside limits is a conformance failure. Diagnostics include `Fix: align bytecode validation, stack accounting, and instruction semantics with the CPU interpreter`.