You can't trust assembly an AI wrote. Here's the 3-command gate.

When writing about AI-generated assembly code, it's crucial to remain vigilant and verify every technical claim. AI assistants may invent non-existent mnemonics, flip operand orders, drop immediate values, or misread bytes. Real incidents often share a signature: the code appears plausible yet never errors. To prevent these errors, a mechanical gate can be implemented.

This gate involves assembling the code, disassembling it, and comparing it to the original. If the disassembly doesn't match the original instruction, mnemonic, operands, or size, then the instruction was not written correctly. Two examples of this gate catching errors include a non-existent mnemonic (movqad) and a truncated immediate value (38 dropped by the parser).

Additionally, switching between AT&T and Intel syntax can cause semantic changes in the code. To avoid these issues, it's essential to assemble the code, explicitly set the syntax dialect, and never mix AT&T and Intel syntax. Verifying the raw bytes for unusual instructions and consulting manuals like Intel SDM, ARM ARM, or RISC-V ISA is also crucial.

If the toolchain is unavailable, the code should be marked as UNVERIFIED and include the command that would check it. This discipline should be applied everywhere, including verifying real parallelism (thread counts and wall time) and ensuring APIs actually exist (using cargo search, for instance). The verification process should be self-validating, with a harness that cannot fail serving as evidence.

The repository "low-level-skills-trothbyte" provides 124 verified skills for various programming languages and assembly, with all claims source-traced and continuously updated.