Every optimization has a shadow regression — guard commands make the shadow visible

In any iterative optimization loop, improving one metric risks silently degrading another. Karpathy’s Autoresearch verifies val_bpb after every experiment, but the original design has no mechanism to check whether VRAM usage exploded or code complexity ballooned. The guard command pattern — a separate invariant check that runs alongside the primary metric — makes these shadow regressions visible before they compound.

This is Invert, always invert — many problems are best solved backward applied to optimization: instead of only asking “did the metric improve?”, also ask “what could have gotten worse?” The guard is the structural answer to that inversion. It connects to Verification is the single highest-leverage practice for agent-assisted coding because verification without guards is single-dimensional — you’re measuring what you optimized while ignoring what you didn’t. And it reinforces why A mediocre agent inside a strong harness outperforms a stronger agent inside a messy one: the agent doesn’t need to understand multi-objective tradeoffs if the harness enforces them mechanically. At a strategic level, Amplification widens the judgment gap — AI magnifies clear thinking into compounding advantage and confused thinking into accelerating waste shows this pattern applies beyond metrics — AI amplifies whatever thinking quality you point it at, so optimizing speed without guarding judgment quality creates a shadow regression in strategic direction.