The quality gate should be a second adversarial model, not a human

The quality gate is the layer where most teams instinctively put a person. Blomfield argues it “probably should not be a human in all but the most extreme cases,” and that the gate is often better served by a second adversarial LLM: one that inspects the output and asks whether prompt injection is happening, or — in a bank — whether the response constitutes financial advice, which the system is not permitted to give. For engineers the easy version is a second model doing code review.

The move is what makes unattended operation possible, so it is inseparable from A human in the loop is a gating mechanism — it caps the loop at human waking hours: a human gate does not merely cost money, it reimposes waking hours on the whole loop. This is the enforcement arm of the policy layer described in Without a learning mechanism you have an AI pipeline, not an AI loop, and it is a specific instance of the broader claim in Verification is the single highest-leverage practice for agent-assisted coding that checking is where the leverage lives. The implication worth flagging is that an adversarial checker is only as good as what it is told to catch — which is where Private evals should measure business outcomes that matter — not external benchmarks becomes the binding constraint rather than model quality.