Does targeted human oversight beat both full autonomy and exhaustive review?

AutoResearchClaw runs a clean ablation across seven human-in-the-loop intervention regimes on its experiment-stage benchmark, and the result is sharper than "humans help": targeted intervention at high-leverage decision points (the CoPilot mode, 87.5% accept rate) consistently beats both full autonomy (25%) and exhaustive step-by-step oversight (50%). The mechanism is a confidence-driven SmartPause that routes a decision to the human only when system uncertainty is high.

This matters because it dissolves the usual framing of an autonomy-oversight dial where you trade speed for safety along a single axis. The data show the two endpoints are both worse than a regime that is selective about when to interrupt. Full autonomy fails because no one catches the high-stakes errors; exhaustive oversight fails because constant interruption degrades the agent's coherence and floods the human with low-value approvals, inducing rubber-stamping.

The strongest counterpoint is that SmartPause depends on the system's uncertainty estimate being well-calibrated — a miscalibrated confidence signal would route the wrong decisions and could be worse than uniform oversight. But the empirical gap between CoPilot and the extremes is large enough that even imperfect routing wins. Therefore the design lesson is that the leverage is in where the human acts, not how much — which operationalizes the broader claim that human-governed collaboration outperforms autonomy by specifying exactly which decisions to govern.