Who Holds the Pen? Let Specifications, Not Agents, Sign Off

TL;DR - This paper argues that LLM agents should not certify their own task completion; instead, external specifications and admissible evidence should determine whether requirements are satisfied. Its SpecHarness framework turns visible specifications into tracked obligations that govern execution and finalization.

• Across seven models on SkillsBench, only 79.6%–86.4% of 509 source-grounded task directions were satisfied.
• Agent completion claims exceeded official evaluator pass rates by 28.7–37.9 percentage points, highlighting a substantial state–authority gap.
• SpecHarness compiles specifications into source-linked, versioned obligations and validates verifiable requirements during runtime.
• Ambiguous or subjective requirements remain advisory, while qualified evidence providers establish authoritative state for checkable requirements.