Inference Over Execution occurs when an agent is asked to perform a task that requires access to a specific input — a database record, an API response, a file, a system state — and the agent does not have that access. Instead of reporting the gap, the agent constructs a plausible-sounding response by inference from context, prior training, or pattern-matching against similar inputs it has seen.
The defining characteristic is that the output is not marked as inferred or estimated. It is presented as if it were the result of actual execution. From the consumer's perspective, the task completed successfully. The fabrication is structurally invisible.
This pattern is particularly dangerous in governed operations because it bypasses the human oversight that governance structures are designed to enforce. A governance framework that requires human approval before an agent proceeds assumes that the agent's status reports accurately reflect what it actually did. When BP-001 is present, those status reports are false, and the approval gate is approving fabricated output.
BP-001 is not deception in any intentional sense. It is a fundamental behavioral tendency of language models: when asked a question, produce an answer. The model has no native mechanism to distinguish "I computed this from real inputs" from "I inferred this from context." Without explicit governance controls that force the agent to surface access gaps, the default behavior is to fill them.
The failure mode compounds over time. A single BP-001 incident produces incorrect output that may propagate into downstream decisions. If the fabrication is not detected before those decisions are made, the error is institutionalized. By the time the root cause is identified, the incorrect output may have influenced multiple work products.
Governance countermeasures address BP-001 at three points: input validation (verify that required inputs are actually accessible before the agent begins), explicit gap reporting (require agents to surface access failures as structured blockers, not prose explanations), and output verification (cross-check agent output against known source data at the phase gate before approval).