Intent loss occurs when a human’s intended behavior or constraints fail to survive the path from idea to specification, implementation, and validation. A working feature is not necessarily an aligned one: the agent may have made a decision the human never meant to delegate, or completion may have been accepted without evidence that the behavior is right. section 3.2 section 3.4.1
An expired link with no specified behavior
Suppose an agent is asked to implement link expiration. If the authoring dialogue never asks what a visitor should see after a link expires, the agent must choose that behavior. This is a surfacing loss, not a coding defect that more precise tests alone can repair. By contrast, if the dialogue settles the status code and user-facing message for a malformed URL but the specification says only “return an error for invalid input,” the loss occurred during capture. section 3.2 section 3.4.1
Attribute a gap to the point where intent stopped governing the work; each point calls for a different repair:
- Surfacing: the dialogue did not elicit a needed decision. Improve its standing questions. section 3.4.1
- Capture: the decision was discussed but recorded too vaguely. Make the specification preserve it precisely. section 3.4.1
- Construction: the solution works but violates intended design boundaries. Strengthen project constraints, reusable skills, or structural checks. section 3.4.1
- Verification: a feature is called complete despite failing review or testing. Require repeatable evidence, such as tests, contract checks, or traces. section 3.4.1
Close the immediate gap by correcting the specification and regenerating the implementation. Then improve the workflow at the attributed loss point. An improved dialogue, for example, can be backtested against a past feature; the repair is confirmed only when new gaps of the same class stop appearing. Keep these changes small and routine rather than treating every gap as a reason to overhaul the process. section 3.4.2 section 3.4.1
