humphreys-law ADG-0013

Humphrey's law

“Users do not know what they want until they see working software.”

Also: Ziv's law · Requirements are discovered, not given

Origin

Watts Humphrey, Software Engineering Institute, 1990s. Ziv's law (Hadar Ziv, 1996) is the stronger form: specifications will never be fully understood. About requirements in human software projects, where the mechanism is that the act of seeing a system reveals what was wanted.

Why it should, or should not, apply to models

An agent receives a task description and, unlike a human developer, does not push back or ask for the missing half. If the law transfers, the under-specification that people resolve by iterating with users becomes silent failure in an agent: it builds confidently to the wrong spec. The prediction is that effort spent turning the raw request into an explicit, testable specification before building pays off more for an agent than the same effort would for a person, because the agent will not discover the gaps on its own.

Standing: Untested

Standing counts claims a person has reviewed. Claims reviewed by AI alone are listed and marked, and do not move it.

Where it holds

Claims showing the heuristic applying to models.

On repository-level issue-fixing tasks (SWE-bench Lite/Verified/Pro), spending effort upstream — expanding the issue text into a structured requirement spec from retrieved repo context, then testing and refining that spec via generated code and tests — raises resolve rate over agents that consume the raw issue description, and raises the share of patches that even apply cleanly by a larger margin than it raises correctness.
Generating and editing working codeReviewed by AI
Expanding the raw issue into a structured, tested specification before building raised resolve rate over agents that worked from the issue as given — the requirements had to be discovered first.

Add a case where it holds or breaks

A break is worth more than a hold. If you have seen this fail for a model, with the setup written down, that is the most useful thing you can add here.

Notes

Wanted: a measurement of how often an agent asks a clarifying question when the task is under-specified, against how often a person would. The abstention-and-self-knowledge capability is where such a claim would sit.