hyrums-law ADG-0014Hyrum's law
“With enough users of an interface, every observable behaviour of the system will be depended on by somebody, whatever the contract promises.”
Origin
Hyrum Wright, Google, circa 2012, from maintaining widely used C++ libraries. About API consumers depending on incidental behaviour, so that any change breaks someone.
Why it should, or should not, apply to models
A model is an interface with millions of users and no contract at all. Every prompt in production depends on incidental behaviour — the exact way a model formats a list, how it treats an ambiguous instruction, which of two readings it prefers — and a model upgrade changes those behaviours without notice. So the law should hold with unusual force: there is no documented contract to hold the provider to, and the dependence is on statistical tendencies rather than discrete behaviours, which makes it harder to even enumerate what broke.
Standing: Untested
Standing counts claims a person has reviewed. Claims reviewed by AI alone are listed and marked, and do not move it.
No evidence filed either way.
Searched 2026-09-12: Nothing in the catalog measures how many production prompts break across a model version change, or characterises which incidental behaviours get depended on. The claims nearest by meaning are about instruction-following failures within one model, not drift between models. Nearest misses: verifiable-instructions-such-length-format-constraints-strong-models.
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 study that runs a fixed prompt suite across successive versions of one model family and counts silent behaviour changes. Vendors' own migration guides are the nearest published evidence and would count as vendor-doc sources.