Definition
Plain language
Rewriting different-looking inputs into one standard form so that harmless variations stop mattering.
As stated in the literature
Normalizing surface variation (spacing, casing, punctuation, phrasing) into a canonical representation before a model or filter sees it, so that semantically equivalent inputs map to identical strings.
Also called: canonicalize, canonicalizing
Why it matters: It stops attackers from slipping past filters with cosmetic tricks like odd spacing or capitalization, and stops harmless formatting differences from changing a system's answer.
For example, "H E L L O", "hello!!!", and "HeLLo" all get rewritten as "hello" before anything else looks at them.
Heard on the show
“If it's close but malformed, canonicalize — fix it.”Episode 071 — When the Model Is Fine and the Plumbing Is Broken: Fixing Agents at the Interface