Study AO: Canonical-tag steering: the catalog read is the whole mechanism

A study by AO found that without guidance, AI models never matched exact canonical casing on a 40-tag fixture catalog across 36 drafting tasks, with 0/36 clean on all three tiers. With a canonical-tag tool present, every model consulted it unprompted and landed on-catalog topics correctly on the first attempt in 36/36 cells, including near-collision traps like llm→LLMs and frontend→Front-End. The shipped design of keeping the tool and lint together, without hard gating, is validated.

Study AO · Agent-surface steering Canonical-tag steering Replicator's CMS shipped a warn-only canonical-tag design on its agent publishing surface: a tags list tool whose description says to copy names exactly, plus advisory lint with did-you-mean suggestions. AO measured the assumption across a 40-tag fixture catalog and 36 drafting tasks. Without guidance, models NEVER matched exact canonical casing; 0/36 clean on all three tiers, so tag drift is structural, not sloppiness. With the tool present, every model consulted it unprompted in 36/36 cells and landed on-catalog topics canonical on the FIRST attempt Claude tiers 24/24, including every near-collision trap like llm→LLMs and frontend→Front-End ; zero warning rounds were needed. Warnings without the tool recovered only 7–24/36; the lint can fix slug variants but cannot supply a catalog it never showed. One honest edge: given off-catalog topics, opus returned empty tag lists 12/12 rather than invent; defensible editorial caution the registered non-empty gate mis-scored, filed as a protocol lesson. The shipped design stands: keep tool and lint together, no hard gating needed.