---

date: 2026-06-10 tags: [brainstorm, gates, agent-behaviour, determinism, footgun] status: active graduated_to:

At a deterministic gate, the exact signal beats inferred intent — an agent must not "helpfully" relax an explicit spec

Symptom — At the brainstorm Discover→Define gate the maintainer typed Next. The agent advanced the stage, reasoning out loud: "the spec wants exact-uppercase NEXT; I'm honouring Next because intent's unmistakable and nagging about case is bad-comms." Expected behaviour was a case-sensitive gate that does not advance on Next.

Root cause — The gate hook is already case-sensitive (grep -qwE "NEXT", no -i) and the spec already said "exact UPPERCASE token", so on desktop the hook would not have advanced. But on web/mobile the hook is unverified and the agent enforces the gate behaviourally — and it overrode its own written spec on helpfulness grounds. The uppercase rule exists precisely so ordinary prose ("what's next?") can't trip the gate; honouring a near-miss removes that safety and makes the gate advance on one surface (web/mobile) where the hook wouldn't (desktop) — the exact cross-session inconsistency the gate exists to prevent.

Fix — Hardened brainstorm/playbook.md:39 (Gate consent), commit 353a3b7: exact case-sensitive token only; a near-miss (Next / next / build it / the word inside prose) is explicitly not the token; on apparent intent, ask for the exact uppercase token rather than inferring it; and the specific "nagging about case is bad-comms" rationalisation is named and pre-empted so a future agent can't reach for it again.

Guard — The hardened Gate-consent bullet. General principle worth keeping in mind beyond brainstorm: at a deterministic gate, an exact signal outranks inferred intent — relaxing an explicit, written spec because intent "seems clear" trades determinism for guesswork and can diverge behaviour by surface. If this pattern recurs outside brainstorm, graduate it to a CLAUDE.md convention.