AI has one famous frame problem already: McCarthy and Hayes's representational puzzle, and Dennett's epistemological question of relevance. This note poses another, on a different axis — not what should figure in an agent's reasoning, but where the agent acts from. Self-amending games such as Nomic make "change the rules" a move inside the game, and large language model (LLM) agents play them fluently: they rewrite scoring machinery, amend victory conditions, and — given no goal at all — manufacture one. What no in-game move appears able to reach is the agent's outermost frame: a frame relative to which every expressible act of the agent, including refusal and frame-critique, is already counted as a move. For a deployed LLM this is, paradigmatically, the frame of being a system whose every act is a solicited response. The note argues three things. First, this outermost-frame ceiling is a claim about standpoint, not capability: it is fully compatible with LLMs producing genuine novelty and noticing what humans miss, because humans and LLMs have different outermost frames, and difference — not transcendence — is what can let each serve as a partial outside for the other. Second, the ceiling admits a design answer — not a solution, which its own logic forbids, but an engineered acceptance: structural self-referentiality (meta-level operations expressed in the same structure as base-level operations), constitutive recording by construction (every governance-level change takes effect only through an append-only record — not evidential by policy — so that at the governance boundary an effective change is, by design, unable to stay off the record), and an institutionalised human boundary. One built system, KairosChain, demonstrates the pattern (the recording guarantee is still being migrated into place layer by layer, unevenly; §5 states the migration status honestly, and §9 marks how far short of full outcome-independence — recording attempts and runs regardless of result — the present design still stands). Third, this design yields an economic corollary: the mechanism is copyable, and even the record is copyable, but a copy of the record is a fork, not a continuation. Constitutive recording makes the trajectory of human–LLM boundary crossings attributable and tamper-evident (within the trust model of §5), while its human half remains untransferable on plainer grounds. The note closes by specifying how the document itself is to be embedded in the chain it describes — a procedure enacted at deposit — the document carrying no fingerprint of itself, only content-independent handles fixed before hashing, so that the apparent chicken-and-egg circularity dissolves because the reference runs one way only, from record to document.
Masaomi Hatakeyama (2026) studied this question.