The claim
CLAIM-169 holds that an agent is many interests over a choice only as far as its valuations of that choice cross within the choice’s horizon. Put the interests in the sheaf picture of CLAIM-153. An interest is a section over the contexts of choice in which its valuation ranks first. The question is whether these sections glue into one self (TERM-045), in the self-governing sense.
Without an overlap they glue. Suppose the interest in the diet governs the mornings and the interest in the cake governs the evenings, and nothing in a morning bears on an evening. Then the windows are disjoint, and by CLAIM-197 the sections glue: the global section is a preference indexed by time of day. This is the “tall” example over time. A reversal of preference, on its own, is no more a division of the self than “tall” is a division of English.
Division needs an overlap. The interests conflict only on a choice that both bear on at once. The evening’s cake is that choice only if the morning’s interest is present in the evening’s context, or the evening’s in the morning’s.
Self-prediction makes the overlap. On Ainslie’s account (THEORY-067) the agent comes to see the present choice as a forecast of its later choices in the same series. The present choice is valued as the whole series’ expected value. This brings the later interest’s valuation into the present context. Bargaining, side bets and personal rules all happen on this overlap, and nowhere else.
So the population of interests that CLAIM-169 describes becomes a divided self only where the agent models itself. Where it does not, the interests take turns and do not contend.
What it answers
CLAIM-163 objects that “synchronization” of interests (TERM-050) has no coupling, no order parameter and no transitions. It names self-prediction as “a candidate” coupling. This claim says where that coupling acts: on the overlap, and only there. Interests with disjoint windows are uncoupled, as the objection says, because there is nothing for them to be coupled through. The objection’s other two parts, an order parameter and transitions, are not answered here. The objection is answered in part and stays open.
What it does not say
- That an agent without a self-model has no conflict. It may have conflict in a context both interests reach for other reasons, such as two drives cued by the same stimulus. The claim is that intertemporal division of the Ainslie kind needs self-prediction. It does not say that every conflict does.
- That the cover is settled. Which contexts of choice count, and how finely, is the open piece CLAIM-197 names, and CLAIM-153 leaves cover-dependence open as well.
- That self-prediction is conscious. It is the agent’s forecasting of its own later valuations, however that is implemented.
Note of 2026-10-12: a coupling by self-observation that unifies
CASE-074, the democracy-mode variant of Twitch Plays Pokémon, makes the claim’s coupling concrete at crowd scale. With the tally hidden the crowd has no model of its own current state. With it visible the crowd has one, and the parts are coupled through it. That fits the claim’s view that self-modelling is what couples the parts.
Two things in the case do not simply confirm the claim:
- The coupling unifies here. The claim has self-prediction create the overlap on which interests contend. In the case, the self-observed aggregate makes the voters converge, by herding. So the coupling alone does not settle whether the result is conflict or lock-in. That depends on what each part does with what it sees. QUESTION-077 asks whether the convergent outcome is a different kind of integration.
- A parallel to the precedent effect. With the tally visible, early votes weigh more than late ones. Ainslie’s lapse that counts as a precedent has the same shape. This is a parallel of form and is not derived.
The case is stipulated and its parts are voters, not interests. It bears on the claim without grounding it. No status change and no typed edge.
Drawn by the assistant and filed at the owner’s request.