THEORY-070: How much cognitive control to exert, and on what, is decided by a cost-benefit computation (the expected value of control) in a specification system distinct from the structures that implement the control

Source

Shenhav, Botvinick & Cohen (2013), LIT-585, read in NOTE-446: the model (Eqs. 1–3), §§4–7, and the open questions.

What was actually shown

The model. Shenhav et al. (LIT-585) separate deciding on control from exerting it. A control signal has an identity, the parameter it targets, and an intensity, how far it displaces that parameter from its default. Its expected value of control is the sum over outcomes of their probability given the signal and the state, times their value, minus a cost that rises with intensity. Value is recursive, immediate reward plus discounted future EVC, so it “generalizes what is referred to as a Q-value”. The optimal signal maximises EVC. The cost is intrinsic: “like physical effort, mental effort is assumed to carry intrinsic disutility”.

The mapping. dACC monitors control-relevant states and outcomes and specifies the signal. Lateral prefrontal cortex and subcortical structures implement it. The review collects evidence for each part:

• dACC responds to conflict, errors, pain, loss and reward, and its responses fall as practice lowers the demand for control;
• dACC activity predicts post-error adjustment and the engagement of task-relevant regions, and cingulotomy abolished behavioural conflict adaptation;
• dACC activity tracked both trial difficulty and block-level incentive, and greater dACC response during a demanding task predicted later avoidance of it;
• dACC task selectivity leads lateral prefrontal cortex after a switch and lags it with repetition;
• exploration, foraging and patient intertemporal choice recruit dACC and are treated as overriding a default.

What could have come out otherwise. The account predicts dACC activity that rises with stakes as well as with difficulty, and apathy, not loss of control capacity, when the specifier is damaged. The cited incentive and lesion findings are of that kind. But the paper runs no test of its own. Its results are mappings onto existing findings, the model is stated and not fitted, and the paper calls control-relevance selectivity “not well-tested” and the dissociation of specification from regulation “not definitive”.

What this does not say

• Not what the cost is. “What exact form does the cost function assume?” is left open, as is whether it is intrinsic or a proxy for opportunity cost.
• Not how EVC is computed, or what computing it costs.
• Not that it governs emotion regulation. The paper lists “modulators of emotion” among the parameters a signal can set, and treats override of emotion-driven responses as within dACC’s remit. It does not show either.
• Not that the specifier is of a different kind from what it governs. It weighs value, as the competing tendencies do. What makes it distinct is that it is dedicated, and that it decides about control rather than about the task.

Whether it rivals THEORY-056

THEORY-056 holds that most emotion regulation is one motive state checking another, with no distinct regulating system above them, and says that “a distinct system that does the work generally would” refute it. This account posits a distinct system, though a specifying one and not a regulating one. On the record’s reading the two can both be right, so no rivals is declared:

• They divide the cases. THEORY-056’s claim is about “most” regulation, which it says proceeds effortlessly, by concurrent impulses. EVC is about effortful control, the override of a default. THEORY-056 keeps a role for reflection, and its own promote_when presupposes that deliberate regulation recruits control networks.
• The collision is conditional. They would be rivals if EVC’s specification were shown to decide emotion regulation generally, including the effortless kind. The paper asserts its reach to emotion and does not show it, and asks it as an open question.
• Both deny a regulator of a different kind. THEORY-056 denies reason governing emotion. EVC’s decider is a value-maximiser, not reason. What divides them is whether control is dedicated, not whether it is of another kind.

So the test is the one THEORY-056 already names: whether effortless regulation by a competing emotion works without recruiting the networks this account describes.

Connections

• The unity it is about. behavioral-integration: how control is allocated among tasks. It is silent on whose control it is, so it does not bear on self-governance or on THEORY-029 (ADR-024).
• THEORY-067 (Ainslie). The précis rejects “organ” models of will, and EVC’s treatment of patience as costly default override is close to one. For patience the two cannot both give the mechanism. That collision is in one application of this account, so it is stated here and not declared.
• THEORY-062 (model-based and model-free control by reliability). Different reasons for the same prediction, that concurrent demands favour habit: cost here, unreliability there.
• THEORY-061. EVC’s control signals have identities that include thresholds and attention, which are that account’s accumulation parameters. This account says who sets them; that one says what is set.
• Self-determination theory. EVC has one cost function of intensity, with no term for whether the regulation is endorsed. SDT reports that “true choice was not depleting” (LIT-560, as read in NOTE-441). Neither text addresses the other.
• Norman and Shallice (LIT-600, unread). The supervisory override of habitual responses that EVC formalises descends from that tradition; the paper as retrieved does not cite them.

Note of 2026-10-28: what holds it back

Kept Proposed by the sixth status review. NOTE-446 read the paper in full, and it “reports no new data”. Neither promotion result is in the record: no fit of the EVC computation in which control intensity tracks payoff and cost better than a model without the cost term, and no dACC/lateral-PFC dissociation of specifying from implementing beyond the cingulotomy half-result. THEORY-878’s selective regions are co-activation, not that dissociation. It now bears on QUESTION-054, whose summary already said it asserts one answer.

Note of 2026-10-28: the proposal split off

What the paper proposes, as distinct from whether it is true, is now THEORY-1110, Active, so that CLAIM-159, which places this account in the top-down cell, can stand on the proposal alone. This entry keeps the question whether the proposal holds, and its promote_when.

Note of 2026-10-30: what holds it back

Neither promotion result is held. There is no fit of the EVC computation in which control intensity tracks payoff and cost better than a model without the cost term, and no dACC/lateral-PFC dissociation of specifying from implementing beyond the cingulotomy half-result. CLAIM-159 went Active on 2026-10-29 on THEORY-1110, the proposal split off; that bears on what the paper proposes, not on whether it holds. The title asserts the thesis, not a report of it, so it keeps its condition (ADR-065 §6).