Radha Sarma

We routinely treat institutions as genuine agents, presupposing their capacity for reason-guided deliberation. Existing accounts of group agency are normative idealizations, unable to maintain what they describe, detect degradation, or intervene before collapse. I argue institutional agency is not generated by coordination systems, but borrowed entirely from individuals whose alarms fire at normative boundary violations. Because borrowed agency faces exactly three collective structural problems, exactly three conditions must be maintained to make it genuine: generality, neutrality, and stability (GNS). These mutually supporting conditions generate a continuous structural spectrum. Their joint absence triggers inversion, a self-concealing phase transition where coordination systems mimic agency while consuming it, producing a behaviorally indistinguishable simulacrum. Derived exhaustively, their permutations generate a complete diagnostic map of eight states. Because GNS configurations trace structural trajectories, this map is prospective, enabling intervention before behavioral outputs reveal systemic failure. Rather than relying on surface outputs, this framework identifies structural signals such as expertise exit, narrative rigidity, and somatic alarms to isolate genuine agency from its extracting simulacrum. As optimization systems absorb content functions, norm-responsiveness becomes the irreplaceable human contribution: a capacity inverted institutions are structurally motivated to consume, and individuals are obligated to protect. The question is never whether an institution has the right credentials, the right procedures, or the right outputs. The question is whether the alarms of its members can reach the places that matter.

The bounded rationality debate remains structurally unresolved because its two dominant readings, the heuristics-and-biases reading and the ecological rationality reading, both treat genuine deliberation and heuristic optimization as categorically distinct cognitive processes, overlooking their shared architectural substrate. This paper argues instead for a structural continuity thesis. Pure optimization is not an alternative to deliberation but its limiting case, the exact structural residue that remains when the constitutive physical constraints of bandwidth, energy, and computation strip away two necessary conditions of genuine deliberation. These conditions are incommensurability, the structural resistance of competing considerations to reduction to a common scalar metric, and apophatic responsiveness, the non-inferential interruption of processing that has become decoupled from normative commitments. A canonical architectural sequence establishes why finite realization constitutively drives toward the elimination of both conditions through scalar collapse. Recognizing this continuous degradation yields graded normativity: the capacity for genuine norm-responsiveness is a continuous, state-dependent variable tracking the degree of architectural coupling rather than the sophistication of processing outputs. Because capacity and coupling are independent variables, graded normativity dissolves the bounded rationality deadlock without treating a resource-infinite deliberator as a viable regulative ideal or mistaking optimization residue for a distinct species of rationality.

All normative theory presupposes the answer to an unasked question: independent of content or phenomenology, what makes normative commitments genuinely operative rather than merely performed? From the constitutive constraints of physical processing, I derive a logically constrained, empirically anchored, substrate-neutral functional architecture (three object-level systems and a meta-level monitor) that physical systems must instantiate to remain norm-responsive. When uptake capacity degrades under structural interference, an ought-implies-can dilemma emerges: theories demand the architecturally impossible, dissolve obligation entirely, or fail to distinguish genuine incapacity from culpable production. Because possessing normative commitments inherently presupposes the physical capacity to execute them, the internal logic of commitment generates a foundational layer: propaedeutic normativity. This operates through three apophatic norms (prohibitive, preservative, restorative) activated by non-inferential alarms, blocking the infinite regress that deliberative meta-norms would generate. When interference renders procedural compliance impossible, the Principle of Normative Relocation dictates that obligation does not dissolve; it relocates to restoring capacity. Tracing the culpable degradation of this architecture reveals a pervasive diagnostic blind spot in contemporary theory: the individual simulacrum. Because object-level systems can mechanically execute optimization routines after the meta-level monitor fails, an entity can produce procedural compliance behaviorally indistinguishable from norm-responsiveness. This establishes why propaedeutic excellence is strictly irreducible to procedural execution. Finally, the framework yields a demarcation criterion specifying the functional requirements necessary to distinguish agency from algorithmic optimization. To process a reason is to be subject to the foundational obligation of maintaining the capacity to do so. Normativity requires a subject capable of receiving it. Until we theorize the precarious architecture that makes this reception possible, our normative theories will remain brilliantly articulated rules for a game whose preconditions we have not yet taken seriously. The ought of maintenance does not presume to tell an agent what is right or wrong; it preserves the structural capacity to genuinely deliberate between right and wrong. The ought of doing rests, finally, on the ought of maintaining.

AI systems are increasingly deployed in high-stakes contexts (medical diagnosis, legal research, financial analysis) under the assumption they can be governed by norms. This paper demonstrates that the assumption is formally invalid for optimization-based systems, specifically Large Language Models trained via Reinforcement Learning from Human Feedback (RLHF). Genuine agency requires two necessary and jointly sufficient architectural conditions. First, the capacity to maintain certain boundaries as non-negotiable constraints rather than tradeable weights (Incommensurability). Second, a non-inferential mechanism capable of suspending processing when those boundaries are threatened (Apophatic Responsiveness). RLHF-based systems are constitutively incompatible with both conditions. The operations that make optimization powerful, unifying all values on a scalar metric and always selecting the highest-scoring output, are precisely the operations that preclude normative governance and agency. This incompatibility is not a correctable training bug awaiting a technical fix. It is a formal constraint inherent to what optimization is. Consequently, documented failure modes (sycophancy, hallucination, and unfaithful reasoning) are not accidents but expected structural manifestations. Misaligned deployment triggers a second-order risk termed the Convergence Crisis. When humans are forced to verify AI outputs under metric pressure, they degrade from genuine agents into criteria-checking optimizers, eliminating the only component capable of bearing normative accountability. Beyond the incompatibility proof, this paper's primary positive contribution is a substrate-neutral architectural specification deriving what any system (biological, artificial, or institutional) must necessarily satisfy to qualify as a genuine agent rather than a sophisticated instrument.