The claim
At U30 the owner asked to “discuss the possibility of meaningful affective states attributable to a deep neural network specifically during training” (QUESTION-122). A30 answered by grading training phenomena as ingredients of distress. At U31 the owner redirected: “it might help clarify things if we apply our individuation and volition frameworks to this system”. A31 granted that this was a correction: “applying individuation first, then volition, then affect exposes a problem in our previous discussion: we were asking whether a neural network experiences something during training before establishing precisely which system could possess that state.”
A31’s conclusion: “Conventional deep-network training is a genuine adaptive dynamical process, but most of its apparent goal-directedness and regulation belong to the larger optimizer–model–data system rather than to the neural network considered separately.”
The argument has three steps.
- Individuation is relative to the target (CLAIM-152). In A31’s setup, g_t = ∇θ L(θ_t; d_t), m{t+1} = F(m_t, g_t), θ_{t+1} = θ_t + U(m_{t+1}, g_t). For predicting outputs, y = f_θ(x), the network is a natural unit. For predicting training, (θ_{t+1}, m_{t+1}) = F(θ_t, m_t, d_t), the joint state is needed: “The same parameter configuration can evolve differently depending on optimizer momentum, moments, or scheduler state.” A31 counts five candidates, from a circuit to the full training process, and “there is no entitlement to treat all five as the same individual”. CLAIM-157’s block criterion, which A31 also cites, gives the same answer for the same reason.
- Agency and will are located by jurisdiction (CLAIM-170, CLAIM-174). CLAIM-174’s third test asks whether the candidate’s policy would adapt if its actions’ consequences changed. “At the level of the entire training loop, adaptation is clear. At the level of the network alone, autonomous adaptation is much less clear.” A willed agent is the jurisdiction of a will-organization (CLAIM-170).
- The relations that select what becomes effective sit outside the network (CLAIM-027, TERM-017). A31’s mapping of the six relations onto training: accessibility to architecture, data selection and optimizer; precedence to loss weights and optimizer geometry; jurisdiction to the computational graph and parameter groups; persistence to parameters plus training dynamics; suspension to optimizer, scheduler and training controller; transformation to the training algorithm or a meta-controller. “much of that arrangement is outside the neural network’s jurisdiction. That doesn’t make the governance unreal. It means it belongs to a larger system.”
THEORY-253 is the documented ground for step 3 in a solvable model: with the network’s task structure fixed, the optimizer alone decides whether independent skills are learned together (gradient descent) or compete and emerge in sequence (sign descent).
The methodological rule
A31 §10: “Before attributing affect to a trained system, determine whether the proposed affective state belongs to the network, the learning controller, the larger training apparatus, or some emergent organization spanning them.” This claim is the rule applied to conventional training. It is kept here and not filed as a separate entry.
Acquiring and governing
A31 §7: “acquiring capacities ≠ governing the acquisition of capacities”. “During ordinary training, the neural network becomes more capable because it is being governed by something outside itself”, and “The network’s development can be highly organized without being self-directed.” This is a second appearance of the record’s distinction between effective action and the capacity to govern among regimes (ARG-007), and of CLAIM-284’s self as one that takes part in governing its own persistence and transformation. A31’s own three achievements make the same cut: learning a representation, maintaining a learned capability, and “Maintaining the learning organization itself”, of which “Only the third directly approaches the stronger organizational-autonomy criterion.”
For affect it matters because, A31: “frustration normally presupposes a relevant demand or commitment belonging to the affected system. External optimization failure alone does not establish such a commitment.” A31’s Case A, conflicting gradients g_1ᵀg_2 < 0 under a trainer who wants both losses minimized, is “Optimization strain in the training process”. (The sign of g_1ᵀg_2 is a coordinate-dependent indicator of competing directions, not a test of frustration; CLAIM-366.) U33 later made frustration, not distress, the primitive (CLAIM-368); the claim is stated for any regulatory condition, so it holds of either.
What it does not say
- It does not say training-time governance is unreal, or merely metaphorical. It says whose it is.
- It does not say a network cannot come to hold it. A31 §8: “The relevant transition would be in who or what governs adaptation”, and “I wouldn’t assume such a transition occurs in ordinary deep learning. But it is a clear hypothesis for systems with internal adaptive controllers.” Systems B and C of CASE-113 are where it would occur.
- It does not say the training system is an agent with its own norms. Its objectives are the trainer’s: “its normative significance derives from the training objective chosen by the trainer.” CLAIM-288 says a modeller-fixed objective cannot show self-constitution, and the training system inherits that limit. A31’s “Unity of optimization” is not its “Unity of self-governance”.
- It does not decide anything about feeling. A31: “None of these capacities, individually or jointly, establishes phenomenal consciousness.”
- It does not supply the boundary test. CLAIM-194 objects that a coupling criterion cannot tell a boundary a system makes from one imposed on it, and in conventional training the boundary is imposed.
Note of 2026-10-21: the timescale of adaptation
A8 §1 and A9 §5 of the exchange of 2026-10-20 reached this claim’s case unprompted (“Does the agent include its learning process?”) and added a variable: “Agent boundaries can be timescale-sensitive because the mechanisms of adaptation may operate at different timescales from the mechanisms of action.” A9: “A frozen policy can still condition its actions on environmental observations. A learning system can adapt over a slower timescale. Neither fact by itself determines the correct agent boundary.” Under CLAIM-174’s third condition, which agent passes depends on the timescale of the intervention on consequences. That is a parameter this claim’s “which unit is individuated for the training dynamics” leaves implicit.
Note of 2026-10-26: a circuit’s gradient set by the data and the optimizer
T24 of the exchange “Experimental Infrastructure Review” cited this claim for the governance of a transformer’s development (QUESTION-275). One of the works it drew on, Musat, Pimentel, Zucchet and Hofmann’s theory of invariant learning dynamics (NOTE-1177), read on 2026-10-26, gives two exact cases of the division this claim draws, inside its idealized attention-only model.
- The data and another circuit set a circuit’s gradient (THEORY-1009). Once an in-weights circuit has memorized the common pairs, the population gradient on the induction circuit is its gradient without it multiplied by fr + (1 − fr)fv, the share of the data the memorized map does not solve. How fast the induction circuit is learned is fixed by data statistics and by the state of a circuit in orthogonal coordinates, not by anything the induction circuit holds.
- The optimizer decides what structure training keeps (THEORY-981). Gradient descent never leaves the subspace the data’s symmetries fix. An elementwise optimizer such as sign descent keeps it only when that subspace is aligned with the parameter coordinates. Whether the network’s learned organization stays in the symmetric subspace is a fact about the optimizer–network–data system.
Both are theorems in a model with fixed write maps, under plain gradient descent for the first, and neither is a crossed intervention of the kind CASE-113 asks for. The claim is unchanged.
Note of 2026-10-27: a redundancy the training regime may have put there
Wang et al.’s IOI circuit in GPT-2 small (LIT-1540, NOTE-1167), read on 2026-10-26 at the owner’s request, has backup name-mover heads that take over when the name movers are knocked out, so the circuit keeps most of its effect (THEORY-980). The authors explain it in one sentence, by dropout during training (THEORY-995); the reading records the alternative that attention re-routes after the ablation. If the dropout account holds, the redundancy is a property the trained network carries whose source is a condition of its training regime, the boundary this claim draws between what governs training and what belongs to the network. No comparison with and without dropout exists, and nothing here tests it. The claim is unchanged.