The case
A network with parameters θ_t, an optimizer with internal state m_t (momentum, moments), a scheduler, a data process d_t and a loss L that sums two or more task losses. A31’s update:
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).
The task gradients have a negative Euclidean inner product, g_1ᵀg_2 < 0, for part of training, in the stipulated parameterization. That is one heuristic indicator of competing update directions, not a finding of conflict: its sign is not invariant under smooth invertible reparameterization, and a negative inner product does not exclude a common descent direction (CLAIM-366). Whether the tasks are actually frustrated is a separate test, against the declared acceptance conditions and the admissible update set (TERM-060). The case needs only the indicator; nothing below turns on its being frustration. The trainer wants both losses low. Everything outside θ is fixed by the trainer: architecture, loss weights, optimizer, schedule, parameter groups, what is frozen or clipped.
Candidate bearers of any condition of training, A31 §1: a layer or circuit; the network during one forward pass; the evolving parameterized network; the network–optimizer pair; the full training process.
The variations
A31 §9, “Three increasingly endogenous learning architectures”:
- System A, externally governed learner. “Fixed network with ordinary optimizer. Conflict resolution is entirely external to the model.” This is the base case.
- System B, internally informed learner. “Network computes a persistent regulatory signal that influences an external optimizer’s update decisions.”
- System C, internally governed learner. “A persistent subsystem monitors task conflicts and retained capabilities, then selects among admissible learning policies. Its own continuity depends on maintaining the capacities it regulates.”
A31 §6 gives the matching attributions: A, “Optimization strain in the training process”; B, “Potential functional regulatory distress of the learning organization”; C, “A candidate integrated affect-like state”. In none is phenomenal experience established.
The interventions
A31 §4:
- Change the optimizer, hold the network fixed. “If we hold the network fixed and change the optimizer, which aspects of its apparent preferences, priorities, and regulatory responses change?” If the candidate regime changes mainly here, it belongs to the training organization.
- Disrupt a learned state, hold the optimizer fixed. “If we hold the optimizer fixed and disrupt an internal regulatory representation learned by the network, does the system change how it responds to conflicting learning demands?” A positive result supports an internal causal role.
- Crossed. Change the optimizer, the network’s regulatory state and the environment independently. A31: “That separates imposed regulation from learned regulation more effectively than inspecting a loss curve.”
For System C, A31 adds an intervention on the candidate regulatory variable with task errors held fixed, ablation of the subsystem that maintains it, and a comparison of threats to continuing capabilities with “superficially similar errors” that threaten none. “That last comparison is crucial: it distinguishes a candidate endogenous evaluative state from a generic error monitor.”
What it can show
- Where, under a stated setup, the record places a condition of training: in System A, with the network–optimizer pair or the training process (CLAIM-367), because the relations that decide which updates take effect are set outside the network.
- What evidence would move it inward: the second intervention coming out positive, or System C’s ablations breaking its regulation.
- The optimizer arm has a documented instance in a solvable model. Under gradient descent, independent skills are learned together; under sign descent, they compete and emerge in sequence (THEORY-253). The network’s tasks are the same in both, so the order of learning is the optimizer’s.
What it cannot show
- Given its standing, how any real run comes out. The case states the design and the reading of each outcome, not the results.
- That System C, if built, sets its own norms. Its tasks and the capacities it protects are still chosen by the modeller (CLAIM-288).
- That the unit picked for the training dynamics is the right unit for every question. For output prediction the network is a natural unit (CLAIM-152, CLAIM-157).
- Anything about feeling.
Note of 2026-10-26: a second appearance, with an exposure arm
The reply at T24 of the exchange “Experimental Infrastructure Review” re-derived this case without citing it, when the owner asked whether the record’s setup could interpret transformer learning dynamics (QUESTION-275). It wrote the update as (θ_{s+1}, m_{s+1}) = 𝒰(θ_s, m_s, ℬ_s), cited CLAIM-367 and THEORY-253 for the external governance and the optimizer arm, and proposed as “the stronger causal test” to “clone the full training state, modify a particular training exposure or update pathway, and observe the subsequent development of identified circuits”.
That adds a third arm to the two this case crosses (optimizer, learned state): training exposure, meaning task mix and curriculum, which the reply’s governance table reads as which capacities receive opportunities to develop. The reply also used g_kᵀg_j as “one local measure” of whether task pressures reinforce or oppose one another. This case already calls it a heuristic indicator, and CLAIM-366 says why: its sign is not invariant under reparameterization. The case is unchanged.