Why it is a question
Suppose influence and learning opportunities are coupled:
influence → opportunities to demonstrate competence → learned competence → influence.
A specialist that is temporarily unpopular gets fewer observations and less training. Its measured competence declines, and the decline justifies further suppression. In the end it may become genuinely incapable of the role it was specialized for. That is different from underuse. The organization has changed the specialist’s future competence by denying it the opportunities it needed to keep that competence.
CLAIM-277’s note of 2026-10-18 names the loop at the level of estimates. CLAIM-356 distinguishes passive, selective and regulated retention of a learned capability. A representation can stay stored in the weights without being protected against interference from later training. A specialist can stay nominally present without being able to reassert influence. CLAIM-362 asks that a process keep access to consequential roles. This question asks what makes that access real: recovery requires that the competence survive the period of disuse.
The niche is partly endogenous. Operators’ influence changes which evidence is later encountered and used to evaluate them (CLAIM-297). A system evaluated by probes that also supply its training signal can become excellent at the probes without keeping the intended capacity (CLAIM-266).
What would answer it
An experiment separating three things:
- loss of influence: the specialist’s weight falls;
- loss of competence: its performance in its niche, measured on held-out evaluation it was not selected on, falls;
- loss of access: the training and evaluation opportunities it receives fall.
It needs two distributions: one that generates the system’s adaptive selection pressures, and another that independently checks what capabilities it retains. The decisive result would be a specialist whose competence on the independent distribution is lost through deprivation and does not return when its contexts recur, against a control whose opportunities are decoupled from its influence.
What would not answer it
A specialist that loses influence and keeps its competence. That is under-use, which QUESTION-132’s recovery exponent already covers.
Note of 2026-10-26: inside one network, access follows influence by construction
The reply at T24 of the exchange “Experimental Infrastructure Review” brought this question’s split into a single transformer (QUESTION-275): “A component might have a large effect on current predictions but receive little additional training signal. Conversely, a newly developing feature might receive substantial gradient updates while remaining behaviorally weak. We should measure these separately.”
Within one network the coupling is not a design choice. Backpropagation scales the gradient that reaches a component by the weight of the path leading out of it.
- For y = w·h(θ), ∂L/∂θ = (∂L/∂y)·w·h′(θ). For y = w·tanh(θx) the gradient norm on θ was 1.95, 0.137 and 0.013 at w = 1, 0.1 and 0.01.
- In softmax attention, the gradient on a value vector v_j is exactly α_j times the gradient on the output (checked to 1.5 × 10⁻¹⁰).
So influence-dependent access is the default. A low-influence path’s upstream parameters are starved while its downstream weight still learns, and THEORY-186’s stages from small weights have the same product structure. CASE-117’s decoupled-exposure control has no counterpart in one network without an intervention on the gradient path itself, such as an auxiliary loss on the component or a stop-gradient. The question is unchanged.