nucleation

CLAIM-288: In a population whose operators each maximize an objective the modeller fixed, the norms the organization appears to maintain are set by the modeller, not by the organization: they can become its goals through their role in selecting what persists, but it does not produce or revise them

4 min read

Properties13
StatusActive
Version7
Rolecounter
Warrantinterpretive
Tagsself-governance, agency
Linedistributed-agency
Worksself-as-recursive-ecology
GroundsTHEORY-1119 — Barandiaran, Di Paolo and Rohde propose that minimal agency requires individuality, interactional asymmetry and normativity set by the system's own organisation, each necessary and together, they say, sufficient; that individuality without the other two is not agency; and that a system optimizing an externally fixed function is not a model of agency, LIT-1017 — Autopoiesis, Adaptivity, Teleology, Agency, THEORY-1166 — Di Paolo argues that autopoiesis as originally defined entails conservation and not active homeostasis, so it gives a system only an all-or-nothing norm, to go on or to die, which an observer and not the system can use; that a graded norm the system itself has needs adaptivity, a capacity to detect tendencies toward the boundary of its viability and to act so as to reverse them; and that sense-making and agency need adaptivity
Objected byCLAIM-356 — An objective that originates outside a learning system can become a goal of the training organization, through its causal role in selecting, stabilizing and reproducing the learned structure it rewards; exogenous origin does not disqualify a goal, but it does not show that the network governs or endorses the objective, and what is selected is performance on the updating distribution, not generalization
UndercutsARG-015 — A population of operators that maximize fixed objectives and comes to hold durable authority would show self-constitution
ComplementsCLAIM-367 — Whether a training-time regulatory condition belongs to a network depends on which unit is individuated for the training dynamics and whose jurisdiction governs the updates; in conventional training most of that governance falls to the optimizer–model–data system, so a condition of training is not thereby a condition of the network
Corroborated byTHEORY-219 — Agency in the sense that bears on responsibility needs a standpoint that settles and carries forward revised evaluative standards when its inherited ones break down, and planning, autonomy, memory and coordination do not supply one, THEORY-220 — A far-from-equilibrium system that contributes to its own maintenance has functions relative to itself, with no history or observer needed, and one that switches among self-maintaining processes has indications of success that can be false by its own lights, THEORY-221 — Organizational closure is a circular organization of efficient causes by which a system produces the conditions of its own existence, distinct from feedback, self-organization and catalytic cycles; systems whose loops do not regenerate the constraints that make them possible approximate closure without instantiating it
Rendered intodocs/arguments-by-conclusion.md

Graph View

The objection

CLAIM-284 is meant to be tested by building populations. A10 §7 proposes “Sufficiency tests. Construct systems satisfying the proposed organizational conditions without adding predefined self-models or executive controllers.” A7’s ideal first result is a population that develops “specialized niches, persistent but revisable trust, and higher-level governance structures solely through local transformation, temporally discounted policy evaluation, and constrained endogenous endorsement.”

In the model the exchange built, each operator is rewarded for something the modeller wrote down. In A2 §6 it is J_i = E[Q_i(a*, c)] − λ_i Cost(E_i). In A6 §1 it is discounted reward r_i. Barandiaran, Di Paolo and Rohde (THEORY-1119, read in NOTE-458) derive that “Systems optimizing an externally fixed function are not models of agency” (C5), because normativity must be “Regulation against a condition that can be failed, with the condition set by the system’s organization.” Di Paolo (2005, LIT-1017) makes the same point about norms: they must be norms for the system, which it can use.

So a population of reward-maximizing operators that came to hold durable, differentiated authority would show that endogenous governance can arise from fixed local objectives. It would not show that the organization sets the conditions it maintains, and that is the constitutive part of the thesis. A10 §6.1 named the boundary case and left it open: “a system that learns to maintain itself only because an external designer rewards that behavior.”

What would answer it

  • Operators whose objectives are themselves produced, weighted and revised by the organization, such as the viability of the coalition they belong to. A10 §3.2 moves this way: “Instead of imposing a global reward function defining what the system ought to do, ask whether its own organization generates conditions whose violation threatens its continued existence.” CLAIM-075 warns that viability alone does not settle what ought to persist.
  • An argument that the modeller’s objectives are only initial conditions that the organization then revises. A10 §6.1 calls this the difference between the historical origins of agency and its current constitutive organization.

What it does not say

It does not say the simulations are worthless. They can test CLAIM-277 and CLAIM-279, which make no constitutive claim.

Note of 2026-10-18: applied to training, and objected to

The exchange of 2026-10-17 met this objection three times without citing it.

  • Applied to training. A30 §8 made the loss the modeller’s norm: “the task loss is generally chosen by an experimenter. Lower loss is desirable to the trainer, but that doesn’t mean it has the same status as an endogenously maintained condition of an organism.” A34 kept the guardrail (“Loss minimization does not automatically mean the model has its own goals”).
  • Objected to. The owner, at U35: “the externally assigned objective can motivate identification and persistence of learned features”, so “it seems reasonable to describe the training process as having endogenous goals”. A35 conceded in part: “Yes, in a qualified but important sense. … I was treating the origin of a goal and the organization through which it becomes effective as though they were the same question.” It kept the half this claim needs: stable structure “does not alone show that the learned network governs or endorses the criterion that selected it”. That is filed as CLAIM-356, which objects to this claim in part. It grants it for self-constitution and denies that exogenous origin bars a goal from becoming the training organization’s regulation. This claim’s own answer condition, objectives “produced and revised by the organization”, is what CLAIM-356 calls the governance grade.
  • Used as a control. A38 §6 asked for comparison “against an external optimizer that implements the same protective policy: otherwise we won’t know whether the governance genuinely belongs to the candidate learning organization”, and A40 §2.3 said ordinary gradient descent gives an effective goal and maintained representations but not “the governance and revision of the processes that determine what remains effective”. Both are this claim’s distinction, used as a design control.

Note of 2026-10-28: split under ADR-061

The claim held two things. That the norms of a population of fixed-objective maximizers are the modeller’s is about the subject matter, and it stays here; CLAIM-356 disputes exactly that clause. That such a simulation “cannot show self-constitution” is a verdict on the sufficiency tests as evidence for CLAIM-284, and it moved to ARG-015, form evidence, Proposed, with CASE-081 as its premise. This claim undercuts that argument. The text above is otherwise unchanged.

Backlinks