The design
Contexts c_t ~ Π_t, a shared state x_t, and K operators with parameters θ_i, identical in form and differing only by random initialization and experience. Each proposes a_i = f_{θ_i}(x_t, c_t). The collective produces a_t = 𝒜(a₁, …, a_K; w_t), where w_t is the operators’ current influence. Competence updates from experience, and influence from estimated contribution. The exact update rules are experimental variables, not assumptions about the organization.
| dimension | conditions |
|---|---|
| feature adaptation | fixed features; learned features |
| influence update | uniform; historical endorsement (CLAIM-277); context-sensitive contribution |
| competence exposure | training opportunities independent of influence; influence-dependent |
| diversity mechanism | none; fixed usage balancing; endogenous recovery feedback |
| environment | stationary; recurring regimes; abrupt distribution shifts |
| specialist loss | temporary suppression; complete removal; reintroduction |
| recovery opportunities | none beyond a small positive start; externally supplied reintroduction; internally maintained exploration or reactivation |
| authority-update rule | fixed by the modeller; revisable by the organization |
| specialist parameters during recovery measurement | frozen (recovery of the retained capability); trainable (recovery with re-adaptation) |
Baseline. Ordinary mixture-of-experts learning with contextual routing. Without it, the study risks rediscovering specialization that existing architectures already explain (QUESTION-083’s note of 2026-10-18).
What is measured
No one metric substitutes for the rest.
| quantity | reads |
|---|---|
| feature Gram spectrum | representational redundancy (QUESTION-135) |
| I(operator; context) | statistical specialization |
| conditional marginal contribution | whether a specialist adds unique value |
| rare-operator recovery exponent, resident and conditional | recovery when rare (QUESTION-132) |
| recovery time after suppression | access to consequential roles (CLAIM-362) |
| held-out task risk, on a distribution independent of the selection loop | what is retained beyond endogenous evaluation (QUESTION-134, CLAIM-266) |
| changes in the compatibility complex of the operators’ demands | relief or creation of structural frustration (QUESTION-133, CLAIM-347) |
| viability under specialist removal | robustness and capacity to reorganize (THEORY-134) |
An operator’s functional niche can be defined relative to a context distribution, as a proposed operational definition and not a standard ecological one:
𝒩_i(Π) = {c : ΔR_i(c) > 0},
where ΔR_i(c) is the marginal improvement from letting operator i take part in the decision in context c. It is underspecified without the joint task distribution and the downstream learner class. The organization can change the distribution through its decisions, so the niche is partly endogenous. Context-conditioned representations (THEORY-239) give a way to tell a functional niche from the mere frequency of its contexts.
The decisive comparisons
- Same participation, different mechanism. Hold participation roughly constant while varying what maintains it: fixed usage balancing against endogenous recovery feedback. Matching the overall usage distribution is not enough. Two systems can share it while giving experts different contexts, training different distributions or performing differently. So match, as far as they can be matched together, context-conditioned participation, task performance and resource budget. If systems matched on these differ in recovery dynamics, the difference between equal allocation and stabilizing coexistence is measurable.
- Fixed rule against revisable rule. Learning within a fixed authority-update rule can show endogenous differentiation under fixed learning machinery. An organization that revises the rule itself is the stronger result, the one relevant to CLAIM-284’s self-governing organization. The design keeps them as separate conditions.
- Active, dormant and reconstructed specialists. Compare keeping a specialist through active participation, keeping it as a dormant but recoverable capability, and rebuilding it by later learning, on cost and on resilience (QUESTION-132). Measure recovery with the specialist’s parameters frozen and with them trainable. The first tests whether the retained capability is still valued, the second whether the specialist can be re-adapted. A positive result in the second does not imply one in the first.
- Coupled against decoupled exposure. This tests QUESTION-134: whether influence-dependent training destroys a marginalized specialist’s competence.
- Representational novelty against dynamical stability. Check whether they change at the same parameter thresholds (QUESTION-125).
Main threat to inference. An external mechanism that supplies rare opportunities can look like endogenous recovery. Zero is absorbing under multiplicative dynamics, so recovery is first measured from a small positive start, with no further intervention. Externally supplied reintroduction and internally maintained exploration are separate conditions. The external one cannot show endogenous maintenance. The internal one may be an organizational achievement in its own right, if its origin, cost and causal role are established independently. The first version of this design placed every such mechanism outside the organization; the owner’s review corrected that.
What the outcomes would show, and what they would not
A positive result, meaning stable niches, endogenous recovery and reorganization after loss, would answer QUESTION-132 for this class of models. It would not show self-constitution: the objectives are the modeller’s (CLAIM-288), and the unit that governs is the optimizer–model–data system as much as the operators (CLAIM-367). The synthesis proposed four levels of achievement, as a conceptual ordering and not a theorem that each implies the earlier ones:
- differentiation: distinct roles appear through adaptive dynamics;
- coexistence: differentiated components persist through their interactions;
- endogenous regulation: the organization maintains and revises the conditions of its own coordination;
- constitutive autonomy: its continuing organization helps generate, maintain and govern the norms by which it regulates itself.
The study can reach the first two, and with influence rules that the population revises, perhaps the third. The negative outcomes are as informative:
- roles learned and not kept;
- roles kept only by imposed balancing;
- stable coexistence with no endogenous regulation;
- better global performance through systematic domination of some operators.
The relation between persisting and reorganizing is the contrast that THEORY-514 examines, and resilience is not speed of return (THEORY-132).
Note of 2026-10-23: what mixture-of-experts work already shows
A5 of the exchange of 2026-10-23, “Framework Overview Request” cited three works, registered unread: Switch Transformers (LIT-1412), auxiliary-loss-free load balancing (LIT-1414), and a theory of mixture-of-experts in continual learning (LIT-1411), which by its abstract shows experts diversifying to specialize in tasks while the router learns to route and balance load. So specialization and its dependence on routing are shown. A5: what remains is “whether a specialist can remain recoverably competent but practically excluded; how endogenous recovery differs from externally imposed balancing; and whether the organization can revise its routing standards in response to failures those standards helped cause.” CLAIM-369’s coexistence result (with CASE-118’s growth-rate condition) is the record’s own formal piece.
Note of 2026-10-29: the three mixture-of-experts works, read
The three works the note of 2026-10-23 calls registered unread are now read, so that no longer holds. Each narrows what it shows.
- Switch Transformers, LIT-1412 (NOTE-1363). The load-balancing term is there (eq. 4), but it is imposed on the router’s dispatch fractions and mean probabilities, not on what experts learn, and the paper never measures specialization or diversity among experts. It is the standard case of imposed balance of load, and says nothing about emergent diversity, which it did not study. Two small data for this study: sampling the router is worse than argmax and jitter slightly better (Appendix C), which bears on the exploration regimes; and rerouting overflow tokens to a second-choice expert did not help (Appendix B), which bears on practical exclusion.
- Auxiliary-loss-free load balancing, LIT-1414 (NOTE-1388). The exchange cited it as a contribution-based control of influence. It is not one. The bias responds to load, the count of tokens an expert was selected for, and to nothing about how much its output helped. In this study’s terms it is a usage-based quota on access, a regime of imposed balance that, unlike an auxiliary loss, does not enter the gradient. It does bring an excluded expert back: an expert receiving few tokens has its bias raised until it is selected. It does not measure whether experts specialize.
- Mixture-of-experts in continual learning, LIT-1411 (NOTE-1380). It proves specialization among identically initialized experts, with symmetry broken by routing noise and the order of task arrivals, which matches this study’s operators that differ only by initialization and experience, with the caveat that its tasks carry a signal that makes routing separable. Its result that the gate must stop updating for convergence follows from its own gate objective, whose cumulative load-balancing term keeps pulling every gate output together. It is not shown to hold for a router without a balancing term in its gradient. The condition this study can test is the narrower one: when balance is imposed through the routing gradient, does specialization survive continued routing updates? In the paper’s model it does not, unless updating stops. Its tasks arrive uniformly, so rare contexts are not modelled.
The design is unchanged.