From Microscopic Capability to Collective Function v0.5
An Evidential Ladder and Reducibility Audit for Interacting Units
Author: Kai Wang
Version: 0.5
Status: Preprint
Type: Methodological / Conceptual Perspective with computational audit examples
This paper proposes an evidential ladder and reducibility-audit workflow for claims about interacting units, active matter, physical learning, and collective function. Its central methodological question is not merely whether a richer microscopic mechanism changes a system, but whether the claimed macroscopic effect survives the strongest justified simpler or lower-level alternative.
The computational audit examples emphasize adversarial architecture comparison, held-out evaluation, separation of model selection from final inference, and explicit claim downgrade when a reducer closes the supposed explanatory gap. The paper does not treat increasing microscopic architectural capability as sufficient evidence for increasing collective capability.
Relation to Effective System Capability
Across the current website, Effective System Capability is the broader system-level construct. In this paper, collective function is used more narrowly for a candidate system-level function whose attribution to interacting components is being tested.
The audit is deliberately defeasible. It cannot prove that no imaginable reducer exists. It can, however, strengthen an organization- or collective-level attribution when a claimed effect survives strong matched reducer families, and it can downgrade that attribution when a simpler family closes the explanatory gap.
Porting the audit to AI organizations
The v0.5 paper develops a domain-general methodology using active-matter and physical-learning examples. When the same audit is applied to an AI-organization claim, the reducer families should be instantiated for the AI domain rather than copied literally from active matter.
Depending on the claim, relevant reducers can include:
- one stronger node;
- one model with additional sampling or best-of-N selection;
- majority vote, self-consistency, or another simple centralized aggregation;
- matched compute, context, tool access, or observation budget;
- prompting or scaffolding supplied equally to the lower-level alternative;
- an external human, lead orchestrator, or other operator whose contribution lies outside the declared system boundary.
The last item is a boundary-accounting question. If a human manager is declared as part of the system, that contribution is internal to the analyzed organization. If the human is outside the declared system while supplying the capability being credited to the AI organization, it is an external-operator reducer.
Independent or preregistered reducer construction can increase confidence that the search over simpler alternatives was sufficiently adversarial. It is not a logical requirement for a valid reduction. The logical question is whether the reducer actually reproduces the target effect under the matched conditions; independent construction primarily protects search completeness and empirical credibility.
These portability notes describe how the published audit can be instantiated in another domain; they do not alter the contents or claims of the archived v0.5 PDF.