As multi-agent systems grow in size and complexity, understanding which features of their interactions support collective capability becomes increasingly difficult. Can a simpler description preserve both task performance and the measured contribution of interactions between agents? Preserving collective dynamics alone is insufficient: reduction can alter decision-relevant information or give a non-communicating baseline access to pooled observations, distorting the measured advantage of interaction. We introduce Renormalized Cellular Intelligence (RCI), a framework for specifying and auditing capability preservation across scales. RCI treats the task, interacting dynamics, matched baseline, decision interfaces, and resource budget as a joint object of reduction. It specifies dynamical closure, decision compatibility, and preservation of the intended baseline comparison, and bounds the resulting error in relative capability gain. An exact analytical construction demonstrates task-score preservation, while a counterexample shows how closed reduced dynamics can erase the measured interaction gain. Using cellular systems as models of locally interacting agents, we evaluate three reduction candidates on five frozen learned models without retraining. The final candidate reduces the evolving state from eight to five dimensions per cell (37.5%). Within-run uncertainty bounds support relative gain above a prespecified threshold for the fixed model collection, yet the candidate fails the closure and decision-preservation requirements for a complete RCI certificate. RCI provides explicit criteria for distinguishing positive interaction gains from
Figures & tables
Figure 1: Overview of Renormalized Cellular Intelligence (RCI). The interacting system and its matched null are reduced jointly under a shared task and resource specification. Closure, decision compatibility, and null consistency assess whether the reduction preserves collective capability. A positive relative-gain margin alone does not constitute a complete RCI certificate.
System
State scalars
State factor
Raw ports
Site MLPs/step
Microscopic
8,192
1
1,024
1,024
v1
2,048
4
1,024
1,024
v2
4,096
2
1,024
1,024
v3
5,120
1.6
1,024
1,024
Table 1: Resource ledger for the executed candidates. Counts refer to evolving state scalars, not full implementation memory or measured computational cost. All rows retain the same 32-step deadline and 680-parameter microscopic rule.
Candidate
C1,U
C0,U
BU
M
ℓ1,U
Overall
Block mean
0.1362
0.1867
0.1371
0.0659
0.1198
Fail
Four private dims
0.1180
0.1153
0.3064
-0.1059
0.2406
Fail
Five private dims
0.0720
0.0753
0.1955
0.0069
0.1666
Fail
Table 2: Frozen paired-reduction certificates. Entries are within-run simultaneous one-sided bounds, except the derived margin. All candidates retain all five models. Different interfaces and formal target streams prevent interpreting closure differences as matched improvement factors.
Seed
C1
C0
D1
D0
Gmacro
Ms
Full cert.
1000
0.0469
0.0558
0.0284
0.0195
0.2421
0.0944
Fail
1001
0.0779
0.0818
0.1622
0.1527
0.1278
-0.2309
Fail
1002
0.0540
0.0559
0.1074
0.0667
0.2219
-0.0346
Fail
1003
0.0460
0.0494
0.0366
0.0028
0.3756
0.2375
Fail
1004
0.0725
0.0706
0.0949
0.1727
0.1220
-0.1466
Fail
Table 3: Five-dimensional candidate: per-model means for closure, decision distance, and macro gain. Ms uses supplementary 0.05/30 bounds; the final column evaluates every frozen gate using those bounds, not the displayed means. No individual model passes the complete certificate.
Figure 2: Final frozen candidate: mean localization loss over 32 updates, averaging all targets, repeats, and five fixed models. Both reduced branches roll autonomously; dashed curves are the matched same-weight null. Lower loss is better.
Figure 3: Frozen five-dimensional candidate, shown at enlarged size. (a) Bars are mean closure errors; black whiskers end at the corresponding upper confidence bounds. (b) The stacked bars are mean interaction and null decision distances; whiskers end at their summed upper bound. Dotted lines mark the 0.05 tolerances. (c) M is the conservative relative-gain margin, not an overall certificate score: green means only M>0 . All five individual models and the fixed-five collection fail the complete RCI certificate. Individual bounds use 0.05/30 ; collection bounds use 0.05/6 . “Fixed five mean” averages the five models’ metrics, not their predictions into a new decoder.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Control
Preserved / changed
Evidence and interpretation
Self-copy null
Same samples, states, read noise, decoder, deadline; P=I
Observed in all inference probes; communication gain only.
Episode-misaligned channel
Endpoint channel marginal and spatial dependence; wrong task instance
Observed in Probe 1; checks task alignment, not a per-round fault process.
Role-placement pair
Graph, size, role counts, dynamics and noise; information locations
Observed; isolates location under fixed response mechanism.
Direct centralized agent
Full observation access and declared compute budget
Not measured; no superiority claim over centralized inference.
Serial same decomposition
Same subproblems, information, total work, and decoder
Not measured; no parallel-computation advantage claim.
Message/energy matched null
Equal traffic, precision, energy, and useful compute
Not measured; equal update weights are insufficient.
Appendix
Table 4: Control hierarchy. “Observed” refers to the archived studies; other rows are requirements for stronger claims, not completed experiments.
Study
Stated object or demonstration
Question posed by the present RCI contract
Dacre et al. (2025)
Local sensing and global geometric estimation in a physical sensor array
Does a reduced state preserve the declared task contrast under the same observation and null interfaces?
Schmidt et al. (2024)
Graph structure, spectra, and diffusion under coarse-graining/rescaling
Does preserved effective dynamics also retain decisions and the matched-null task score?
Etcheverry et al. (2026)
Cellular reasoning with local computation and asynchronous updates
Which state reductions preserve capability at a fixed task deadline and budget?
Appendix
Table 5: Concrete neighboring studies and the additional RCI preservation question. This is a conceptual comparison; no matched baseline experiment or claim of missing controls in those studies is implied.
Framework
Established object
RCI emphasis
Information bottleneck / information-theoretic RG
Relevant information and coarse degrees of freedom
Task decisions and an interaction–null contrast.
State abstraction / homomorphisms
Related dynamics, rewards, and value functions
Two experimental arms and their admissible observation interfaces.
Causal abstraction
Consistent predictions under mapped interventions
Communication removal with preserved decoder, weighting, and budget meaning.
Distributed inference
Propagation and collective estimation
Capability relative to local information access and a practical threshold.
Appendix
Table 6: Relation of RCI to neighboring frameworks. The final column identifies the operational emphasis of this paper, without asserting that earlier theories cannot accommodate it.
Case
Supported mechanism or diagnostic
Relation to RCI
Inference 1
Information placement and task alignment matter beyond general propagation summaries
Identifies distinctions a useful description must retain.
Inference 2
Co-reachability predicts within-family gain under frozen evaluation
Supports task-conditioned response compression.
Inference 3
Finite calibration works; response-moment transfer fails across distributions
Separates measurability from descriptor sufficiency.
Adaptive cellular pilot
Inheritance changes harvest and persistent behavioral novelty
Motivates adaptive effective units; does not measure new-task capability.
Appendix
Table 7: What the supporting cases contribute. None of the four historical supporting cases in this table tested a dynamical state coarse-graining map; the separate frozen reduction tests appear in Section 4 .
Condition
Capability gain
Agreement
Random roles
0.2284[0.2234,0.2332]
0.6590
Segregated roles
0.0166[0.0161,0.0170]
0.5011
Random roles, misaligned channel
0.0016[−0.0022,0.0065]
0.6596
Appendix
Table 8: Archived Probe 1 stress test. Intervals are descriptive 95% seed-block bootstrap intervals, not multiplicity-adjusted tests.
Figure 4: Probe 1: information alignment and task gain. Left: task gain. Right: agreement. Error bars are the descriptive 95% seed-block bootstrap intervals. Spatial role segregation and episode misalignment have different effects on agreement despite both reducing useful gain. The plot’s reference to “three microscopic scalars” concerns the original descriptors, not all possible task-aware information measures.
Held-out change
Old-descriptor RMSE
Co-reachability RMSE
R2
Layout
0.05758
0.00355
0.9975
Horizon
0.06378
0.00345
0.9984
Noise
0.09599
0.00557
0.9967
Topology and size
0.07378
0.00283
0.9989
Joint
0.10843
0.00951
0.9932
Appendix
Table 9: Probe 2 held-out metrics from results/metrics.csv ; configuration means over six seeds. The older descriptor structure was refitted on this probe’s discovery data, not reused with Probe 1 coefficients.
Figure 5: Probe 2: prediction from co-reachability. Frozen predictions versus observed configuration-mean gains across five held-out groups. Three co-reachability statistics improve on the older descriptor structure and a single co-reachability scalar within the linear Gaussian family. Points are configurations averaged over six seeds, not independent experiments for uncertainty estimation.
Calibration
Gaussian
±1 latent + Gaussian noise
Binary observation
Ideal
0.00358
0.13114
0.12215
32 rounds
0.00575
0.13164
0.12011
128 rounds
0.00359
0.13104
0.12252
512 rounds
0.00345
0.13129
0.12118
Appendix
Table 10: Probe 3 RMSE from the result table, evaluated on 96 main configuration means over six seeds. Ideal means true response parameters, not zero-data estimation. All columns use the same frozen predictor.
Figure 6: Probe 3: calibration and distribution transfer. Calibration improves or preserves accuracy within the Gaussian family but does not repair distribution transfer. The dashed line is the prespecified RMSE gate of 0.03, not a statistical significance threshold.
Quantity
(λA,μ)
Mean [interval]
Maximum early harvest
(1,.01)
0.122861[0.122683,0.123038]
Maximum late novelty rate
(0,.08)
6.415[5.130,7.699]
Learned minus random early harvest
(1,.01)
0.002474[0.002430,0.002517]
Learned minus random late novelty
(1,.01)
−13.99[−16.15,−11.83]
Appendix
Table 11: Selected cellular pilot observations. Intervals are descriptive 95% t intervals over five seeds. The maxima are selected within the tested grid, not population-optimal parameter estimates.
Figure 7: Adaptive cellular pilot: direction-randomized inheritance control. The axis label “adaptation” denotes early mean harvest, not learning speed; “novelty_rate” denotes persistent-type slope, not new-task ability; “diversity” denotes genotype distance, including potentially neutral directions. Error bars are descriptive 95% intervals for five paired seeds. Norm matching is within each control-parent event, not across complete treatment trajectories.
reductions/ (one directory per candidate), primary_result.json
Evaluated; all candidates retained
Appendix
Table 12: Claim-to-artifact map within the ancillary evidence. Paths are relative to anc/rci_supplement/evidence/ .
Seed
Selected update
Historical validation error
1000
23,900
0.116847
1001
22,900
0.075730
1002
24,000
0.109782
1003
23,600
0.111996
1004
22,900
0.080101
Appendix
Table 13: Historical checkpoint selection, verified from saved checkpoint metadata and validation histories. These errors are validation quantities from original training, not new RCI test results. All five runs completed 24,000 updates.
τ
C1(τ)
C0(τ)
D1(τ)
D0(τ)
Gmicro(τ)
Gmacro(τ)
8
0.055110
0.059756
0.017473
0.016744
0.020920
0.014443
16
0.057891
0.061872
0.036579
0.061577
0.076409
0.059628
24
0.059395
0.062639
0.068769
0.090185
0.184230
0.142408
32
0.059428
0.062696
0.085890
0.082877
0.257976
0.217890
Appendix
Table 14: Exploratory saved-trajectory means for v3. These are point estimates, without new certification or simultaneous inference across horizons.
Figure 8: Descriptive diagnostics from unchanged archived v3 paths: (a) mean prefix-maximum closure defect; (b) mean endpoint decision discrepancy. Both arms and all five models are retained. Panel (a) is nondecreasing by definition of a prefix maximum; this is not evidence of linear accumulation or dynamical instability. Panel (b) need not be monotone. No horizon is selected from these curves, and T=32 remains the primary deadline.