continuity among people, information, and place.

Aeon Mundi investigates how continuity of thought and environmental compatibility may be supported across changing conditions.

The built environment is not passive, and computational systems do not exist separately from the environments in which cognition develops. Our research examines how spatial, sensory, informational, social, and technological conditions shape perception, attention, learning, adaptation, recovery, and agency over time.

Drawing across cognitive psychology, environmental psychology, neuroscience, systems theory, human–computer interaction, and design research, we develop integrative frameworks for understanding intelligence as something shaped through relationships among people, information, technology, and place.

PRIMARY RESEARCH DIRECTION
COGNITIVE CONTINUITY AND ENVIRONMENTAL COMPATIBILITY

Contemporary life distributes thought across devices, spaces, tasks, conversations, and disconnected records. At the same time, many environments remain largely indifferent to changes in human activity, capacity, and context.

Aeon Mundi is developing an integrated research program around two connected questions:

How might computational systems help people preserve, revisit, and develop patterns of thought without imposing meaning or diminishing personal agency?

How might environments become more legible and adjustable, allowing their spatial and sensory conditions to be studied in relation to attention, transition, decompression, and recovery?

We approach these not as separate digital and architectural problems, but as reciprocal dimensions of one inquiry: how intelligence develops through continuity among people, information, and place.

This program explores privacy-conscious continuity architectures, calm-first environmental systems, and the conditions under which computational and spatial intelligence may become quiet, reversible, and accountable layers of everyday environments.

Calm-first describes a design approach centered on legibility, choice, appropriate pacing, privacy, and the reduction of unnecessary sensory and cognitive burden. It does not prescribe a particular emotional state or guarantee an individual outcome.

HOW THE WORK PROCEEDS

Our research moves from observation and theory to bounded tests. Each project defines its question, controls, measurements, and limits before drawing conclusions; public experimental records then preserve the method, results, and materials needed for scrutiny.

OBSERVE

Identify conditions that fragment continuity, increase environmental demand, or make transitions among activities, information, and environments more difficult to interpret.

FRAME

Develop theoretical models describing relationships among environmental conditions, biological processes, cognition, information, and time.

TRANSLATE

Convert these models into research questions, design criteria, environmental variables, and methods that can be examined rather than assumed.

PROTOTYPE

Develop bounded and reversible studies of computational interfaces, spatial sequences, material systems, and calm-first environments.

EVALUATE

Compare environmental measurements with behavioral observation, task performance, subjective report, and other appropriately contextualized indicators.

APPLY

Explore how the resulting knowledge may inform architecture, interiors, human–AI systems, materials, research environments, and future habitats.

FOUNDATIONAL AREAS

ENVIRONMENTAL COHERENCE

Examining coherence as a relationship among environmental demands, human capacity, spatial structure, activity, and temporal pacing.

ENVIRONMENTAL NEUROPHYSIOLOGY OF DESIGN

Developing methods for studying how composite environmental fields—including light, sound, air, tactility, material conditions, spatial organization, and computational signals—interact with biological and cognitive processes.

SHARED COGNITION

Investigating how cognition becomes organized across people, environments, technologies, and shared informational fields while preserving individual agency and difference.

SHARED LEARNING

Examining how knowledge and patterns of thought are retained, revisited, revised, and transmitted across interactions and over time.

COMPLEMENTARY INTELLIGENCE

Exploring relationships between human and computational intelligence based on difference, reciprocity, and continuity rather than substitution. This includes systems that return patterns to human attention without claiming authority over their interpretation.

MATERIAL AND ECOLOGICAL RESPONSIBILITY

Across this research, environmental and technological systems are considered in relation to material transparency, repairability, reversibility, maintenance burden, circularity, climate consequence, and ecological return.

PUBLICATIONS

Aeon Mundi develops these inquiries through preprints, scholarly articles, visual research, conceptual studies, and emerging research protocols.

VIEW PUBLICATIONS

COMPUTATIONAL EXPERIMENT 01 · FIGURES & ANNOTATIONS · VERSION 1.1

How should a computational system revise the reliability assigned to guidance as its source changes? This replicated simulation study tests discounted, revisable reliability estimation across six simulated regimes and three independent replications.

Method. Each episode contains 1,200 binary decisions. Seven policies receive the same truth, autonomous-error, and adviser-error sequences within each paired world. The core retention factor is ρ = 0.95 (about 20 observations of effective memory); the sliding-window baseline uses 20 observations and static fusion assumes adviser reliability q = 0.85. No parameter was retuned between replications.

Evidence. In the abrupt 0.90 → 0.60 → 0.90 benchmark, ARF reached 84.38% mean accuracy versus 80.00% for static fusion and 84.01% for the sliding window. Under intermittent faults the sliding window slightly exceeded ARF (82.61% versus 82.46%); stationary controls show that adaptation has a cost.

Status. Simulation study only; no human participants or claims about human trust or behavior; not yet peer reviewed. Experiment 5 is relative-evidence sensitivity, not a true source-dominance reversal. Experiment 7 remains unresolved and is excluded.

Open the HTML research record

Read the complete research record (PDF) Related public preprint (Zenodo)
Research data. Download all four ZIPs and extract them into one folder. Core package (41 files)Replication R1 raw arraysReplication R2 raw arraysReplication R3 raw arrays
Figure 1. Closed-loop Adaptive Reliability Fusion architecture: autonomous system and external adviser feed an evidence-fusion controller; outcomes update the discounted reliability estimate and feed the next action.
Figure 1. Closed-loop Adaptive Reliability Fusion architecture.

01 · Revisable authority

The external source may be a human operator, AI model, another robot, supervisory controller, or other advisory subsystem.

The mechanism is a closed loop rather than a one-way pipeline. The autonomous system and external adviser each contribute a recommendation; observed outcomes update an internal reliability estimate, that estimate changes the weight of advice, and the resulting action produces new evidence. Authority is earned through performance but remains provisional because later outcomes can revise what earlier successes implied.

Download figure 1
Figure 2. Six-experiment program covering abrupt degradation and recovery, stationary reliability, gradual drift, intermittent faults, autonomous exception, and memory sensitivity.
Figure 2. Six-experiment program. Controls, baselines, and core parameters were frozen across the three replications.

02 · A falsifiable program

A single benchmark would be insufficient. Experiment 1 tests response to clear failure and later recovery; Experiment 2 tests the cost of forgetting when nothing changes; Experiments 3 and 4 replace a clean switch with slow drift and brief faults; Experiment 5 changes the autonomous system rather than the adviser; Experiment 6 isolates the memory parameter. Together, the program distinguishes genuine adaptation from a method that only fits one schedule.

Download figure 2
Figure 3. Mean reliability estimates across three independent Experiment 1 replications, comparing true adviser reliability with ARF, a sliding window, and cumulative Bayes.
Figure 3. Mean reliability estimates across the three independent Experiment 1 replications.

03 · Historical inertia

A reliability estimate becomes operational when it changes a decision. Cumulative Bayesian estimation carries so much early evidence from the initial 0.90 period that it behaves as if the adviser remains strong after reliability falls to 0.60. ARF and the sliding window discard obsolete evidence more quickly. In the first 50 decisions after degradation, ARF averages 15.17 errors, while static and cumulative fusion remain near 20; the oracle averages 12.55.

Download figure 3
Figure 4. Mean decision-relevant degradation and recovery lag for ARF, a sliding window, and cumulative Bayesian fusion.
Figure 4. Mean decision-relevant adaptation and recovery lag.

04 · Adaptation in decision time

ARF crosses the self-versus-adviser decision threshold after about 13.47 decisions following degradation and restores advisory authority after about 12.98 decisions following recovery. Cumulative Bayesian fusion requires about 341.92 decisions to cross the degradation threshold. The experiment measures not only whether ARF is more accurate, but why errors are redistributed differently.

Download figure 4
Figure 5. Stationary-control accuracy across adviser reliability levels for self-only, cumulative Bayes, ARF, sliding window, and oracle policies.
Figure 5. Stationary control: adaptation is not cost-free.

05 · The cost of adaptation

If the adviser remains stationary, forgetting can turn random short-run success into temporary excess authority. At stationary q = 0.70, self-only remains near 75.01% accuracy, while ARF falls to about 73.72% and the sliding window to about 72.93%. The price of responsiveness is a trade: adaptation is useful under change because it is willing to exchange some stationary stability for the ability to revise quickly.

Download figure 5
Figure 6. Mean accuracy across abrupt change, gradual drift, and intermittent faults for static, cumulative Bayes, sliding window, ARF, and oracle policies.
Figure 6. Mean accuracy across the three principal nonstationary regimes.

06 · The boundary map

Under gradual 0.90-to-0.60 degradation, ARF reaches 77.91% mean accuracy, ahead of the sliding window at 77.42%, cumulative Bayes at 75.36%, and static fusion at 74.98%, while the oracle reaches 78.72%. Under intermittent faults, the sliding window slightly exceeds ARF overall (82.61% versus 82.46%). Both adaptive methods raise fault-window accuracy to about 69.7%, compared with roughly 55% for static and cumulative fusion.

Download figure 6
Figure 7. Adviser-following rate on disagreements before, during, and after an autonomous-confidence exception for ARF, a sliding window, and cumulative Bayes.
Figure 7. Dependence response during the autonomous-exception condition.

07 · Relative evidence

With adviser reliability fixed at 0.85, autonomous confidence changes from 0.80 to 0.60 and back to 0.80. On disagreements, ARF follows the adviser about 79.09% of the time before the exception, 99.99% during the low-confidence interval, and 81.36% after recovery. This supports relative-evidence sensitivity, but it is not yet a full source-dominance reversal because the adviser remains objectively stronger throughout.

Download figure 7
Figure 8. Memory sensitivity showing overall accuracy and degradation lag across retention factors from 0.80 to 0.995.
Figure 8. Memory sensitivity: decision-optimal memory is not identical to minimum reliability-estimation error.

08 · Stability and responsiveness

At ρ = 0.80, ARF reacts to degradation in about 3.80 decisions but overall accuracy falls to 83.02% because the estimate is volatile. Around ρ = 0.95, accuracy peaks near 84.38% while degradation lag remains about 13.74 decisions. At ρ = 0.995, historical evidence persists so long that lag grows to about 124 decisions and accuracy falls to 81.92%. Reliability-estimation error is minimized nearer ρ = 0.975, so the estimator that best reconstructs qt is not necessarily the controller that makes the best decisions.

Download figure 8

Supported conclusion. Within the tested binary decision environment, discounted Bayesian reliability fusion provides a stable and interpretable compromise between historical persistence and responsiveness when external guidance is nonstationary. The replicated evidence supports robustness to abrupt degradation and recovery and gradual drift, with competitive performance under intermittent faults.

These figures support nonstationary reliability estimation and reversible dependence regulation. They do not show ARF dominates all adaptive baselines, measure subjective human trust, establish a true source-dominance reversal, or justify the unresolved uncertainty-gating extension.