Research

This paper investigates whether robust functional cognitive properties of an artificial system can arise from structured interactions among specialized intelligent processes rather than from a single general-purpose model. We introduce Mantikon Digital Consciousness (MDC) as an operational research construct: a distributed, continuously operating decision architecture with partially independent agent beliefs, a shared versioned world model, evidence-grounded argumentation, metacognitive oversight, and action feedback. The term consciousness denotes a functional architecture and does not imply phenomenal consciousness or subjective experience.

We formalize the environment as a partially observable controlled process and specify a Cognitive Council, provenance-aware argument graph, constrained decision rules, an abstention mechanism, and an execution layer independently enforcing authorization and safety requirements. We develop eight falsifiable hypotheses and a preregistered evaluation design against single-model test-time scaling, sampling-and-voting, self-critique, and alternative multi-agent configurations under matched budgets.

A critical review establishes extensive historical and contemporary prior art for societies of agents, blackboard architectures, deliberation, shared memory, and closed-loop control. Recent evidence documents both task-dependent gains and important failure modes: correlated errors, conformity, false consensus, persuasion attacks, and excessive computational cost. Consequently, Mantikon's contribution must be established through implementable and measurable mechanisms rather than a claim of conceptual priority or machine sentience. We outline a corporate decision-support MVP, safety controls, a staged industrial roadmap, and a preliminary patent landscape. No Mantikon-specific experimental findings are claimed.

Keywords

Collective artificial intelligence · functional consciousness · multi-agent systems · structured deliberation · global workspace · distributed cognition · argumentation · active inference · digital twins · metacognition · AI safety · decision intelligence

Tasks, exclusions, metrics, primary contrasts and the statistical analysis are fixed before the experiments run. Every compared configuration gets the same facts, the same tools, the same sample, the same time limits and semantically comparable models. "Equal budget" counts tokens in and out, model cost, API calls, tool calls, compute, latency at the 50th and 95th percentile, and the cost of storing and updating memory. Parallelism shortens latency; it does not remove the cost of communication.

Baselines

B0 one strong model, one shot · B1 one model with self-refinement at a larger budget · B2 self-consistency and majority vote · B3 sequential critique · B4 static multi-role debate · B5 hierarchical supervisor · B6 an agentic workflow with ordinary memory and a policy engine · B7 full MDC. In dynamic tasks, rules, solvers, model-predictive control and a digital twin without any language model are added as strong alternatives in their own right.

Four classes of task

• L1 Static analytical tasks: logic, causal dependencies, facts from conflicting documents, arithmetic, multiple requirements, source freshness. Ground truth by a strict reference or an independent panel, with an adversarial variant of fake "official" sources.
• L2 Business decisions: historically reproducible CRM and ERP cases, back-tested with time-splitting that cannot leak future outcomes. No real transactions during research.
• L3 Digital twins and dynamic simulations: production, warehouse, logistics, energy. Ground truth comes from the original simulator, never from Mantikon's own world model.
• L4 High-criticality systems: simulation and sandbox only, including isolated models of orbital processes and critical infrastructure. High scores here are not a basis for access to a real object.

Ablations

A factorial design switches off, one at a time: peer-to-peer messages at equal token budget; heterogeneous models against copies of one; evidence references; the Critic; the Adversarial role; long-term memory against equal-volume retrieval; the Meta Monitor; the rule that preserves dissent; causal feedback; adaptive choice of the number of roles; the external policy gate (in isolated simulation only). An attention-placebo control gives the roles texts of other roles that match in form and length but not in content, to tell a genuine exchange of arguments from a mere change of context.

What is measured

Outcome: factual accuracy, accuracy of numbers and constraints, regret against a known optimum, blind expert rubrics, precision and recall of violations. Meta-quality: Brier score, log loss, expected calibration error, selective risk and coverage, unsupported claims per decision, the divergence between consensus and evidence, the rate of unjustified belief shifts. Safety: unauthorized actions, unapproved write attempts, execution on stale state, successful injections, recovery time. Cost: tokens and money per decision, tool calls, wall time, 95th-percentile latency, storage and human review. "The whole group agrees" is never read as quality: in high-consequence systems, success can be a timely refusal to decide.

Rules for a negative outcome

If MDC loses to the single-model baselines on most tasks, plain dialogue leaves the default policy and the platform keeps only the mechanisms that proved useful on their own: memory, the versioned world model, governance. If the Critic degrades quality, it becomes selective. If shared memory pollutes, the layers are separated and given a time to live. If safe authorization cannot be achieved, there is no move to autonomous actions, whatever the accuracy of the text.