Digital
Consciousness
A decision system whose mind is a dialogue. Several specialized models hold partially independent beliefs over one versioned model of the world, argue a case from evidence under a protocol, and settle on a decision that carries its reasoning, its dissent and its review. The term names an architecture, not an experience.
The thesis
In the human brain, no single system decides. Perception, memory, appetite, fear and planning speak at the same time, each pressing its own need, and the mind that emerges is the conversation between them. Consciousness is not a part of the brain. It is what the parts do together.
Mantikon rebuilds that process from specialized AI models. A case is put before a council of specialized models, and every model holds one function: one observes, one forms hypotheses, one attacks them with evidence, one counts the risk, one checks the policy, one writes the recommendation with the dissent attached. No model holds two functions, and no model decides alone. The decision is what the dialogue settles on.
That is Digital Consciousness: not a bigger model with a longer prompt, but a mind made of a dialogue. The hypothesis behind it is narrower than the name and testable: that a structured dialogue, with evidence provenance, independent critique, a calibrated right to abstain and a closed loop from action back to observation, can decide better than a strong single model at the same cost, on the classes of decisions where interests genuinely conflict.
Everything else on the platform, the data it ingests and the actions it is allowed to take, exists to feed this dialogue and to carry its decisions out safely. What the dialogue is for is the subject of Foresight: modelling the futures a decision would create, and the metacognitive duty of checking our own forecasts.
The
definition
Mantikon Digital Consciousness (MDC) is an architecture for continuous, distributed decision-making in which a set of partially independent specialized computational processes maintains a versioned probabilistic model of a controlled environment; produces hypotheses and objections whose provenance can be verified; conducts resource-bounded collective deliberation; monitors its own errors and uncertainty; chooses, rejects or defers decisions under hard constraints; and updates its internal state from the verifiable consequences of authorized actions.
This is an operational definition of a class of architectures. It says what the system must do, not what it is. The word "consciousness" has several meanings, from wakefulness to the availability of information for report and control, to self-monitoring, to subjective experience, and these are not the same thing. Mantikon is built as a functional cognitive architecture. Whether it experiences anything does not follow from the definition and will not follow from any test it passes.
Eight requirements
A system is an instance of MDC only if it meets all eight. A system without the first and the seventh is a debate engine for one session; a system without the third to sixth is an ordinary orchestration or a voting ensemble.
- D1
Durable state
A long-lived identity of the environment's state, with versions, time and the provenance of every datum.
World model - D2
Real independence
At least two genuinely distinct mechanisms for forming hypotheses and one independent channel for refuting them. A role in a system prompt does not count as independence.
Hypotheses - D3
Typed exchange
Discrete, verifiable claims and counter-evidence are exchanged, not unbounded text.
Messages - D4
Recorded contradiction
A mechanism that keeps unresolved contradictions on the record instead of smoothing them over.
Dissent - D5
Adaptive stopping
An explicit decision whether to keep reasoning, fetch more data, or stop.
Budget - D6
Calibrated abstention
Uncertainty calibrated against external criteria, and the right to refrain from acting.
Abstain - D7
Causal journal
A causally and temporally linked log of decisions, actions and observed consequences.
Audit - D8
Separate shield
An execution layer for safe authorization that does not overlap with the language models.
Policy gate
When a
council wins
Not always, and not by itself. The research on multi-agent debate is mixed: on some tasks a structured exchange improves accuracy; on others, voting alone explains the gain, homogeneous groups imitate each other into a false consensus, and one persuasive wrong voice can drag the rest. Mantikon does not assume that ten models are better than one.
A council is expected to win under three conditions. The task divides into parts that need different tools and knowledge. The parts' results are partly independent, so that disagreement surfaces before execution. And the cost of communicating and verifying is smaller than the expected cost of the errors it prevents. It is expected to lose on tightly sequential reasoning, on correlated models, under time pressure, with unreliable sources, or when the roles defer to each other's authority.
So the architecture is built around the conditions, not around the number of roles: independence that is measured rather than declared, a budget for deliberation that is allocated by the risk of the case, and evidence weighed by its source rather than by how many roles repeat it.
Lineage
The idea of a mind as a society of parts is not new, and Mantikon does not claim it. Minsky's Society of Mind described intelligence as the organization of many simple mechanisms. Hearsay-II and the blackboard architectures put specialized modules around a shared working structure. Hutchins showed cognition distributed across people, tools and artefacts. Global workspace theory tied conscious access to the broadcast of a few representations to many processes, and the Conscious Turing Machine gave that competition a formal definition. LIDA and Soar built cognitive cycles with long-term memory.
What Mantikon adds is a particular, implementable combination for live organizations: a versioned world model with the provenance of every claim, independent verification that is measured rather than assumed, a deliberation budget allocated by the expected harm of being wrong, calibrated abstention, and an execution shield that the models cannot talk their way through. Whether that combination earns its keep is the subject of the research programme.
- Not a chatbot.Nobody talks to the council in prose. A case goes in, a dossier with a decision, its reasoning and its dissent comes out, and the dialogue happens between the roles.
- Not one model with a clever prompt.A single model asked to "consider all sides" still weighs them with one mind. The council keeps the functions apart on purpose and measures how independent they really are.
- Not a dashboard.Dashboards show what is. The council argues what to do, says why, and records who disagreed.
- Not a claim of machine experience.No test the system can pass would show that it feels anything, and none is claimed. The word names a functional organization and nothing more.