A system of
evidence,
perspectives
and action.
Mantikon is being developed as an operational decision layer for complex organizations. Its data plane assembles authorized information into a working world model. Its cognitive layer evaluates alternatives through specialized AI agents and independent critique. Its governance layer records the rationale, preserves dissent and requires the right permissions before any action.
The console is open.
Decision Inbox, World View, Scenario Lab, Decision Ledger, Dependencies and the Council Room, running on our own web index rather than on a mock-up. In this build the roles are deterministic functions rather than models, and every number on the screen traces back to the source that produced it.
Signal, context, perspectives, evidence, trade-offs, decision, authorization, outcome.
The same eight steps run whether the subject is a late shipment, a maintenance window or a quarter's pricing. Each step leaves a record.
- 01
Signal
Something happens in an authorized source: a delayed delivery, a financial deviation, a change in load, a parameter leaving its range.
Observe - 02
Context
Mantikon links the event to objects, dates, dependencies, obligations and the data it is permitted to read.
World model - 03
Perspectives
The council independently assesses the consequences for cost, risk, safety, schedule and people. No role holds two interests.
Council - 04
Evidence
Participants support claims with references to source data; the critic marks gaps, contradictions and repeats from one lineage.
Provenance - 05
Trade-offs
The system models not one best answer but several admissible alternatives, with their sensitivity to the assumptions behind them.
Scenarios - 06
Decision
The operator receives a decision brief with the recommendation, the dissent that was not resolved and the conditions under which the system declines to decide.
Decision Brief - 07
Authorization
An independent policy gate checks rights, limits and whether a human confirmation is required. In the first product every external change requires a person.
Policy gate - 08
Outcome
The actual change is linked back to the decision that caused it and updates the record of how good the reasoning was.
Ledger
The flow
The critical boundary runs between the decision engine and the executive gateway: the models form proposals, they never hold the key to execution.
- Data and eventsERP · CRM · MES · documents · sensors
- Mantikon Connectconnectors · streams · identity resolution · trust labels
- Evidence registry and provenancesource · time · hash · independence group
- Mantikon Worldobjects, relations and constraints, versioned in time
- Mantikon Council · Deliberatespecialized agents · typed messages · independent critique
- Mantikon Decide + monitoralternatives · scenarios · abstain · calibration · cost
- typed proposal, never a writethe boundary the models cannot cross
- Mantikon Guard · policy as codedeny · abstain · approval required
- Mantikon Control · executive gatewaysigned approval · dry run · idempotent · logged
- External system or APIthe organization's own
- Outcome events → Mantikon Ledgerwhat actually happened, against what was predicted
Eight named parts, one decision path.
This is the intended module structure. The maturity of each part is stated in the project status rather than implied by the name.
- Connect
Sources, streams, documents, events
Connect the data behind every operational decision — without assuming every record must be copied into one more warehouse.
Data plane - World
Objects, relations, change, constraints
A living model of how your organization works: what exists, how it depends on everything else, and what was true at which moment.
Data plane - Council
Agents with different interests and competencies
Let every critical perspective be heard: cost, schedule, safety, risk, people and law, each argued by a role that holds nothing else. Watch a case being argued
Cognitive - Deliberate
Evidence, debate, dissent
Challenge assumptions before they become actions. Typed messages carry their sources; an independent critic attacks the weakest link.
Cognitive - Decide
Alternatives, scenarios, trade-offs
Choose with reasons you can inspect. The Decision Room ranks admissible options and may decline to decide.
Cognitive - Control
Permissions, workflows, confirmations
Turn approved decisions into controlled work: typed commands, dry runs, idempotency, compensating actions.
Governance - Ledger
Memory of decisions and real outcomes
Remember why you chose — and what happened next. Decision Briefs stay linked to their evidence, approvals and results.
Governance - Guard
Rights, policies, audit and security
Accountability built into the decision path. Policy as code decides what is admissible, independently of any vote.
Governance
Decision
Brief
Every finished case produces the same reproducible structure. The screen is called the Decision Room, the records are Decision Briefs, the history is the Decision Ledger. A customer is not buying an AI opinion; they are buying a lower cost and a lower risk on one management decision, and the brief is where that is visible.
The console shows this file on live cases, with the age of every piece of evidence, the position of every role and the objections nobody answered.
| Field | Content | Why it matters |
|---|---|---|
| Problem | The question and its boundaries | Everyone argues the same task |
| Evidence | Sources, references, time, rights | It is clear where the conclusion came from |
| Alternatives | At least two relevant paths, or a reasoned refusal | The decision is not the model's first idea |
| Stakeholder views | Effects on schedule, cost, people, risk | The conflict of interests is visible |
| Critique and dissent | Counter-arguments and unresolved objections | False agreement is detectable |
| Scenarios | Ranges, assumptions, sensitivity | A forecast is not mistaken for a fact |
| Decision and confidence | Recommendation or refusal, and its limits | The conditionality of the output is explicit |
| Approval | The authorized person and the right to act | Responsibility does not disappear |
| Outcome | What actually happened | Learning becomes verifiable |
Three
planes
Data plane
Connectors and ingestion, batch and stream processing, normalization, object models, search, quality control, the provenance of every source. Mantikon is designed to integrate with the infrastructure a customer already has — a lake, a warehouse, SQL systems, BI and enterprise APIs — rather than requiring every dataset to be duplicated. The scale of "big data" is proved by a measured volume and rate on a test, not by the words "ingests everything".
Cognitive plane
The world model, specialized models, the deliberation protocol, the critic role, the scenario engine and the Decision Brief. Two senses of "agent" stay separate: an interest (cost, safety, schedule, people, risk) and a function (retrieval, verification, forecasting, arbitration). One agent may perform a function, but a conflict of interests is never hidden inside a single synthesized answer.
Action and governance plane
Admissible-action logic, approvals, separation of powers, an immutable event audit and correction from outcomes. The first commercial release is default-deny with human approval for every write. In industries with physical risk the executive loop is architecturally separated from the language models and validated by independent procedures.
The ability to propose is not the right to execute.
Mantikon supports several levels of engagement with real processes. Moving between them is a decision about risk, not a feature flag.
| Level | What the system may do | Condition |
|---|---|---|
| Analytical | Observes the organization, detects change, forecasts and explains. | Default on connection |
| Advisory | Develops alternatives and hands them to the responsible people. | The first product starts here |
| Operational | After a decision is approved, performs the permitted actions through enterprise interfaces. | Per-action human approval |
| Limited autonomy | Executes pre-defined low-risk operations without confirming each step, inside approved policies and technical limits. | Only after a separate security gate |
| Critical infrastructure | Actions that can affect physical safety, health or substantial public interests. | Separate admission, independent protective loops, mandatory engineering verification |
Design ambition, stated plainly: one protocol across many operational worlds. A pilot in one domain does not demonstrate fitness for a power grid or an orbital system; each domain brings its own validation. What the whole path is for is set out in Foresight: modelling the futures a decision would create, and searching for the one worth choosing.