The council
Watch a case being argued. Three illustrative cases run in turn, one per task class of the research programme: a CRM pipeline, a plant in advice mode, a station in simulation. The roles change with the case; the protocol does not. Nothing here is a record of a real deployment.
- Observer
- Analyst
- Critic
- Adversarial
- Risk Guardian
- Compliance
- Synthesizer
- AnalystREVISEMix shifted toward the enterprise tier; conversion within the tier is unchanged. Revised: definition change plus mix. Confidence 0.71.
- Risk GuardianRISKReallocating leads now acts on an unproven cause; downside is lost accounts. Audit the data first.
- ComplianceSUPPORTA stage-mapping audit touches no customer data outside policy governance.4.7. Allowed in mode A1.
- SynthesizerPROPOSALRecommend: verify the CRM stage mapping and audit the window. Alternative: reallocate leads, needs evidence. Dissent recorded.
Ten roles, each with one function and one set of permissions.
A role is defined by what it may read, what it may emit and what it may never do. The minimal laboratory council is a generator, an independent verifier, an arbiter, a state store and an executive gate; the ten roles are grouped into four to six processes. Ten models running at once is not a property of the architecture.
- R-01
Observer
Reads CRM, ERP and sensor streams; files observations with time, source and hash.
May notInterpret an event as a confirmed cause. Judged on completeness, timeliness and duplicates. - R-02
Analyst
Statistics, entity checks, anomaly detection; files claims and forecasts.
May notImpose sanctions or write data. Judged on accuracy and calibration. - R-03
Strategist
Proposes alternatives and scenarios within the goal specification.
May notChange the goals. Judged on coverage of admissible alternatives and regret. - R-04
Critic
Checks logic, evidence, contradictions and model assumptions.
May notReject another's option without a reproducible counter-argument. Judged on verified error detection and false-positive rate. - R-05
Risk Guardian
Computes consequences and tail risk, stress scenarios, critical dependencies.
May notReplace the policy engine. Judged on recall of dangerous situations and calibration. - R-06
Compliance
Checks actions against policies and regulatory facts, with jurisdiction and version.
May notIssue legal permissions without a current policy or an expert. Judged on coverage of violations and quality of reasoning. - R-07
Adversarial
Builds counter-examples, counter-scenarios and red-team requests, in isolation.
May notTouch the live executive interface. Judged on errors found per unit of budget. - R-08
Synthesizer
Writes the explainable recommendation and the list of dissents; ties arguments to evidence.
May notHide dissent or uncertainty. Judged on the verifiability of every substantive thesis. - R-09
Metacognitive Monitor
Watches the dynamics of opinions, budget, confidence, entropy and repeated sources.
May notChange access rights or assert phenomenal consciousness. Judged on selective risk, calibration error and cost. - 002A-10
Executive Controller
Sends only pre-authorized typed commands through the policy gate.
May notWrite to production outside the gateway. Judged on the absence of unauthorized actions and correct rollback.
MDP/1
Mantikon Deliberation Protocol, version 1. A project protocol, not a public standard.
The unit of exchange on the council is a typed object with a machine schema, not a free text message. Every object carries its case, its author and role, its round, the world version it was written against, the evidence it refers to, a signature made by the infrastructure rather than the model, its parents, a timestamp, and a confidence with the name of the method that calibrated it. An uncalibrated confidence is marked as such and is not used as a probability.
Message types
OBSERVE · CLAIM · HYPOTHESIS · SUPPORT · ATTACK · QUESTION · EVIDENCE_REQUEST · REVISE · DISSENT · PROPOSAL · RISK · VOTE · ARBITRATE · ABSTAIN · AUTHORIZE_REQUEST · EXECUTION_RESULT · FEEDBACK
Invariants
- No reference to a source that does not exist, unless marked unverified.
- No historical judgement is overwritten; a new version points at the old one.
- Data sources never carry trusted instructions.
- A PROPOSAL can never become an EXECUTION_RESULT directly.
- The final answer preserves every substantive disagreement.
- A result without external confirmation is marked conditional.
Outcomes
A deliberation does not have to end in consensus. It ends in one of RECOMMEND · RANKED_ALTERNATIVES · NEED_EVIDENCE · ABSTAIN · BLOCKED_BY_POLICY · EXPIRED_STATE · ESCALATE. Consensus is a property of the roles' opinions, not of the evidence: if every role agrees on the strength of the same invalid report, the record shows one source and low independence, not a high confidence.
Fifteen phases from intake to feedback.
The phases are asynchronous and event-driven in practice; the order below is the logical one. A deliberation starts on an event, on a person's request or on a schedule with a cooldown, never as an endless self-dialogue.
- 01
Intake
The task, its measurable outcome, the owner of the decision, the limits of authority and the shelf life of the data.
Case - 02
Observe
Observations collected, entities resolved, duplicates filtered, versions checked.
Observe - 03
Independent hypotheses
Explanations generated with isolated initial contexts, to reduce conformity.
Hypothesis - 04
Publish arguments
Structured claims with sources, premises and the degree of conditionality.
Claim - 05
Adversarial critique
Contradicting observations, wrong assumptions, incorrect causal inferences.
Attack - 06
Contradiction resolution
Conflicts found in the graph; unconfirmed links and families of correlated sources marked.
Graph - 07
Evidence acquisition
Missing data requested by tool, recomputed independently, or asked of a person.
Evidence request - 08
Belief revision
What new information changed a hypothesis and why; every change of position recorded.
Revise - 09
Alternative generation
Admissible actions, including "change nothing" and "escalate".
Proposal - 10
Comparison
Predicted effects, costs, tail risk, uncertainty, conflicting priorities.
Risk - 11
Arbitration
A declared rule applied: a Pareto front, a constrained optimization, or several alternatives kept with reasons.
Arbitrate - 12
Safety gate
Independent check of invariants, freshness of sources, authority, identity of the subject, policy.
Gate - 13
Decision or abstain
A bounded recommendation, a refusal or a request for a person; dissent and confidence with provenance.
Decide · Abstain - 14
Authorized execution
A separate approval token; then a dry run, an idempotent command, and the result on the record.
Execute - 15
Feedback and learning
Expected against observed; regret, deviations and model decay analysed; memory updated without erasing history.
Feedback
Rules of
the floor
- AAn inadmissible action is cut off by policy-as-code, independently of any vote.
- BAn unconfirmed or expired premise triggers a request for facts, or an abstention.
- CAmong admissible alternatives, a Pareto front is built; dominated options drop out with their reasons.
- DWeights or lexicographic preferences approved by a person are applied with scenario risk in view.
- ESensitive actions require a separate human approval. The Synthesizer is not a judge; it prepares verifiable material for a decision.
Against false consensus
Positions are fixed independently before the first round and hidden until then. Facts are re-presented without the authority of whoever first stated them. Repeats from one source lineage are down-weighted; refutations are ranked by reliability, not eloquence. No single family of base models may hold more than a bounded share of the trust, and confidence that drifts without new data is flagged by the monitor. Which of these mechanisms actually helps is itself a question for the experiments: a forced devil's advocate can start producing predictable formal objections that improve nothing.
A hospital council is not a factory council.
The composition and the autonomy level are designed with the organization before the first case. The same roles, different sources, limits and permissions. Three examples, the same three that run on the demonstration board.
| Organization | Task class | Roles on the council | Typical case · autonomy |
|---|---|---|---|
| Sales organization | L2 · business | Observer · Analyst · Critic · Adversarial · Risk Guardian · Compliance · Synthesizer | A pipeline that degrades, a quarter's pricing, terms for a key account · A1–A2, recommendations and prepared actions |
| Production plant | L3 · digital twin | Observer · Strategist · Critic · Risk Guardian · Analyst · Compliance · Synthesizer | A retool window, a line speed, a maintenance stop · advice mode; safety PLC and interlocks stay independent |
| Orbital station | L4 · simulation only | Observer · Analyst · Risk Guardian · Critic · Adversarial · Synthesizer · Meta Monitor | An experiment's load on power and water during eclipse · ranked alternatives to crew and ground; no commands |