forme / System architecture

Cognition, proactivity, and action governance share one runtime.

forme contains three decision kernels and one harness runtime. Cognitive evolution updates long-term understanding. Coordination forms action constraints. Proactivity decides when to intervene. The harness applies authority, policy, verification, and evidence requirements to every real action.

Runtime contractforme
Experience
Cognition
Action intent
Verified outcome

One subject, one event chain, one governed exit

Decision kernels

Three kernels answer how to understand, how to coordinate, and when to intervene.

The kernels share memory, evidence, and runtime state, but hold no execution privilege. They produce cognitive candidates, action constraints, or intervention proposals. Every ActionIntent returns to the harness.

forme kernel topologySingle cognitive subject
formeJUDGMENT
Harness runtime
Run / Session / TurnPolicy / ApprovalExecution / RecoveryVerification / Event Store
Cognitive Evolution Kernel

How experience changes future judgment

Which experiences are strong enough to enter stable cognition?

Events, outcomes, and feedback enter the memory graph. Reflection, provenance, conflict, and evidence checks form cognitive candidates. Governed promotion can place them in the user, self, partnership, or cognitive model.

RESPONSIBILITY

Organize long-term evidence into cognitive structures that inform later judgment.

STATE TRANSITIONEvent -> cognitive candidate -> stable model
INPUT
Events and outcomesMemory graphOwner feedback
OUTPUT
UserModelAgentSelfModelPartnershipModelCognitiveMap
Submit candidate updates only. No direct execution or authority expansion.

Action formation

The decision trace remains reviewable from observation through verification.

DecisionTrace records the goal, situation, evidence, constraints, resource choices, authority, and outcome. Hidden model reasoning is not treated as a system fact.

Observation / Context

Facts enter context

Goals, runtime state, prior evidence, and environmental change enter the current context.

Governed evolution

New interpretations enter a candidate layer first.

Provenance, evidence, conflict, impact, confidence, and counterevidence determine whether a candidate can be promoted. Stable cognition supports decay, retraction, and rollback. A single outcome cannot overwrite the long-term structure.

Event / Verification

What happened

Events, actions, outcomes, and feedback become traceable facts.

Governance boundaries

User control, execution policy, event audit, verification, and rollback form stable boundaries.

User controlExecution policyEvent auditVerification and rollback

Evolving strategies

Resource composition, action patterns, intervention style, and strategy versions evolve through evidence.

Resource compositionAction patternsIntervention styleStrategy versions

Capability layer

The capability layer supplies execution. The coordination kernel decides the composition.

Models, tools, skills, plugins, MCP, subagents, and local or remote executors are replaceable capability modules. The coordination kernel produces a ResourcePlan from the goal and constraints. Real invocation still passes through the harness.

Reasoning and collaboration03
ModelsSubagentsAction patterns
Tools and environments07
ToolsSkillsPluginsMCPBrowserComputerShell
Execution placement04
Local executorRemote executorSandboxApp API

Runtime governance

The harness is the only exit for real action.

The harness binds each action cycle with Run, Session, and Turn. It rechecks authority and policy at the exit, then handles approval, execution, verification, recovery, and event evidence.

Runtime state

Run, Session, and Turn support waiting, recovery, cancellation, and handoff.

Run / Session / Turn

Action exit

Policy, authority, competence, plan, and approval are rechecked before real execution.

Policy / Approval / AutonomyEnvelope

Outcome evidence

Results are verified against DoneContract, retaining failure and recovery evidence.

Verification / FailureEvidence

Reconstructable facts

Inputs, proposals, actions, and outcomes are stored as events for replay, correction, and audit.

EventStore / Replay
Proactive exit conditionValueGate ∧ CompetenceGate ∧ Policy / AutonomyEnvelope

forme / architecture

Cognition may evolve. Governance boundaries remain stable.