superconscious

Superconscious is the reference governed cognition loop for recursive agents across the SocioProphet / SourceOS stack.

It coordinates task trees, safe operational traces, skill activation, tool use, memory decisions, model routing, policy admission, approvals, benchmarks, replay plans, and AgentPlane-compatible evidence without becoming the authority for schemas, execution, runtime placement, model governance, or workspace topology.

Position

Superconscious is the visible cognition/control-loop layer. It makes the estate behave like one governed agent operating system while preserving clean authority boundaries across the existing repositories.

Task input
  -> validate
  -> plan
  -> request policy admission
  -> request model route
  -> activate skill
  -> call tool adapter
  -> record observation
  -> decide memory handling
  -> request approval when needed
  -> emit safe operational trace
  -> emit AgentPlane evidence
  -> emit replay plan
  -> run benchmark assertions

Trust Surface Protocol

Superconscious now seeds the SourceOS Trust Surface Protocol for the estate.

The principle is direct: no invisible authority. Any repo that starts a process, installs a service, opens a socket, controls a browser or terminal, launches a container, stores credentials, runs agents, or routes model/provider traffic must declare that authority in TRUST_SURFACE.yaml.

This repo includes:

TRUST_SURFACE.yaml
schemas/trust-surface.schema.json
docs/trust-surface-protocol.md
examples/TRUST_SURFACE.node-commander.yaml
scripts/validate-trust-surface.py
.github/workflows/trust-surface.yml

The first target is not bureaucracy. The target is inspectability, purgeability, and provable cleanup across local agent runtimes.

SVF validation-history consumer

Superconscious may consume Sovereign Validation Fabric validation history as memory input for recursive agency planning. Subconscious is part of Superconscious in this lane: it may optimize inference over validation history, failure patterns, missing observations, and plan usefulness, but it is not a separate authority plane.

This consumer is read-only. It may remember validation debt, recommend observed validation before autonomous continuation, bias planning toward report-only behavior, and route summaries to AgentPlane evidence or Sociosphere backlog records.

It must not execute SVF Actions, run Sociosphere commands, issue or verify receipts, promote advisory validation to blocking validation, mutate policy, override guardrails, or grant agent autonomy from validation history alone.

Relevant files:

docs/SVF_VALIDATION_HISTORY_CONSUMER.md
examples/svf-validation-history-event.selected-missing-observation.json
scripts/validate-svf-validation-history.py

Validate locally:

make validate-svf-validation-history

What this repo owns

• Reference recursive reasoning loop implementation.
• Safe operational trace assembly.
• Local deterministic demo runner.
• Adapter interfaces for AgentPlane, Agent Machine, Model Router, Guardrail Fabric, Agent Registry, Memory Mesh, SocioSphere, sourceosctl, BearBrowser, TurtleTerm, and Socios.
• Example reasoning runs and benchmark fixtures.
• Product-facing cognition-console semantics.
• Trust-surface protocol seed, examples, and validation workflow until canonical schema ownership moves to SourceOS-Linux/sourceos-spec.
• Read-only SVF validation-history memory consumer fixtures and validation.

What this repo does not own

• Canonical schemas: owned by SourceOS-Linux/sourceos-spec.
• Execution, placement, evidence, replay authority: owned by SocioProphet/agentplane.
• Workspace topology, manifests, locks, registry governance: owned by SocioProphet/sociosphere.
• Runtime substrate and AgentPod activation: owned by SourceOS-Linux/agent-machine.
• Local model profile carriage: owned by SourceOS-Linux/sourceos-model-carry.
• Model promotion, consent, and personalization authority: owned by SocioProphet/model-governance-ledger and SociOS-Linux/socios.
• Policy authority: owned by Guardrail / Policy Fabric.
• Agent identity and grants: owned by Agent Registry.
• SVF authority, validation execution, receipt issuance, and receipt certification.

M1 deliverable

M1 is a deterministic, no-network, no-model-call, no-side-effect reference loop that emits:

.runs/<run-id>/events.jsonl
.runs/<run-id>/reasoning-run.json
.runs/<run-id>/agentplane-evidence.json
.runs/<run-id>/replay-plan.json
.runs/<run-id>/benchmark-result.json

The first goal is not model quality. The first goal is lifecycle discipline, evidence compatibility, replay shape, safe trace semantics, and declared trust surfaces.

Repository layout

superconscious/
  README.md
  ARCHITECTURE.md
  THREAT_MODEL.md
  ROADMAP.md
  AGENTS.md
  TRUST_SURFACE.yaml
  docs/
    trust-surface-protocol.md
    SVF_VALIDATION_HISTORY_CONSUMER.md
  schemas/
    trust-surface.schema.json
  examples/
    TRUST_SURFACE.node-commander.yaml
    svf-validation-history-event.selected-missing-observation.json
  scripts/
    validate-trust-surface.py
    validate-svf-validation-history.py
  packages/superconscious-core/
  tests/

Strategic framing

Subconscious optimizes recursive inference.
Superconscious governs recursive agency.

Superconscious does not claim machine sentience. It defines operational consciousness as explicit awareness of task state, tools, memory, models, policy, runtime, evidence, and feedback loops.

The trust-surface layer extends that frame to local authority: a governed agent operating system must know what it can start, read, write, execute, expose, update, remember, and remove.