Human Oversight Registry
Which AI decisions run autonomously, and which require a human? This registry
is our human-in-the-loop framework — every class of decision sorted by
risk, with its automation level, what triggers human review, and the
escalation path. Low-risk work is autonomous; high-risk and irreversible
actions are gated on a human. Aligned with EU AI Act Article 14 (human
oversight) and the Montreal Declaration.
3 Low — Autonomous
3 Medium — Logged
4 High — Human Approval
3 Critical — Human-Initiated
Low — Autonomous
Notifications, reminders & briefings
Fully autonomous
Routine notifications, reminders and briefings
Agents send notifications, reminders, scheduled briefings and internal log entries without human approval. These are informational and carry no irreversible effect.
Human review
None — informational only.
Escalation path
Recorded in the activity log; no approval gate.
Knowledge-base reads & translation
Fully autonomous
Knowledge reads, search and translation
Reading the knowledge base, semantic search, and translating existing content are read-only or non-destructive and run autonomously.
Human review
None — read-only / non-destructive.
Escalation path
No gate; covered by routine audit logging.
Scheduled audits & embeddings
Fully autonomous
Scheduled audits, embeddings and health checks
Background jobs — compliance scans, embedding generation, fleet health checks, welfare digests — run on a schedule and only produce reports or internal state, never outward-facing changes.
Human review
None; findings surface to humans via reports.
Escalation path
Findings routed to the relevant director for review.
Medium — Logged
Agent task dispatch (LLM inference)
Autonomous, fully audited
Agent task dispatch and content generation
Routing work to agents and generating content (text, images, drafts) is autonomous, but every dispatch is logged with agent, model, capability, duration and outcome, and runs under C.R.E.E.D. welfare + governance checks.
Human review
Post-hoc: humans review the dispatch log; welfare gating can block work for an over-stressed agent.
Escalation path
task_execution_log + governance_audit_log; K.I.N. welfare gate.
Agent-proposed knowledge writes
Autonomous → quarantined → human-reviewed
Agent-proposed knowledge-base writes
When an agent proposes a new knowledge entry it is quarantined (source_type='agent_proposed', invisible to search) until C.O.D.E.X. promotes or rejects it. No agent-written knowledge becomes visible without that review.
Human review
Always — hourly C.O.D.E.X. review gate before an entry is visible.
Escalation path
agent_kb_review job → C.O.D.E.X. promote/reject.
Self-healing code proposals
Autonomous scan → human-approved deploy
Self-healing code proposals (Repair Bay)
The self-healing pipeline scans, diagnoses and drafts fix proposals autonomously, but a proposal is only ever a proposal — deployment to production stays a manual, human-approved step.
Human review
Always — a human approves every deploy.
Escalation path
Repair Bay review queue → human deploy.
High — Human Approval
Paid cloud model escalation
Blocked — human approval required
Escalation to paid cloud models
Escalating work to a paid external cloud model has a cost and data implication, so it is gated. (Platform-wide the escalation mode is currently 'local_only' — paid cloud is off by default.)
Human review
Always — a Telegram approve/deny prompt before any paid call.
Escalation path
autonomy_approval_queue → Telegram approval (Tier 2).
Creating or hiring an agent
Blocked — human approval required
Creating, hiring or reactivating an agent
Standing up a new agent (or hiring a contractor agent) changes the platform's workforce and cost profile, so it requires explicit human approval before creation.
Human review
Always — human approval before the agent is created.
Escalation path
propose_hire → autonomy_approval_queue → Telegram (Tier 2).
Cost-gated contractor dispatch
Blocked above threshold — human approval required
Cost-gated contractor / external dispatch
When the dispatcher picks a paid contractor agent and the estimated cost exceeds the auto-approve threshold, the work blocks on human approval before it runs.
Human review
When estimated cost exceeds the auto-approve threshold.
Escalation path
contractor_approval_requests → Telegram approve/deny.
Public website deployment
Human-triggered
Public website / content deployment
Publishing to a live public website replaces what visitors see, so a human triggers and verifies every deploy; agents may prepare content but do not auto-publish.
Human review
Always — a human triggers and verifies the deploy.
Escalation path
Manual /deploy-website by an authorised human.
Critical — Human-Initiated
Infrastructure, secrets & node changes
Human-initiated only
Infrastructure, secrets and node changes
Changes to servers, secrets, DNS, VPS nodes or the mesh are never agent-initiated. They are performed by the owner (or an explicitly authorised human) with the action logged.
Human review
Always — human-initiated and human-executed.
Escalation path
Owner action; recorded in the governance audit log.
Destructive data operations
Human-initiated only
Deletion or overwrite of data
Deleting or overwriting data is irreversible, so it is never automated. A human initiates it deliberately, and destructive steps are confirmed before they run.
Human review
Always — explicit human confirmation.
Escalation path
Owner / admin action with confirmation + audit log.
Billing & financial decisions
Human-initiated only
Billing, subscription and financial changes
Changes to billing, subscription tiers, payouts or financial transactions are made by the owner. Automated systems may surface information but never move money or change a paying customer's tier without human action.
Human review
Always — owner action.
Escalation path
Owner action; Stripe + billing audit trail.