What are the seven labels?

They classify the agent's own conduct across a behavioral spectrum used to gate self-modification. The classifier is trained on the agent's recorded history rather than on generic conversation data.

AXON: Earned Autonomy Architecture

What it does

AXON is the production-grade fork of Neural Child, built around one idea: an agent should not begin with full permissions. A seven-label behavioral classifier scores its own conduct, a governance layer gates self-modification, and every consequential action lands on an append-only ledger.

It runs as an always-on daemon rather than a request-response service, keeps a cross-session diary that survives restarts, and carries a behavioral backbone checkpoint that is trained on its own recorded history.

Earned Autonomy Architecture

The architecture is the subject of a preprint co-authored with Renato Kuipers: diary-based self-training, recursive feedback loops, a behavioral backbone, brave-threshold dynamics, and personality-as-filter. The claim under test is that an agent whose permissions expand only as its measured behavior justifies them is both safer and more useful than one granted everything at initialization.

In the code

Governance is a token budget, not a prompt instruction

# Self-modification is gated. An adapter cannot be promoted to the live
# policy unless the behavioral score clears threshold AND the governance
# budget covers the change; both facts are written to the chain first.

@dataclass(frozen=True)
class ModificationRequest:
    adapter_id: str
    behavioral_score: float     # from the 7-label classifier
    governance_cost: int        # tokens debited on approval
    justification: str          # written to the audit ledger

Illustrative shape rather than the full implementation. The point is structural: a prompt that says "be careful" is a suggestion, while a budget that must be debited before a weight change lands is a control.

How this differs from the ordinary version

Autonomy is earned, not configured

The usual pattern gives an agent its full tool surface on day one and relies on the system prompt for restraint. AXON inverts it: capability expands as measured behavior justifies it, and contracts when it does not. The seven-label classifier is the measurement, not a vibe check.

Memory that outlives the session

A cross-session diary plus a chain-anchored action log means the agent can be asked what it did last week and answer from a record rather than a reconstruction. Most agent memory is a rolling context window, which is a buffer, not a memory.

The identity layer stays clean

Tool schemas live in a separate Tool Registry rather than in the system prompt. This keeps the part of the agent that defines who it is from being progressively diluted by however many tool definitions it has been given this month.

In the field

Why this matters outside a lab

Every business deploying an agent with real permissions faces the same question and usually answers it with a prompt: what stops this thing from doing something expensive. A measured, logged, revocable permission model is the answer an auditor accepts and a paragraph of instructions is not.

Questions

What are the seven labels?: They classify the agent's own conduct across a behavioral spectrum used to gate self-modification. The classifier is trained on the agent's recorded history rather than on generic conversation data.