How the factory behaves

Autonomy has a perimeter.

Value classes

Process should scale with what can be lost.

A throwaway experiment should not carry commercial process. Customer money, private data, valuable work and physical safety change the required evidence.

Class 00 · ThrowawayExplore freely; discard cheaply.
Class 01 · ExperimentPreserve learning; add useful checks.
Class 02 · ProjectProtect source, repeatability and records.
Class 03 · ProductProve users, access and production quality.
Class 04 · CommercialControl money, promises and rollback.
Class 05 · TreasuredNarrow authority; prove recovery and safety.

Controls increase with value, irreversibility and consequence. Classification is a working judgement, not an automatic risk score.

Authority

Observe more than you mutate.

Agents can collect production evidence and propose a change. A consequential action needs a prepared diff, rationale, qualification and rollback path. Human approval is a durable system state, not a chat prompt an agent can ignore.

Observe

Read permitted state and source material.

Recommend

Explain what changed and what remains uncertain.

Prepare

Produce a reviewable change with evidence and rollback.

Execute by policy

Bounded sandbox work may proceed; financial, destructive and physical decisions keep a human gate.

Failures teach the factory.

Contain the failure. Record it. Investigate why the controls missed it. Fix the product, then decide whether the failure belongs in a regression check, release gate, policy, architecture or documentation.

A recurring failure should become harder to repeat because the factory learned from it. Public claims remain smaller than the evidence supporting them.

A good factory is not maximally autonomous. A good factory knows where autonomy is appropriate.Core doctrine