Concepts
The mental model behind Numinance - what an autonomous engineer is, how authority is separated from intelligence, and what you can check afterwards.
Give it context. Let it engineer. Your policy decides, not the model, and the evidence lets you check the result.
Numinance treats intelligence as one component of an engineering system. Context, authority, execution, verification and evidence are separate boundaries, each with a contract you can read. These ten pages build that model in order. Read them once, then use the architecture, configuration and protocol sections as reference.
The model in one pass
objective -> context -> understand -> plan -> act -> verify -> evidence
| |
policy decides before every privileged action| Idea | One-line claim | Page |
|---|---|---|
| Autonomous engineer | Work is a bounded unit with an objective and an outcome, not a chat session. | Autonomous engineer |
| Context first | Context is retrieved under authorization and produced, not hand-authored. | Context first |
| Change lifecycle | Every change moves through the same durable stages. | Change lifecycle |
| Autonomy levels | Autonomy is a ladder you configure, capped by floors. | Autonomy levels |
| Governed execution | Policy, grants and identities sit outside the model. | Governed execution |
| Harnesses vs runtime | The harness owns intelligence. Numinance owns authority. | Harnesses vs runtime |
| Reliability and recovery | Repair is bounded, checks are independent, rollback is planned. | Reliability and recovery |
| Evidence and trust | Don't trust the agent. Check the evidence. | Evidence and trust |
| Open vs proprietary | The brain is private. The contract is open. | Open vs proprietary |
The invariants
These hold across every page. If a page seems to weaken one, treat it as a documentation bug.
- Policy is evaluated outside the model before every privileged action.
- Authority is expressed as capability-based grants, bounded by scope, risk, environment and expiry.
- Four identities are distinct: author-agent, reviewer-agent, release-pipeline and human approver. No identity approves its own work, and only the release-pipeline deploys.
- Evidence is append-only, written by the runner identity, and denials are recorded too.
- Protected paths cannot be changed by an agent.
- Extensions are signed, digest-pinned and declare their permissions.
- A customer-local circuit breaker can stop or downgrade the system.
- Orchestration payloads carry references and digests, never content.
- Break-glass is human-only.
Designed versus available
Most of the runtime behavior in this section is designed and lives in the proprietary engine. What you can run today is the open numi CLI for schema validation and evidence verification. Check the status page before relying on any capability.
Where to go next
Status and supported surface
A source-backed matrix of what the public Numinance tree ships, what is designed, and what is deliberately not claimed. Dated 2026-11-09, spec revision 3c74d59.
The autonomous engineer
What separates an autonomous engineer from a coding agent - bounded work units, an objective-to-outcome contract, and explicit limits on what is built.