Documentation
Plugins & MCP Apps
API Reference
SDKReference
Glossary
Terms used across the architecture and API docs
Status: AVAILABLE. See status labels.
Short definitions. Follow the linked pages for detail.
Documentation Index
Fetch the complete documentation index at: /llms.txt
Use this file to discover all available pages before exploring further.
Terms used across the architecture and API docs
| Term | Meaning |
|---|---|
| Agent ISA | The instruction set of seven groups (observation, decision, mutation, execution, version control, verification, control) that graphs are made of. |
| ABI package | The frozen set of kernel schemas, registry and examples, version 1.0.0. |
| Action receipt | A durable record that an effect happened. Required before retrying an external effect. |
| Attention | A request for a person to answer or approve. |
| Cassette | A recording of a run’s external results for replay. Preview (internal). |
| Compute graph | The only workflow type; nodes are primitives and capabilities. |
| Context projection | The compiled set of chunks one component sees for one step. |
| Decision primitive | A typed question over candidates answered by a decision backend. |
| Divergence | A difference between a replay and its cassette. |
| Effect class | The declared kind of change a tool can make, from NONE to PERMISSION_CHANGE. |
| ExecutionEnvironment | The resolved working directory, environment allowlist and scope a node runs in. |
| Fencing token | A value that stops a former lease holder from writing. |
| Gate | Allternit’s policy check in front of every executor. |
| Idempotency key | A caller-supplied key that makes a repeated request safe. |
| Lease | Time-bound permission for one holder to run a node. |
| Primitive | One named operation in the registry. |
| Receipt | A signed, append-only, content-addressed record of an action. |
| S0 to S3 | The cognition levels, from deterministic code (S0) to the strongest model (S3). See cognition levels. |
| Shadow | A mode where a decision backend records what it would decide and takes no action. |
| System lift | The measured improvement of the system over the same model run without it. Design. |
| Trust class | A label of how far content is believed, including UNTRUSTED. |
| Verifier-owned completion | Completion decided by the system from verifier evidence, never by a model’s statement. |
| Part | Status |
|---|---|
| Glossary | AVAILABLE |