Concepts

Harness

A portable specification of how intelligence performs a class of work. It names required context providers, capabilities, policies, verifiers, protocol instructions, artifact expectations, and execution limits.

Harness IR

The normalized, content-addressed representation that the runtime executes. Source syntax is not runtime semantics.

Objective

The requested external state or outcome. An objective is data, not a system prompt.

Context Provider

Resolves a typed slice of the world for the current run step. Context providers do not take actions.

Capability

A boundary through which the harness can affect or query the world. Capability invocation is an explicit action and therefore passes through policy.

Policy

Evaluates a proposed action and returns allow, deny, or require_approval. Multiple policy results combine conservatively: deny > require_approval > allow.

Protocol

Domain or method guidance describing how work should be performed. In v0.1 a protocol is instructions, not a deterministic workflow DSL.

Verifier

Independently evaluates whether produced artifacts satisfy an objective. A verifier returns pass, fail, uncertain, or requires_human plus structured evidence.

Artifact

A durable output of work: patch, memo, report, dataset, protocol, file, etc. Artifacts are content-addressed using a digest.

Evidence

Structured support for a verification or claim. Evidence should reference durable sources or artifacts when possible.

Run

One execution of a compiled harness against one objective. Runs produce events, artifacts, verification results, and an outcome.