API reference
The core modules are importable without the CLI dependencies and never import Harbor.
| Module | Purpose |
|---|---|
schema |
Versioned pydantic models for every artifact |
dataset |
Intake import, canonical dataset, validation |
build |
Canonical dataset to Harbor task directories |
environment |
Shared image pins, Dockerfile context, build/check commands |
judge |
Judge protocol, AnthropicJudge, FakeJudge |
grade |
Deterministic checks, rubric judging, grade outputs |
scoring |
Versioned scoring rules |
agents |
Supported agents: providers, credentials, hosts, closed-book settings |
run |
job.yaml generation, allowlists, provenance manifest |
canary |
Network canary task |
report |
Aggregation and leaderboards |
provenance |
Digests, versions, git and image identifiers |
trialdesignbench
TrialDesignBench: a thin evaluation framework for AI agents in clinical trial design.
Harbor runs agents; this package owns the task schema, task materialization, grading, scoring, aggregation, and provenance.