Skip to content

Verifiable tasks

A task is an instruction, an environment, and a test. It has passed when the test passes.

Instruction, environment, test

A task is a directory: instruction.md, task.toml, environment/, tests/, and a reference solution/. Nothing implicit, everything reviewable in a PR.

The verifier produces the reward

tests/test.sh runs after the agent and writes /logs/verifier/reward.txt (1/0) or reward.json. The number comes from code.

A trial is reproducible

Each attempt is a trial: start environment, run agent, capture trajectory, verify, emit reward. The same task should agree across machines and providers.

Aligned with Harbor Framework

Everything here is derived from the public Harbor Framework repository and docs.

An English and Chinese field guide to Harbor Framework, focused on one thing: verifiable tasks.

Official entry points:

Line Question Output Owner
Verify Did the task get done? reward (0/1 or metrics) this site / Harbor
Audit What happened, provably? evidence chain, trajectory trajectory audit
Score How good was it? ordinal number trajectory scoring