Harbor

Task layout

Create this package before any Harbor trial:

The files have these contracts:

• Task.md is the human-reviewed control-plane spec. Keep the exact request, initial state, hidden truth, complete scoring rules, fairness analysis, and open decisions here. Never mount or copy it into the agent image or workspace.
• task.toml follows the installed project-supported Harbor schema. Record the agent and Verifier timeouts, Environment resources and services, network policy, and runtime variable names needed by this Task. Never store secret values in it.
• instruction.md contains the exact approved agent input. State the goal, agent-visible context, required output or effect, paths, and real limits. Do not mention hidden tests, scoring, expected answers, or the solution.
• environment/ contains one usable Environment definition: a Dockerfile, or a docker-compose.yaml whose primary service is main. Include only the dependencies and seed state visible to the agent. Never copy Task.md, tests/, solution/, hidden truth, judge rules, or credentials into it.
• tests/test.sh always exists. Harbor uploads tests/ only after agent work ends and runs this file as the Verifier entry point. It may call optional helper tests and hidden fixtures. It must read independent evidence and write a valid reward to /logs/verifier/reward.txt or /logs/verifier/reward.json on every completed Verifier path. It must not exit before writing the reward, turn an infrastructure failure into a zero, or expose hidden truth or secrets in output.
• solution/solve.sh is optional. When present, it performs the real reference work against the same Environment and proves the intended result is reachable. It does not write a hard-coded answer only to satisfy the Verifier.

Use optional Verifier helpers, fixtures, or a solution only when the Task needs them. Do not create empty placeholder files or directories.

A Harness adapter can bind approved dependencies and translate I/O. It must not decide the answer or fabricate actions.

Keep generated jobs outside task source, for example under evals/jobs/. Keep them until the user accepts, revises, or drops the eval. Before retaining private transcripts, define who can read them, what must be redacted, how long to keep them, and how to delete them. Keep only the evidence needed to audit the Task.

Audit package completeness before running

Use harbor --help and subcommand help to confirm current flags. Resolve the configuration without a scored run when the installed version supports it. Do not start any Harbor trial, including an Oracle or model trial, until this audit passes for the exact Task package:

1. Confirm task.toml, instruction.md, tests/test.sh, and a usable Environment definition all exist at the paths Harbor will load. Confirm Task.md exists beside them for human review.
2. Parse task.toml with the supported Harbor version. Confirm exact Task selection, Harness or adapter, model, resources, network policy, runtime variable names, trial count, concurrency, timeouts, judge, and output paths.
3. Build and start the Environment through Harbor. Confirm mounts, the main service when Compose is used, readiness, agent-visible seed state, allowed operations, cleanup, and trial isolation.
4. Inspect the built agent image and workspace. Confirm they do not contain Task.md, tests/, solution/, hidden evidence, expected answers, judge rules, or secret material.
5. Confirm Harbor can upload and execute tests/test.sh with the required file permissions and working paths. Exercise each Verifier dependency.
6. Run known-good, alternative-valid, wrong, shortcut, collateral-change, and missing-or-corrupt-evidence cases through the real Verifier command. Confirm each completed path produces a parseable reward and criterion evidence.
7. Run solution/solve.sh through Harbor when it exists, then run the Verifier. Confirm the reference result is reachable and receives the intended reward.
8. Fix every missing file, invalid path, parse error, build error, startup error, missing dependency, permission error, missing reward, or leaked file before starting a trial.

Docker is the default boundary, not proof that isolation works. Test allowed and denied access when network policy matters. Report a limit if the backend cannot enforce or expose it. Never put secret values in source, images, prompts, fixtures, or logs.

Run contract

Before model trials, run Environment checks, the reference path, and focused Verifier fixtures through the same images and commands Harbor will use. Increase timeouts only when evidence shows valid work exceeds the current limit. Keep a finite bound and record the reason.

For each trial, retain:

• resolved configuration;
• Harness messages, calls, results, retries, and errors;
• Environment startup, requests, state, reset, and cleanup evidence;
• Verifier criteria, evidence, verdict, reason, reward, and errors; and
• phase timing and termination reason.

Wrong agent work receives zero. Build, adapter, credential, reset, timeout, judge, Verifier, or cleanup failure receives no agent score. Every attempted trial must end as completed, cancelled, or infrastructure error. Do not use a pending trial as evidence.

For multi-turn runs, also prove that the first Harness input equals instruction.md, later turns came from the declared user policy, one approved session was reused, future messages were not preloaded, and no model call occurred after termination. See the multi-turn reference for implementation details.