Build

`trialdesignbench.build`

Materialize canonical dataset tasks as Harbor task directories.

Each task gets:

<tasks_dir>/<task_id>/
  task.toml          Harbor config (network allowlist placeholder)
  instruction.md     prompt template + question skeleton + source document
  environment/       empty (prebuilt image) or our Dockerfile context
  tests/
    Dockerfile       verifier image: shared image + /tests files
    test.sh          runs `tdb grade`
    rubrics.json     hidden grading spec

Network policy has two phases. [environment] is the baseline during agent setup (and the healthcheck); [agent] applies during agent.run(). Both allowed_hosts lists are written empty (deny all) with marker comments; tdb run fills them per agent and auth mode: the model API hosts for the agent phase, plus Harbor's install hosts for the setup baseline when the agent is not preinstalled in the image.

`build_task(record, rubrics, task_src, out, *, options, dataset_version, dataset_digest, template)`

Write one Harbor task directory.

`build_tasks(dataset_dir, out, *, options, task_ids=None, template_path=None)`

Build Harbor tasks for a dataset and write tdb-build.json.

`default_template()`

The packaged prompt template (copy of the intake system prompt).

`render_instruction(record, document, template)`

Agent instruction: template, sandbox note, question block, outputs, document.

`render_task_toml(record, *, options, dataset_version, dataset_digest, task_digest, template_sha256)`

Harbor task.toml for one task.

`render_test_sh(grader_source, version)`

Verifier test.sh for the chosen grader source.

`render_tests_dockerfile(options)`

Verifier image definition: shared image plus /tests files.