Outputs and Evidence

The output directory is declared by runtime_info.output.harbor_tasks_dir.path, normally artifacts/tasks/.

That path is a symlink to the newest immutable publication under artifacts/tasks-versions/<batch-id>/. A passing batch is copied into a new version, then the output symlink switches atomically. Tracer resolves the version path when staging; later publications leave its existing inputs unchanged. Retain versions still referenced by downstream jobs.

artifacts/tasks/
├── verifiable_tasks.txt
├── registry.json
└── <task-id>/
    ├── instruction.md
    ├── task.toml
    ├── environment/Dockerfile
    ├── solution/solve.sh
    ├── tests/test.sh
    └── tests/test_outputs.py

The manifest is a newline-delimited list of task IDs that passed this block's fresh Oracle/NOP gate. It is an allowlist, not a progress log. Only those IDs may be staged into Tracer. A generated candidate that has not passed both trials is never eligible for downstream inference.

runtime_info.output.eval_registry.path names the generated registry.json. Its terminal-curator@1.0 entry lists the same accepted IDs with absolute paths into this immutable batch. Evaluator loads those local paths through its standard Harbor registry flow; no task upload or registry service is needed. The registry and manifest switch together with the publication symlink. Keep referenced versions available on the evaluation host. See Evaluator hand-off for configuration.

Each task describes a terminal problem, its reproducible environment, the reference solution, and a verifier that writes a Harbor reward. Schema normalization targets Tracer's pinned Harbor version. The adapter preserves task provenance and validation evidence without putting API credentials in task metadata or instructions.

The canonical verifier runs tests/test_outputs.py through pytest and records JUnit evidence. Validation rejects zero collected assertions, collection/runtime errors, all-skipped suites, a failing Oracle, and a NOP that passes. The local candidate import mode therefore expects Terminal-Lego's pytest layout rather than an arbitrary Harbor verifier.

Runtime acceptance is followed by the operating agent's semantic audit of the instruction, assertions, and reference behavior before downstream use. Keep its requirement/test mapping and correction provenance with the task. Any repaired candidate needs both new fresh Oracle/NOP trials; see the workflow for the audit and correction procedure.

Progress and failures

artifacts/state.json is the live phase/count snapshot. At run completion or failure, artifacts/reports/latest.json contains the machine-readable report. The report records counts, per-candidate outcomes, and links to validation evidence. The dashboard reads an allowlisted view of live state and the published manifest. Inspect the actual JSON rather than assuming all discovered directories passed.

Intermediate source data, generated candidates, and Harbor validation jobs remain under artifacts/ outside the published task pool. These are useful for diagnosing rejected tasks and service failures. They are local runtime materials and are ignored by Git.

Run archives

start.sh installs an EXIT trap so success, errors, and interruption produce an archive and an entry in artifacts/index.yaml:

artifacts/archives/run_NNN/
├── metadata.yaml
├── config.yaml
└── scripts/

The archive records the effective configuration, with secrets redacted, including an overlay passed through --config. Metadata records exit state and repository SHAs. It does not copy entire vendored repository trees. A fresh checkout has no index; the first archived run creates it.

An archive entry does not establish a successful task or rollout by itself. Use its report, manifest, Oracle/NOP results, and separately generated Tracer trajectory to make those claims. Review artifacts before sharing them and keep raw credentials, private infrastructure details, and generated datasets out of a code PR.