laborix is a fully offline orchestrator for autonomous experiment pipelines, in the spirit of Agent-Laboratory-style research automation. It implements and verifies the mechanics of automated research programs — versioned plan schemas, dependency-safe deterministic execution, tamper-evident provenance ledgers, budget-respecting scheduling under uncertainty, ablation bookkeeping, and report synthesis — rather than the creativity of any particular language model.

At runtime the package depends only on NumPy. PyTorch (CPU-only) is an optional extra powering the trainable surrogate-scheduling demos. Every default test and example runs without network access; "experiment steps" in the demos execute real NumPy computations on synthetic data.

python -m venv .venv
.venv/bin/pip install -e ".[dev]"          # development env (hatchling, pytest, ruff, mypy)
.venv/bin/pip install -e ".[dev,torch]" \
    --extra-index-url https://download.pytorch.org/whl/cpu   # for surrogate-scheduling demos

.venv/bin/laborix demo demo-output

The demo executes a complete synthetic research program offline — plan → schedule → run → ablate → report — and prints measured regret against the oracle, ablation coverage, and the artifact ledger summary (for example {"coverage": {"method": true, "seed": true}, "regret": 4.5, "runs": 24, "synthetic": true}). Everything it produces is reproducible from the printed seeds and config hashes.

The laborix command-line tool provides one subcommand per lifecycle stage:

Exit codes are documented in docs/exit-codes.md.

The test suite (837 fast tests plus 4 slow integration tests) pins the properties that make an automated pipeline trustworthy:

• Executor determinism — the same plan and seeds produce byte-identical artifacts across runs and across run directories.
• Crash recovery — a real subprocess is killed mid-run; resuming from the ledger completes to artifacts byte-identical to an uninterrupted run.
• Tamper-evident ledger — every step attempt records input hash, code version, config hash, seed, output hash, wall time, and status; the chained hash verification detects any edited or reordered entry.
• Budget safety — the scheduler never exceeds run/wall-clock budgets, and cumulative regret is sublinear on stationary synthetic surfaces.
• Schema discipline — plan/artifact schemas round-trip exactly, reject unknown keys and wrong versions, and golden files make any drift deliberate.
• Ablation completeness — generated grids carry a witness for every factor variation; incomplete coverage is reported honestly instead of silently dropped.

ablation · adapters · artifacts · cli · clock · demo · errors · executor · http_adapter · ledger · mappings · metrics · plan · plan_queries · records · report · scheduler · schema · sequences · surrogate (torch extra) · svg · transforms — with py.typed and full type hints throughout.

examples/ contains three fully offline, runnable examples, each with its own README and captured output:

• checkpoint_resume — plan validation, deterministic execution, and a simulated crash followed by resume-to-identical-artifacts.
• bandit_regret — UCB1/Thompson scheduling against a synthetic oracle with a regret curve emitted as dependency-free SVG.
• mini_program — a complete miniature research program producing an HTML report from real executed computations.

Documentation lives under docs/: architecture, artifacts and artifact layout, canonical JSON, configuration, adapters, ablation, the API reference, a glossary, and per-command references (demo-command.md, ablate-command.md, exit-codes.md, …).

MIT — see LICENSE. Release history is in CHANGELOG.md.