Skip to content

Repository files navigation

mfg-bench

An open, MIT-licensed benchmark for the white-collar cognitive work of manufacturing and operations — the reasoning, planning, diagnosis, and decision-making done at a desk (quality management, supply chain, production planning, continuous improvement, cost analysis, compliance interpretation), explicitly excluding the machine / physical layer (robot/PLC control, raw sensor telemetry, physical process control, vision defect detection).

Status: bootstrapping. The scaffold and the autonomous build loop are in place; the benchmark itself is built unit-by-unit from docs/roadmap.md. See docs/state.md for current progress.

Why it's credible: generation ≠ grading

Claude generates tasks, but Claude's opinion is never the source of truth. Ground truth comes from computation (a formula/algorithm), a deterministic simulator, or an authoritative cited source. This is the same principle that makes SWE-bench trustworthy (the test suite is the arbiter), applied to operations. See GOALS.md and SPEC.md.

Task families (v1 headline = A + B + C)

  • A — Computed: deterministic formula/algorithm truth (OEE, MRP, EOQ, SPC/Cpk, scheduling, FMEA arithmetic, cost variance…). ~55%.
  • B — Source-grounded: canonical public structures, closed-form, cited (8D, APQP/PPAP, 7 wastes, SMED/5S/kanban…). ~20%.
  • C — Simulated-outcome: a deterministic ops simulator scores the KPI of the model's decisions/orchestration (L4/L5). ~25%.
  • D — Rubric-judged (LLM-as-judge): optional, experimental, reported separately, gated on kappa ≥ 0.7. Not in the v1 headline.

Repository map

GOALS.md            North star (read first).        SPEC.md    Full build spec (authority).
CLAUDE.md           Conventions for every session.  config.yaml  Global config.
docs/roadmap.md     The ordered build queue.        docs/state.md  Living progress log.
.loop/              The autonomous build loop (prompt + budget ledger).
schemas/            task / result / rubric JSON schemas.
generators/ scorers/ simulator/ harness/            The benchmark machinery (filled by the loop).
taxonomy/           Coverage plan across domain × tier × format.
data/public/        Public snapshot (may be memorized).  data/heldout/  Private-seed set (gitignored).

How it's built: the autonomous loop

A Routine fires every 4 hours (fresh session). Each firing reads GOALS.md + the roadmap, checks its self-governed budget (.loop/budget.yaml), advances one coherent unit, opens a documented issue + PR, and auto-merges on green CI. The repo is the state; the loop is stateless. Full contract: .loop/build-loop.md.

Running (once Phase 0 lands)

pip install -r requirements.txt
pytest -q
# python -m harness.run --generator oee --seed 123 --model anthropic   # after unit 0.1

License

MIT — see LICENSE.

About

Benchmark for frontier models on non-operating, non-machine activities in a typical manufacturing environment. Open source, MIT license.

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages