metispy

The metis python referee + research kit. Import name stays metis.

This is the SEMANTIC side of the metis toolchain: the exact measure (forward filter, factorized tables, clique trees), learning (Dirichlet notching, MAP-EM), decision, the surprise record + recover() replay, and the stage-45 formalization harness (metis.lang.formalize.admit_text_pack) — the gate cascade the metis-loop drives. It is NOT the shipping toolchain: the compiler is metisc (metis-lang, OCaml) and the runtime is metis-lang's C++ headers; metispy referees both.

• metis/kernel/ — measure semantics, factorization, IR emission, canonical identity, notching, records.
• metis/lang/ — parser/compiler/typecheck (the python front-end, kept as differential oracle for metisc), the LSP dev server, the formalize harness, and the discovery pair over #-directives (00-language-spec §8): directives.py (the llp.trace / metisc --run replay) and repl.py — the metis <file.llp> shell (readline editing/history/completion; step/sample on the reference sampler, assert/retract, exact query from the current state, reload with degrade-to-last-good, save back to directives).
• metis/fw/ — the python evaluator product loop (truth generator for the C++ Evaluator).
• metis/domains/ — the benchmark playground models (GoL, kitchen/Overcooked, nav, sysadmin, xtraffic, road). Their catalogs live in metis-catalog. The Ceptre love triangle the lineage tests read is metis-catalog's domains/ceptre/love_triangle/: it is under interactive-lp's license, so metispy does not ship it (since 0.1.1).
• tools/ — golden/artifact generators writing INTO the metis-lang checkout (METIS_LANG, default: the sibling ../metis-lang).
• tests/ — the full gate suite, including the cross-toolchain parity gates (OCaml compiles ≙ python referees ≙ C++ interprets).

Install

pip install metispy         # Python 3.12+; the import name is `metis`, the CLI `metis`

What you get: the referee and research kit above — the exact measure, learning, decision, records, the metis <file.llp> shell and the python front-end, with the domains' catalogs it builds from. It is not the compiler: metisc is metis-lang's (OCaml), built from that repo. A domain NAME (rather than a .llp path) needs the catalog's loader: install metis-catalog, or set METIS_CATALOG to a checkout.

For development, beside its siblings:

pip install -e .            # into the metis-lang dev env
METIS_LANG=../metis-lang python3 -m pytest tests/ -q
python3 -m pytest tests/ -q -m 'not slow'   # the dev loop

Optional extras: pip install overcooked-ai scipy (kitchen), pyRDDLGym (sysadmin), pgmpy (graph-evaluator comparisons). Oracle parity vs the wayplan ll2pg lineage: METIS_ORACLE_PATH.

Releasing

metispy goes to PyPI from GitHub by trusted publishing (.github/workflows/publish.yml; no token anywhere): bump version in pyproject.toml, commit, git tag vX.Y.Z (the tag must be v + that version: the workflow refuses any other), push the tag: it publishes straight through (the pypi environment has no required reviewer). The run builds the sdist and the wheel, smoke-installs the wheel in a fresh venv outside the checkout (tools/smoke_wheel.sh, which ci.yml runs on every push too) and uploads. workflow_dispatch is the dry run: everything but the upload.

Narration (narrate, metis/loop/narrate.py)

narrate tells the session's fired events as prose (narrate SEED: a fresh trace sampled with that seed). The narrator sees only certified material — the Foata beats, the final state and the catalog's %% docs — so the docs are its vocabulary: one plain sentence per story row. Four conventions keep a mined or machinery-heavy catalog tellable:

Example, metis-catalog domains/wildfire/wildfire.llp:

  % evidence: "k, the number of its neighbors burning at the start of the step"
  %% [fold tally] the fire counts the burning neighbours of each cell
  see_fire [dyn] : scan(X, Y, S) * tally(X, Y, K) * ... -o scan(X, Y, S2) * tally(X, Y, K2).
  %% X,Y, with K burning neighbours, catches fire | X,Y, with K burning neighbours, stays unburnt.
  catch [dyn] : fresh(X, Y) * tally(X, Y, K) * not_burning(X, Y) -o (burning(X, Y) @w pign(K) | not_burning(X, Y) @w ~pign(K)).

The material asks for a length sized to what is told, and a narration the endpoint cut at its token cap ends on its last full sentence with "(narration cut at N tokens; ask fewer beats)".

License

Apache License 2.0 (see LICENSE, NOTICE).