causal-worlds 0.21.0

Generate fictional-but-coherent causal operations worlds (executable sim + ground-truth answer-key) from a natural-language description, for benchmarking causal-discovery agents.

causal-worlds

Turn a plain-language description of an operation into a fictional causal world with a declared, ground-truth causal graph — then benchmark whether a causal-discovery method can recover it.

Because the structure is declared (not learned from data), it's an answer key: run any discovery method on the generated data and score how well it recovered the world. The worlds are fiction-first — plausible and internally consistent, not models of any real system — so there is no data to leak and nothing to memorize, which is exactly what makes a causal benchmark trustworthy.

from causal_worlds import worlds, grade_spec, InterventionalCiDiscoverer

spec = worlds.get("coffee")                          # a hidden confounder + a regime sign-flip
report = grade_spec(spec, InterventionalCiDiscoverer())
print(report)   # directed_shd=0  skeleton_shd=0  f1=1.0  confounded_reported=0
#                ^ swap in YOUR discoverer to benchmark it against a known truth

Status — v0.6, beta. The full loop works: natural language → an admitted causal world, persisted with provenance. A Claude author proposes the world; an independent Gemini judge (a different model family) plus statistical gates admit only worlds that are valid, recoverable, and not guessable from variable names. The deterministic engine (specify → sample → grade → score) and all grading run with no API key; only authoring needs keys. Worlds are currently tabular SCMs with do() interventions — a Gymnasium env, temporal lags, and counterfactual replay are on the roadmap. See the CHANGELOG.

Install

pip install causal-worlds              # or: uv add causal-worlds
pip install 'causal-worlds[discover]'  # + the baseline discovery stack (PC/GES/FCI/GIES)
pip install 'causal-worlds[llm]'       # + natural-language authoring (Claude + Gemini)

The base install (engine, grading, built-in worlds, CLI) needs only typer, pydantic, numpy.

60-second quickstart (no API key)

causal-worlds worlds                     # list built-in worlds: coffee, ecommerce
causal-worlds gate coffee                # run the validity gates -> admitted=True
causal-worlds grade coffee               # grade the reference discoverer -> directed_shd=0 ...
causal-worlds score benchmark/v0.5/world_01   # grade the reference on a shipped benchmark world

New to it? Walk through the getting-started guide or run the examples.

Benchmark your own discoverer

Implement one method — recover(substrate, *, seed) -> set[(src, dst)] — and grade it against any world's answer key:

from causal_worlds import grade_spec, worlds

class MyDiscoverer:
    def recover(self, substrate, *, seed):
        sample = substrate.sample(2000, seed=seed)        # observational data...
        flows = substrate.sample(2000, seed=seed, do={"price": 1.0})  # ...or interventional
        return {("price", "demand")}                       # your recovered edges

print(grade_spec(worlds.get("coffee"), MyDiscoverer()))

Or from the CLI on a persisted world: causal-worlds score <bundle> --discoverer your_pkg:YourClass.

Benchmark a controller (Stage 2 — control)

The same worlds are a control benchmark: pick lever values to maximise an objective. Because the mechanisms are declared, the best the levers can do is computable — a by-construction optimal policy (scope §1a) — so a policy is graded by regret against it, no external data needed.

from causal_worlds import default_objective, grade_control, optimal_policy, worlds

spec = worlds.get("coffee")
objective = default_objective(spec)            # controllables raise the outcome KPI, quadratic cost
report = grade_control(spec, objective, {"price": 0.0})   # your policy here
print(report.regret, report.optimal_policy)    # regret vs the declared optimum (0 = optimal play)

A pluggable Controller (one method, control(substrate, objective, *, seed) -> {lever: value}, which may do()-experiment on the world but never sees the mechanisms) is graded by grade_controller.

Author a world from a description (needs [llm] + keys)

Set ANTHROPIC_API_KEY and GEMINI_API_KEY (see .env.example; the CLI auto-loads a local .env), then:

causal-worlds generate "a coffee chain with weekend swings and variable lead times" ./my-world

Or describe a world conversationally. A one-shot prompt is underspecified, so elicit runs a short dialogue first — it asks the minimal clarifying questions (entities & roles, what drives what, regimes, hidden causes, the objective), shows the accumulating brief, and authors only once the brief is complete (or you type go):

causal-worlds elicit ./my-world      # interactive: answer a few questions, then it generates

Observability: with the observability extra + Langfuse keys and CAUSAL_WORLDS_LANGFUSE_ENABLED=true, every run is traced (generate → author → gate) in Langfuse.

from causal_worlds import generate
from causal_worlds.author import build_claude_author
from causal_worlds.judge import build_gemini_judge

world = generate(
    "a hospital ED with triage staffing and bed pressure",
    author=build_claude_author(complexity="hard"),   # easy | standard | hard | adversarial
    judge=build_gemini_judge(),                       # independent model family
)
print(world.report.difficulty, world.report.grade)

What the crossover shows (and what it doesn't)

Across the 35-world benchmark/v0.5 set (3 seeds each). The comparison is information-fair: the +do methods get the same interventional budget (pooled observational + per-variable do() environments) as the latent-aware reference — so we compare methods, not data access (full table + bootstrap CIs):

method	data	latent-aware?	mean skeleton-SHD ↓	confounded pair kept as causal ↓
interventional-ci (reference)	interventional	yes	1.44	0
GIES	interventional	no	4.62	17
PC	observational	no	2.72	14.3
PC + interventions	interventional	no	3.31	15.0
FCI	observational	partly	2.68	9.7
FCI + interventions	interventional	partly	3.29	6.7
DAGMA	observational	no	5.73	16.0
DirectLiNGAM	observational	no	5.64	14.7

(DAGMA and DirectLiNGAM run at default hyperparameters, and LiNGAM's non-Gaussian assumption is violated by these linear-Gaussian worlds, so their skeleton accuracy is not their best — but the relevant, robust verdict is confounded-kept, and like every causal-sufficiency method they keep it.)

The honest reading: the dividing line is latent-awareness, not interventions. The decisive row is PC + interventions — given the same interventional budget as the reference, it still keeps the hidden-confounded pair as a causal edge in ~15 worlds (no better than observational PC's 14.3); GIES likewise (17). Only the latent-aware interventional rule reaches 0. The interventional advantage is robust: ΔF1 = F1(reference) − F1(method) is +0.29, 95% CI [0.22, 0.35] for pc+do (every method's CI excludes 0). So this is an identifiability result (you cannot tell confounding from causation without both interventions and a latent-aware method), not "our method beats the toolbox."

Caveats we're not hiding (see evals/ and the issues): (1) the worlds are admitted by the reference grader itself Fixed in v0.15: admission (gate T3) is now grader-independent — a world is admitted iff its declared SCM is faithful by construction (every edge induces a detectable partial correlation; regimes genuinely modulate), computed in closed form from the spec with no discovery method run. The reference grader's score is reported, never gates. (2) Simulated-DAG leakage — synthetic SCMs can leak the causal order through marginal variance (varsortability) and through scale-invariant predictability (R²-sortability). v0.14 generates worlds with internal standardization (iSCM), dropping varsortability to 0.54 and R²-sortability 0.73 → 0.60; both trivial sorting baselines fall to F1 ≈ 0.33–0.37, well under the real methods. The residual R²-sortability (0.60 > 0.5) is disclosed, not yet fully closed. (3) Difficulty vs skeleton-SHD error is descriptive, not a validated predictor: with bootstrap CIs (n=35), the observational methods show r≈0.40 (PC [0.07, 0.68], FCI [0.08, 0.68] — just excluding 0) while the latent-aware reference is flat (r≈0.24, [−0.06, 0.51], includes 0). (4) The shipped benchmark/v0.5 is still name-guessable — being fixed. A name-only LLM baseline scores F1 0.71 vs a 0.20 chance floor (names and roles leak). v0.19 hardens the machinery for the next generation: T4 now admits only worlds with difficulty ≥ 0.5 (named-prior F1 < 0.5, down from the old 0.9 bar) plus a blind control (the name+role-anonymized prior must sit near chance), and an adversarial author tier writes worlds where the obvious name-based guess is wrong (phantom edges, reversed edges, regime sign-flips — keeping every true edge detectable). The v0.5 set predates this; regenerating it under the strict gate is the next scaled run.

What you get per world

1. An executable SCM — sample observational data and do()-intervene, deterministically by seed.
2. A time-series dataset — the observed variables (the input to a discovery method).
3. An answer key — the declared causal edges + the hidden-confounded pairs, derived from the spec.
4. A manifest — full provenance (models, grader version, seed, difficulty) and an honesty label.

Concepts

• Spec / IR — variables (with roles, incl. hidden), linear-Gaussian mechanisms, regime sign-flips.
• Answer key — directed edges over observed variables + the hidden-confounded pairs; derived from the spec, never stored separately, so they can't disagree.
• Gates — T1 validity · T2 sample-sanity · T3 faithfulness (grader-independent: the declared SCM is faithful & non-trivial by construction) · T4 anti-cliché (the named prior recovers < half — difficulty ≥ 0.5 — and a name+role-blind prior stays near chance). A world is admitted only if all pass.
• Reference grader — an interventional-CI discoverer that uses do() data to tell confounding from causation, where PC/GES/GIES/FCI (which assume causal sufficiency) cannot.

Depth: docs/scope.md · docs/hld.md · docs/lld.md · docs/architecture.md · docs/validation.md.

Roadmap

Shipped: NL authoring, independent judge + anti-cliché gate, artifact persistence, the baseline crossover, a structural-difficulty axis, a 35-world benchmark, temporal worlds (lagged edges + autoregression — see the built-in supply), and time-series grading (PCMCI+, LPCMCI, VARLiNGAM, Granger — grade_temporal_spec), authoring temporal worlds (an LLM-authored lagged world, admitted through a PCMCI+ temporal gate), and conversational elicitation (causal-worlds elicit — a dialogue that builds a WorldBrief before authoring). and the control track (Stage 2): a by-construction optimal-policy answer-key with regret scoring (grade_control — see scope §1a). Next: regret-under-perturbation (regime-aware vs static — the stay-optimal thesis) and a Gymnasium env; tightening the anti-cliché gate at scale (#12); a temporal benchmark set (n>1); and scaling to 100+ worlds. Tracked as issues.

Why this is the unoccupied intersection

Today's tools each own one corner — natural-language authoring × executable causal simulator × ground-truth answer-key for discovery is the gap:

Tool	Corner it owns	What it lacks (for this job)
G-Sim	LLM authors a sim + calibrates to data	needs real data; aimed at fidelity, not a declared answer-key
DEVS-Gen	NL → executable discrete-event ops sim	no declared causal-graph answer-key
SD-SCM	LLM fills mechanisms → counterfactuals	needs a user-supplied DAG; tabular, not an executable sim
TimeGraph	known-graph time-series for discovery	parametric/templated; no natural-language authoring

Built on the shoulders of pgmpy, DoWhy, CausalPlayground, causal-learn, and Gymnasium.

Contributing

Issues and PRs welcome. The bar: make validate green (ruff select=ALL, mypy strict, pytest with a coverage floor) — see docs/engineering.md. Atomic, conventional commits.

License

MIT. An open-source project from Noumenal.