auto-causal-lib 0.11.3

Auto-impute, SLM-directed AutoCleanse/AutoEDA/AutoMine suites, agentic causal loop (compact+memory), Kineteq GRAIL soft loop, MCP/AgentHook connective tools, mine, discover, public multi-source causal mining, NLTK NLP hints, behavioral science traces, KPI ML loops (SLM→PyTorch), physics predictive loop + Streamlit demo, direction guides, SLM create/infer, tool suite, and ground exploratory causal relationships from CSV and SQL databases.

AutoCausalLib (auto-causal-lib / import autocausal)

Automatically impute missing tabular fields and discover exploratory causal relationships from CSV / Parquet and SQL databases - with optional SLM-aided creation/inference, a shared tool suite, and a physics predictive / autocausal loop for physical insight grounding.

Scope is intentionally small: impute → role inference → PC-lite / score edges → optional IV → optional physics rollout.
This is not a full AutoML OS and does not guarantee causal identification.

Features

• Load from CSV, Parquet, or SQLAlchemy URLs (Postgres, Vertica, DuckDB, and more via extras)
• Auto-imputation (median_mode, knn, or auto) with strategy reporting
• Role inference (treatment / outcome / instrument / confounder candidates)
• Exploratory discovery: PC-lite + scored edges + optional 2SLS
• Mining - column profiles, associations, KPI hints
• SLM - RuleBackend always; optional HuggingFace for creation (questions/Z/morphemes) and inference (narrative/caveats)
• Direction guides - soft-optional LLMIntent / retracement / Kineteq pivot embeddings / GRAIL → DirectionPlan (see docs/GUIDES.md, docs/GRAIL.md)
• GRAIL (autocausal.grail) - embellished Kineteq Generative Reflective Agentic Imputation Loop; live MCP/module when configured, rich offline stub otherwise; MCP tools autocausal_grail_*
• suite_tools - registry of causal/NLP/KPI/validation adapters (NLTK, gensim, DoWhy stubs, …)
• Physics loop - analytic KPI dynamics (damped oscillator / drift-diffusion / linear ODE), physical insight grounding, PhysicsCausalSuite.loop, optional Streamlit demo (physics ui)
• KPI ML loop - SLM/Rule ModelConstructPlan → median/sklearn/PyTorch MLP impute → discover → FitReport (docs/ML_KPI_LOOP.md)
• Isolates causal - soft bridge to IntentIsolates layer motifs → indication vs IV (docs/LAYER_CAUSAL_IV.md)
• NLP library (autocausal.nlp) - soft-optional NLTK tokenize/POS/sentiment, TextCausalHints, NlpFeatureBuilder for apps/notebooks (docs/NLP_AND_BEHAVIORAL_TRACES.md)
• Behavioral traces (autocausal.behavioral) - habit/nudge/reinforcement demos → panel → mine/discover (docs/NLP_AND_BEHAVIORAL_TRACES.md)
• Public causal mining - multi-source join of bundled/open datasets → mine → discover → report (docs/PUBLIC_CAUSAL_MINING.md)
• Insight suite (autocausal.insight) - InsightReport + optional SLM; closed research loop recommends experiments and mines further (run_loop / ExperimentRecommender) (docs/INSIGHT_SUITE.md)
• Auto suites (autocausal.suites) - SLM-directed AutoCleanseSuite / AutoEDASuite / AutoMineSuite with dedicated action modules + autocausal.skilling tool surface (docs/SUITES.md, docs/SLM_SKILLING.md)
• MCP connective (autocausal.mcp / autocausal.connective) - Model Context Protocol stdio server + in-process AgentHook so other agents can load/cleanse/mine/discover/report (docs/MCP.md)
• Agentic loop (autocausal.agentic) - SLM-guided cyclic FSM: hypothesize → skill → validate → compact → persist → route, with MEM1-inspired memory + ACON-inspired compaction (docs/AGENTIC_LOOP.md)
• Fabric contracts - to_mine_report / to_causal_edges / to_fabric_bundle / to_search_dag aligned with shared Causal Fabric schemas (docs/LIBRARY_API.md)
• Discovery stability & ensemble - bootstrap per-edge stability (honest confidence); multi-method consensus (pc_lite + corr_skeleton + mi_stub)
• QC gate - autocausal.qc.validate_frame before discover (ID leakage / bad keys)
• Panel / join / IV handoff - PanelSpec, join.align, to_causaliv_request, sensitivity + soft refute hooks
• Causal backends (0.11) - soft causal-learn / LiNGAM / gCastle discovery; DoubleML + EconML estimate; real DoWhy refute (docs/CAUSAL_BACKENDS.md)
• Engines surface - autocausal.engines list/status + CLI engines / estimate / refute; MCP autocausal_list_engines / _estimate / _refute
• Markdown / JSON reports and a CLI

Install

PyPI name: auto-causal-lib · Import: autocausal
pip install autocausal is not this project (name rejected as too similar to other packages). Always:

pip install auto-causal-lib
python -c "import autocausal; print(autocausal.__version__)"

From PyPI:

pip install auto-causal-lib
pip install "auto-causal-lib[all]"   # nlp + slm + mcp + ui + ml + causal-extra + web + drivers

Optional extras (soft deps; core works without them):

pip install "auto-causal-lib[nlp]"           # nltk + gensim
pip install "auto-causal-lib[slm]"           # torch + transformers (lazy load)
pip install "auto-causal-lib[mcp]"           # MCP stdio server for other agents
pip install "auto-causal-lib[ui]"            # Streamlit physics demo (+ plotly)
pip install "auto-causal-lib[ml]"            # torch + scikit-learn
pip install "auto-causal-lib[causal-extra]"  # causal-learn, DoWhy, DoubleML, EconML, lingam, gCastle
pip install "auto-causal-lib[postgres]"      # and other DB drivers - see docs/CONNECTIONS.md

First-class modules in every wheel (no heavy deps required): insight, mcp/connective, skilling, cli, backends/engines, agentic, grail, suites.

python -m autocausal engines status
python -m autocausal insight --help
python -m autocausal skilling list
python -m autocausal.mcp          # needs [mcp] for SDK; AgentHook works without it

Docs: docs/INDEX.md (full map) · docs/MODULES.md · docs/CLI.md · docs/MCP.md · docs/CAUSAL_BACKENDS.md · docs/LIBRARY_API.md.

From source (development):

cd research/AutoCausalLib
pip install -e ".[dev]"

Env:

Variable	Effect
AUTOCAUSAL_SLM=1	Prefer HuggingFace SLM
AUTOCAUSAL_SLM_MODEL	Model id (default sshleifer/tiny-gpt2 for tests)
AUTOCAUSAL_TORCH=1	Prefer PyTorch MLP imputer/predictor when installed
AUTOCAUSAL_TORCH_TEST=1	Enable gated torch unit tests
AUTOCAUSAL_LLMINTENT_MODEL	Optional LLMIntent heavy analyzer model
AUTOCAUSAL_KINETEQ_MCP=1 + KINETEQ_MCP_URL	Live Kineteq MCP pivot embeddings / GRAIL
AUTOCAUSAL_GRAIL_MCP=1	Also enables live GRAIL MCP calls

Better instruct SLMs (document only): Qwen/Qwen2.5-0.5B-Instruct, HuggingFaceTB/SmolLM2-360M-Instruct, microsoft/Phi-3-mini-4k-instruct.

Core deps: numpy, pandas, sqlalchemy. See docs/CONNECTIONS.md. Optional path deps: pip install -e ../LLMIntent.

Quick start

from autocausal import AutoCausal

ac = AutoCausal.from_csv("data.csv")
result = ac.run()          # impute + discover
print(ac.report())         # markdown (AutoCausal.report)
print(result.report())     # same via DiscoveryResult.report()
print(result.to_json())    # graph + edges + candidates

# Engines connectivity (list / status / estimate / refute)
from autocausal.engines import list_engines, engine_status
print(engine_status()["n"], "engines")

# 0.8: stability, QC, fabric, NLP→guide
ac.enrich_from_text("Does spend cause sales?")
result = ac.discover(stability=True, bootstrap_n=12, ensemble=True)
print(ac.to_fabric_bundle()["schema"])      # FabricBundle.v1
print(result.to_fabric_bundle()["schema"])  # same via DiscoveryResult
est = result.estimate(backend="builtin_ols")  # also: ac.estimate(...)
ref = result.refute(method="placebo")         # also: ac.refute(...)
print(ac.to_causaliv_request()["schema"])

# SLM/rule creation + inference
print(ac.create(text="Does spend cause sales?").to_markdown())
print(ac.interpret().to_markdown())

# Tool suite validation
print(ac.validate_tools(y="y", d="d", z="z").to_markdown())

# Direction steering (LLMIntent / retracement / Kineteq pivots / GRAIL - soft-optional)
plan = ac.direct(
    text="Does spend cause revenue?",
    backends=["llmintent", "retracement", "kineteq_pivot", "grail", "rule"],
)
print(plan.to_markdown())

# GRAIL reflective loop (offline stub unless Kineteq MCP configured)
from autocausal.grail import GrailEngine
report = GrailEngine().run("Does spend cause revenue?", context={"text": "Does spend cause revenue?"})
print(report.to_markdown())

# Physics predictive / autocausal loop
from autocausal.physics import PhysicsCausalSuite

suite = PhysicsCausalSuite.from_csv("data.csv")
phys = suite.loop(horizon=5, text="what drives outcome?")
print(phys.to_markdown())
# Or: ac.physics_loop(horizon=5) / AutoCausal.auto(..., physics=True)

# KPI-mined ML loop (SLM/Rule constructs torch vs median imputer)
from autocausal.ml import KPIMinedCausalLoop

ml = KPIMinedCausalLoop.from_csv("data.csv").run(
    text="what drives Y?", use_slm=False, use_torch=True, horizon=5
)
print(ml.plan.to_markdown())
print(ml.fit.to_markdown())

Real dataset examples (offline)

Bundled Iris, Wine, Titanic, Gapminder subset, Diabetes, and California housing sample - no network required. Licenses/attribution: DATASETS.md. Full walkthrough: docs/EXAMPLES.md.

from autocausal import AutoCausal, load_dataset

df = load_dataset("iris")  # Fisher Iris CSV from package data
ac = AutoCausal(df)
ac.mine().impute().discover(use_iv=False, min_abs_corr=0.2)
print(ac.report())  # exploratory edges - not scientific flower-causation claims

python examples/iris_causal.py
python examples/iris_causal.py --insight
python examples/multi_dataset_tour.py
python -m autocausal public load iris
python -m autocausal insight demo --dataset iris --no-slm

NLP hints & behavioral traces (library-first)

These are importable modules for apps and notebooks - the CLI is optional.

from autocausal.nlp import extract_causal_hints_from_text, NlpFeatureBuilder
from autocausal.behavioral import BehavioralTraceStore, mine_behavioral_traces

hints = extract_causal_hints_from_text(
    "Randomized treatment leads to higher revenue, associated with age."
)
print(hints.roles.to_dict())          # treatment / outcome / confounder / instrument cues
print(hints.to_guide_context())       # feed guide/discover

features = NlpFeatureBuilder().transform(["because spend increases sales"])

result = mine_behavioral_traces("habit_loop", discover=True)
print(result.report.to_markdown())    # hypothesized stimulus→response / habit→outcome edges

See docs/NLP_AND_BEHAVIORAL_TRACES.md (Python API first, CLI secondary).

Public causal mining (library-first)

Join bundled/open demo sources, mine associations, and run exploratory discovery:

from autocausal import AutoCausal, PublicCausalMiner, mine_public

report = AutoCausal.mine_public(
    ["finance_demo", "demographics_demo", "health_demo"],
    join_on="region",
    discover=True,
    use_iv=True,
)
print(report.to_markdown())

# Explicit miner
miner = PublicCausalMiner(["marketing_demo", "instruments_demo", "demographics_demo"])
report = miner.run(discover=True, validate=True)

# Convenience
report = mine_public(["finance_demo", "climate_demo"], discover=True)

python -m autocausal public list --offline
python -m autocausal public mine --sources finance_demo,demographics_demo --discover
python -m autocausal public causal --sources finance_demo,demographics_demo,health_demo -o report.md

See docs/PUBLIC_CAUSAL_MINING.md.

Auto suites - Cleanse / EDA / Mine (SLM-directed)

Every auto* path is directed by SLMAutoDirector when available; rules always work offline.

from autocausal import AutoCausal, AutoCleanseSuite, AutoEDASuite, AutoMineSuite

clean = AutoCleanseSuite(df, use_slm=True).run()
eda = AutoEDASuite(clean.frame, use_slm=True).run()
mine = AutoMineSuite(clean.frame, use_slm=True).run()

ac = AutoCausal.from_dataframe(df).cleanse().eda().automine().discover()
# or: AutoCausal.auto("data.csv", use_slm=True, cleanse=True)

python -m autocausal suite cleanse --csv data.csv --no-slm -o cleanse.md
python -m autocausal suite eda --csv data.csv -o eda.md
python -m autocausal suite mine --csv data.csv --format json -o mine.json

See docs/SUITES.md and docs/SLM_SKILLING.md.

Agentic causal loop (library-first)

SLM-guided cyclic research loop with compaction + constant-budget memory (SOTA-inspired APIs - not paper clones):

from autocausal import AutoCausal, load_dataset
from autocausal.agentic import AgenticCausalLoop, run_agentic_loop

df = load_dataset("iris")
report = run_agentic_loop(df, text="petal drivers", max_rounds=2, use_slm=False)
print(report.to_markdown())

# Or: AutoCausal(...).agentic_loop(...) / MCP tool autocausal_agentic_loop

See docs/AGENTIC_LOOP.md.

Use from other agents (MCP)

Expose AutoCausal as MCP tools for Cursor, Claude Desktop, and other MCP clients - or call the same surface in-process via AgentHook (no mcp SDK required).

pip install -e ".[mcp]"
python -m autocausal.mcp          # stdio server
# or: autocausal-mcp

Cursor / Claude Desktop stdio config:

{
  "mcpServers": {
    "autocausal": {
      "command": "python",
      "args": ["-m", "autocausal.mcp"],
      "env": { "PYTHONUNBUFFERED": "1" }
    }
  }
}

Library-first (scripts / non-MCP agents):

from autocausal.connective import AgentHook

hook = AgentHook()
hook.call_tool("autocausal_load_dataset", {"dataset_id": "iris"})
hook.call_tool("autocausal_discover", {"use_iv": False})
print(hook.call_tool("autocausal_report", {"format": "markdown"})["markdown"][:400])

Tools include autocausal_load_dataset, autocausal_from_csv, autocausal_cleanse / eda / mine, autocausal_discover, autocausal_insight_loop, autocausal_recommend_experiments, autocausal_public_mine, autocausal_report, autocausal_list_datasets, autocausal_skilling_list. Soft-fail if optional suites are missing. Full setup: docs/MCP.md.

Insight suite (library-first)

from autocausal.insight import InsightSuite, ExperimentRecommender, run_insight_loop, demo_insight

report = run_insight_loop("data.csv", text="what drives revenue?", use_slm=False)
report.write("insight.md")

# Closed loop: mine → guide/SLM → recommend experiments → join/remine → rediscover
suite = InsightSuite(use_slm=False)
report = suite.run_loop(
    "data.csv", max_rounds=3, join_sources=["demographics_demo", "instruments_demo"]
)
print(report.experiments_recommended[:3])
print(report.round_history)

# From a pre-built AutoCausal
from autocausal import AutoCausal
ac = AutoCausal.from_csv("data.csv")
report = InsightSuite.from_autocausal(ac).run(use_slm=False)

python -m autocausal insight run --csv data.csv --no-slm -o report.md
python -m autocausal insight loop --csv data.csv --rounds 3 --no-slm -o loop.md
python -m autocausal insight demo
python -m autocausal insight demo --dataset iris --no-slm

See docs/INSIGHT_SUITE.md and docs/EXAMPLES.md.

Guiding direction with LLMIntent / Retracement / Kineteq pivots / GRAIL

Backends are soft-optional: missing packages soft-fail to stubs/fallbacks and never break core discovery.

python -m autocausal guides list
python -m autocausal auto --csv data.csv --text "what causes revenue?" \
  --guides llmintent,retracement,kineteq_pivot,grail
python -m autocausal direct --csv data.csv --text "..." --guides grail,rule

See docs/GUIDES.md and docs/GRAIL.md.

python -m autocausal discover --csv data.csv
python -m autocausal create --csv data.csv --text "lottery assignment"
python -m autocausal infer --csv data.csv
python -m autocausal tools list
python -m autocausal tools validate --csv data.csv --y y --d d --z z
python -m autocausal physics loop --csv data.csv --horizon 5 --text "what drives outcome?"
python -m autocausal physics rollout --csv data.csv --horizon 5
python -m autocausal physics ui --port 8518
python -m autocausal auto --csv data.csv --physics --horizon 5
python -m autocausal ml loop --csv data.csv --text "what drives Y?"
python -m autocausal ml loop --csv data.csv --torch --guides rule
python -m autocausal ml fit-imputer --csv data.csv --backend median
python -m autocausal nlp extract --text "treatment leads to revenue"
python -m autocausal behavioral list
python -m autocausal behavioral mine --demo habit_loop --discover
python -m autocausal public list --offline
python -m autocausal public mine --sources finance_demo,demographics_demo --discover
python -m autocausal public causal --sources finance_demo,demographics_demo,health_demo -o report.md
python -m autocausal slm-status
python -m autocausal guides list
python -m autocausal auto --csv data.csv --slm

Physics Streamlit demo

Interactive UI for the physics autocausal loop (trajectory charts, edges, physical insights, energy proxies). Soft-optional - core mine/discover/ml loops do not import Streamlit.

pip install -e ".[ui]"
python -m autocausal physics ui --port 8518
# or: streamlit run src/autocausal/apps/physics_streamlit.py --server.port 8518

See docs/PHYSICS_DEMO.md. Caveat: exploratory dynamics only - not true physics ID.

PostgreSQL

pip install -e ".[postgres]"
python -m autocausal discover \
  --db "postgresql+psycopg2://user:pass@localhost:5432/mydb" \
  --table events

Epistemic caveats

AutoCausalLib is an exploratory toolkit. Please read these before treating outputs as science:

• Discovery ≠ identification. PC-lite / scored edges are candidate relationships, not proven causal effects.
• Imputation and joins change the sample. Missing-data fills and multi-source joins can invent associations (including ecological fallacy on region aggregates).
• SLM text is assistance only. Narratives, experiment suggestions, and role hints from rules/HF models are generative - not statistical proof.
• IV / 2SLS paths are soft. Optional instruments need human design review (relevance, exclusion); lite F-stats are not a substitute.
• Bundled public/behavioral tables include MIT synthetic fixtures and real educational CSVs (Iris, etc.) - see DATASETS.md. Exploratory edges on Iris are illustrative, not flower-causation science.
• Physics / KPI loops are predictive rollouts and grounding aids - not true physical or causal identification.

Reports (InsightReport, PublicCausalReport, markdown CLI output) repeat these caveats; keep them in downstream apps.

Docs

• Library API map (0.8+)
• Roadmap (P1-P3 shipped)
• Examples (Iris + real datasets)
• Dataset licenses & paths
• Insight suite (library API + optional SLM)
• Auto suites - Cleanse / EDA / Mine (SLM-directed)
• SLM skilling / tool surface
• MCP connective (agents / Cursor / Claude)
• Agentic causal loop (compact + memory)
• Public causal mining (multi-source join)
• NLP & behavioral traces (library API)
• KPI ML loop (SLM → PyTorch)
• ML Model Hub proposals
• Physics Streamlit demo
• Physics world models + autocausal loop (SOTA)
• Direction guides
• GRAIL (Kineteq adaptation)
• Tool suite registry
• Connection matrix & pip extras
• SOTA context (PC / GES / NOTEARS, imputation)

Related suite

Project	Role
EmotiveVision	Emotion/intent streams → autocausal frames
CausalIVSuite	IV / DiD / AutoML causal suite
CausalSearch	Causal evidence search & DAG infill
CausalBridge	Control plane (status shows SLM/tools)
NextFrameSeq	Vision / next-frame prediction

License

MIT