kaos-ml-core 0.1.0a3

Classical ML primitives for the Kelvin Agentic OS — corpus, clustering, LLM labeling, logistic regression on the kaos-content AST

kaos-ml-core

Part of Kelvin Agentic OS (KAOS) — open agentic infrastructure for legal work, built by 273 Ventures. See the full KAOS package map for the rest of the stack.

kaos-ml-core is the classical-ML layer of KAOS — a typed Python API that takes documents from the kaos-content AST through a complete supervised pipeline (featurize → cluster → LLM-label → split → train → evaluate → tune-threshold → predict → aggregate) and emits predictions that round-trip back to the source document via stable block_refs.

It is built on top of: kaos-content (AST + TabularDocument), kaos-nlp-transformers (dense embeddings, L2-normalized output as of 0.1.0a2), kaos-llm-core (LLM-driven labeling), and kaos-mcp (11-tool agentic surface).

It is dependency-light at the BASE: the install pulls only the core KAOS runtime (kaos-core, kaos-content, kaos-nlp-core) plus numpy, scipy, scikit-learn, and joblib. Optional extras layer in the rest of the pipeline — [transformers] for fastembed-backed featurization, [llm] for LLM-driven cold-start labeling, and [mcp] for the MCP tool surface.

The Rust crate is intentionally a stub in 0.1.0a1 — only a version() smoke test, no hot path. v2.0+ phases land sparse-vectorizer and parallel-cosine kernels there as profiling warrants. We don't claim otherwise.

Install

uv add "kaos-ml-core[transformers,llm,mcp]"   # full pipeline
# or
pip install "kaos-ml-core[transformers,llm,mcp]"

kaos-ml-core requires Python 3.13 or newer. The published wheels are cp313-abi3 — one wheel per OS/arch covers every CPython 3.13+ minor (3.13, 3.14, 3.15, …).

Without the optional extras, you can still use Corpus, evaluate, stratified_split, tune_threshold, aggregate_predictions, and the classifier directly with your own feature matrix — the BASE install is useful on its own for downstream code that already produces embeddings.

Quick start

This is the full real pipeline: featurize → split → train → evaluate → tune threshold → save pipeline → load + predict → aggregate to document level. Backed by the live integration test tests/integration/test_pipeline_endtoend.py against the same fixture.

import numpy as np
from kaos_content.model.blocks import Paragraph
from kaos_content.model.document import ContentDocument
from kaos_content.model.inlines import Text
from kaos_content.model.metadata import DocumentMetadata, SourceRef

from kaos_ml_core import (
    Corpus, Pipeline,
    aggregate_predictions, evaluate, stratified_split, tune_threshold,
)
from kaos_ml_core.features import embed_corpus
from kaos_ml_core.train import train_logreg

# A tiny fixture: 3 contracts, mix of arbitration + non-arbitration
# clauses. In a real flow this would come from kaos-pdf or kaos-office.
arbitration = [
    "Any dispute shall be resolved by binding arbitration in Delaware.",
    "The parties agree to submit any controversy to AAA arbitration.",
    "All claims shall be settled by binding arbitration under JAMS rules.",
    "The parties consent to binding arbitration of any contract dispute.",
]
other = [
    "The Seller indemnifies the Buyer against third-party claims.",
    "Force majeure clauses excuse performance during natural disasters.",
    "This agreement is governed by the laws of New York.",
    "The buyer agrees to make payments within 30 days of invoice.",
    "Indemnification is subject to a cap of 10 percent of purchase price.",
    "Notice of breach must be provided in writing.",
    "Confidential information must not be disclosed.",
    "The agreement may be amended only by written consent.",
]


def _doc(uri: str, paragraphs: list[str]) -> ContentDocument:
    return ContentDocument(
        metadata=DocumentMetadata(source=SourceRef(uri=uri)),
        body=tuple(Paragraph(children=(Text(value=p),)) for p in paragraphs),
    )


docs = [
    _doc("contract://A", arbitration[:2] + other[:3]),
    _doc("contract://B", arbitration[2:] + other[3:6]),
    _doc("contract://C", other[6:]),
]
corpus = Corpus.from_documents(docs, level="paragraph")
labels = np.array([
    "arbitration" if corpus.unit(i).text in arbitration else "other"
    for i in range(len(corpus))
])

# 1. Featurize. Embeddings are L2-normalized (kaos-nlp-transformers KNT-101).
X = embed_corpus(corpus)

# 2. Stratified train/test/control split. The control_frac is held out
#    for threshold tuning (CLAUDE.md hard rule #5).
split = stratified_split(labels, test_frac=0.25, control_frac=0.20, seed=42)

# 3. Train logistic regression on the train slice.
clf = train_logreg(X[split.train_idx], labels[split.train_idx])

# 4. Evaluate on the test slice. evaluate() returns precision, recall, F1,
#    accuracy, ROC AUC, confusion matrix, AND a Wilson 95% CI on recall
#    (CLAUDE.md hard rule #3).
test_proba = clf.predict_proba(X[split.test_idx])[:, list(clf.classes_).index("arbitration")]
metrics = evaluate(
    labels[split.test_idx],
    clf.predict(X[split.test_idx]),
    y_proba=test_proba,
    classes=("other", "arbitration"),
)
print(f"F1={metrics.f1:.3f}  recall={metrics.recall:.3f}  ROC AUC={metrics.roc_auc:.3f}")
print(f"Wilson 95% recall CI: [{metrics.recall_ci_lower:.3f}, {metrics.recall_ci_upper:.3f}]")

# 5. Tune the operating threshold on the held-out control set.
control_proba = clf.predict_proba(X[split.control_idx])[
    :, list(clf.classes_).index("arbitration")
]
tuned = tune_threshold(
    labels[split.control_idx], control_proba,
    target_recall=0.85, pos_label="arbitration",
)
print(f"threshold={tuned.threshold:.3f} (control recall={tuned.achieved_recall:.3f})")

# 6. Bundle into a Pipeline and save it. The .kaos directory is portable
#    across hosts — ship to prod with one path.
pipeline = Pipeline(
    embed_model_id="BAAI/bge-small-en-v1.5",
    embed_revision="5c38ec7c405ec4b44b94cc5a9bb96e735b38267a",
    classifier=clf,
    threshold=tuned.threshold,
    classes=("other", "arbitration"),
    kaos_ml_core_version="0.1.0a1",
    train_metrics=metrics,
)
pipeline.save("/tmp/contracts-arbitration-v1.kaos")

# 7. Load the pipeline back and apply to the full corpus.
loaded = Pipeline.load("/tmp/contracts-arbitration-v1.kaos")
predictions = loaded.predict(corpus)
# predictions is a TabularDocument: one row per CorpusUnit with
# block_ref, doc_uri, predicted_label, score, above_threshold.

# 8. Aggregate paragraph-level predictions up to document level. This
#    is the central operation for due-diligence / contract analytics.
agg = aggregate_predictions(
    predictions, by="doc_uri", method="any", positive_class="arbitration",
)
for row in agg.tables[0].rows:
    print(f"{row[0]}: arbitration={row[6]} ({row[2]}/{row[1]} positive paragraphs)")
# contract://A: arbitration=True   (2/5 positive paragraphs)
# contract://B: arbitration=True   (2/6 positive paragraphs)
# contract://C: arbitration=False  (0/2 positive paragraphs)

For an agent-driven flow over the same operations, see the MCP tool surface below — register_ml_tools(runtime) exposes all 11 lifecycle steps as MCP tools.

Concepts

The package is built around a small set of typed primitives. Build on top of kaos-content for the AST and kaos-nlp-transformers for embeddings.

Concept	What it is
Corpus	Frozen, AST-grounded set of text units with bidirectional mapping between internal row indices and block_refs. Built from one or more ContentDocument instances at four granularities — "paragraph" (default), "sentence", "section" (group by section_ref), "document" (one row per doc). Cross-granularity workflows use aggregate_predictions to roll up.
Pipeline	Bundles (embed_model_id, embed_revision, classifier, threshold, classes, kaos_ml_core_version) with save(path) / load(path). Persistence uses joblib for the classifier + JSON manifest with a magic-byte header (load refuses files without it). Pipeline.predict(corpus) does featurization + classification + threshold in one call.
Metrics + evaluate	Precision / recall / F1 / accuracy / ROC AUC / confusion matrix on a held-out set, plus a Wilson 95% CI on recall (CLAUDE.md hard rule #3). wilson_score_interval(positives, total, confidence) is exposed as a public helper for callers that already have raw counts.
SplitResult + stratified_split	Three-way stratified split (train / test / control). control_idx is the held-out subset that tune_threshold consumes — closes CLAUDE.md hard rule #5 ("never tune on the training set") at the API level.
ThresholdResult + tune_threshold	Find the operating threshold that hits a target recall (or precision) on a control set. Refuses degenerate inputs (hard predictions) with a fix-the-data warning.
aggregate_predictions	Cross-granularity bridge. Takes the TabularDocument from Pipeline.predict() and rolls up by doc_uri / section_ref / any other key with method ∈ {any, all, max, mean, count, majority}. Output preserves supporting block_refs so a UI can drill from doc-level decision back to the triggering paragraph.
register_ml_tools(runtime)	Registers the 11-tool MCP surface. Session-scoped registries for corpora / pipelines / predictions; same shape as kaos-tabular's _ENGINES. Tool descriptions chain explicitly ("call X first; pass output to Y") so an agent can run the pipeline end-to-end without external orchestration.
KaosMLCoreSettings	Typed settings (env prefix KAOS_ML_CORE_) — default_embed_model, default_threshold, recall_target, recall_confidence, profile. Resolves through the standard six-level KAOS settings hierarchy.

CLI

kaos-ml-core ships a stub kaos-ml info administrative CLI in 0.1.0a1. The train / evaluate / predict subcommands land in 0.1.0a2 (the Python API and MCP tool surface are the canonical entry points today).

kaos-ml info --json    # version + settings + Rust extension status

MCP tool surface

Importing register_ml_tools(runtime) registers 11 MCP tools spanning the full classifier lifecycle:

from kaos_core import KaosRuntime
from kaos_ml_core.tools import register_ml_tools

rt = KaosRuntime()
n = register_ml_tools(rt)
assert n == 11

The tools (in lifecycle order):

Tool	What it does
kaos-ml-build-corpus	Build a Corpus from ContentDocuments at a chosen granularity.
kaos-ml-corpus-info	Read-only stats.
kaos-ml-cluster	MiniBatchKMeans + k-medoid seed selection (caches the feature matrix).
kaos-ml-label-seeds-with-llm	LLM-driven cold-start labeling via kaos-llm-core.
kaos-ml-train	Stratified split + train + evaluate; returns pipeline_id + Wilson-CI metrics.
kaos-ml-evaluate	Read-only evaluation on a held-out subset.
kaos-ml-tune-threshold	Tune operating threshold on the control set.
kaos-ml-predict	Apply pipeline → predictions table (TabularDocument).
kaos-ml-aggregate	Roll up predictions to doc_uri / section_ref.
kaos-ml-save-pipeline	Persist pipeline to disk.
kaos-ml-load-pipeline	Load + register a saved pipeline.

Tool descriptions reference prerequisite + follow-up tools so an agent can chain them end-to-end.

Use cases

kaos-ml-core was designed for the legal-tech workflows where predict-at-fine-granularity → aggregate-to-coarse-decision is the central operation. Pick the granularity that matches the partner's question:

Use case	Predict at	Aggregate to	Example aggregation
Contract analytics ("find arbitration clauses")	paragraph	doc_uri	method="any" — list contracts that contain the language
Contract analytics ("classify doc type: NDA / SPA / lease")	document	—	direct doc-level classification
Due diligence ("find indemnification sections")	section	doc_uri	method="any" + supporting block_refs for review
Due diligence ("triage 5,000 docs into financials/IP/leases")	document	—	direct doc-level classification
TAR / ediscovery ("responsive vs non-responsive")	document or paragraph	doc_uri (when paragraph-level)	method="any" for production decisions; method="count" for review queues
Privilege detection	section / paragraph	doc_uri	method="any" + supporting refs feeds the privilege log

Compatibility & status

Aspect
Python	3.13, 3.14 (informational matrix entries for 3.14t free-threaded and 3.15-dev). One cp313-abi3 wheel per OS/arch covers all 3.13+ minors.
OS	Linux (manylinux + musllinux, x86_64 + aarch64), macOS arm64, Windows x86_64, Windows arm64. macOS x86_64 deliberately skipped (Apple ended Intel sales in 2023).
Maturity	Alpha. The public API is documented in kaos_ml_core.__all__ (23 symbols).
Stability policy	Pre-1.0: minor bumps may change behaviour. Every change is documented in CHANGELOG.md.
Test coverage	99 unit tests + integration tests (live LLM round-trip + real-PDF fixtures via kaos-pdf).
Type checker	Validated with ty, Astral's Python type checker.

Companion packages

kaos-ml-core is one of the packages in the Kelvin Agentic OS. The broader stack:

Package	Layer	What it does
kaos-core	Core	Foundational runtime, MCP-native types, registries, execution engine, VFS
kaos-content	Core	Typed document AST: Block/Inline, provenance, views
kaos-mcp	Bridge	FastMCP server, kaos management CLI, MCP resource templates
kaos-pdf	Extraction	PDF → AST with provenance
kaos-web	Extraction	Web extraction, browser automation, search, domain intelligence
kaos-office	Extraction	DOCX / PPTX / XLSX readers + writers to AST
kaos-tabular	Extraction	DuckDB-powered SQL analytics
kaos-source	Data	Government + financial data connectors (Federal Register, eCFR, EDGAR, GovInfo, PACER, GLEIF)
kaos-llm-client	LLM	Multi-provider LLM transport
kaos-llm-core	LLM	Typed LLM programming (Signatures, Programs, Optimizers)
kaos-nlp-core	Primitives (Rust)	High-performance NLP primitives
kaos-nlp-transformers	ML	Dense embeddings + retrieval
kaos-graph	Primitives (Rust)	Graph algorithms + RDF/SPARQL
kaos-ml-core	Primitives (Rust)	Classical ML on the document AST
kaos-citations	Legal	Legal citation extraction, resolution, verification
kaos-agents	Agentic	Agent runtime, memory, recipes
kaos-reference	Sample	Reference module for module authors

Packages depend on kaos-core; everything else is opt-in. Mix and match the ones you need.

Development

git clone https://github.com/273v/kaos-ml-core
cd kaos-ml-core
uv sync --group dev --extra transformers --extra llm --extra mcp
uv run maturin develop --release

Install pre-commit hooks (recommended — they run the same checks as CI on every commit, scoped to staged files):

uvx pre-commit install
uvx pre-commit run --all-files     # one-time full sweep

Manual QA commands (the same set CI runs):

cargo fmt --check
cargo clippy --no-default-features --all-targets -- -D warnings
cargo test --no-default-features --lib
uv run ruff format --check python/kaos_ml_core tests
uv run ruff check python/kaos_ml_core tests
uv run ty check python/kaos_ml_core tests
uv run pytest tests/ -m "not live and not benchmark"

Build from source

uv build
uv pip install dist/*.whl

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md for setup, quality gates, pull request expectations, and engineering standards. By contributing you agree to follow the project conduct expectations and certify the Developer Certificate of Origin v1.1 — sign every commit with git commit -s. Please open an issue before starting on a non-trivial change so we can align on scope.

Security

For security issues, please do not file a public issue. Report privately via GitHub Private Vulnerability Reporting or email [e-mail]. See SECURITY.md for the full disclosure policy.

License

Apache License 2.0 — see LICENSE and NOTICE.

Copyright 2026 273 Ventures LLC. Built for kelvin.legal.