graphite-engine 0.3.0

Open-source claim verification engine for high-stakes decisions

⛏️ Graphite

Open-source claim verification engine for high-stakes decisions.

Turn documents into structured claims with provenance, then verify whether downstream assertions are actually grounded in evidence.

⚠️ v0.3.x — Experimental. Usable and tested, but API may change before 1.0. Pin your version.

---

10 lines to your first verified claim

from graphite import Claim, ClaimStore, ClaimType, Provenance
from graphite.enums import SourceType, ConfidenceLevel

# 1. Extract a claim from a document
claim = Claim(
    subject_entities=["company:TSMC"],
    predicate="SUPPLIES_TO",
    object_entities=["company:NVDA"],
    claim_text="TSMC supplies advanced CoWoS packaging to Nvidia.",
    claim_type=ClaimType.RELATIONSHIP,
    supporting_evidence=[Provenance(
        source_id="tsmc-10k-2024",
        source_type=SourceType.SEC_10K,
        evidence_quote="The Company provides advanced packaging services including CoWoS.",
        confidence=ConfidenceLevel.HIGH,
    )],
)

# 2. Persist to a claim registry
store = ClaimStore(db_path="/tmp/demo.db")
store.save_claim(claim)

# 3. Query: does anyone supply to Nvidia?
results = store.search_claims(object_contains="NVDA")
for c in results:
    print(f"{c.claim_text}  [{c.supporting_evidence[0].source_id}]")

Every claim is traceable: source, quote, confidence. No black-box assertions.

See examples/quickstart_verification/ for a fully runnable example.

---

Quickstart

pip install graphite-engine

Or from source:

git clone https://github.com/graf-research/graphite.git
cd graphite
pip install -e .
python examples/quickstart_verification/run.py

No database. No LLM. No API keys. Start local, stay local.

---

Why Graphite?

Graphite is built for domains where being wrong is expensive: financial research, regulatory compliance, infrastructure risk.

Problem	Most tools	Graphite
Unverifiable claims	Assertions without audit trail	Every claim carries Provenance: source, quote, confidence
No trust scoring	Binary pass/fail	Explainable ConfidenceScorer with named factors
Opaque pipelines	Fragile chains, no persistence	ClaimStore with SQLite-backed registry + search

Core primitives

Primitive	What it does
Claim	The atomic unit of trust — structured assertion with provenance
ClaimStore	SQLite-backed registry for persisting and querying claims
Provenance	First-class evidence source: document, quote, confidence
ConfidenceScorer	Explainable confidence scoring with named factors
BaseFetcher / BaseExtractor	Plugin interface for domain-specific extraction

What belongs where

Graphite (engine)	Your application
Claim schemas & enums	Domain-specific extractors
Claim registry (ClaimStore)	Extraction prompts & calibration
Confidence scoring	Verification logic
Plugin interfaces (Fetcher, Extractor)	UI, API, alerts

---

Also included: graph assembly & propagation

Graphite also provides graph assembly and shock propagation for supply chain analysis:

from graphite import GraphAssembler, top_k_paths_from_source, build_blast_radius

edges = my_extractor.extract(documents)
G = GraphAssembler().assemble(edges)

paths, _ = top_k_paths_from_source(G, source="country:CD", max_hops=3, k=3)
blast = build_blast_radius(paths, k=3)

See examples/toy_battery_demo/ and examples/flood_replay_demo/ for end-to-end examples.

---

Built with Graphite: EdgarOS

EdgarOS is the first commercial application built on Graphite — a verification engine for AI-generated financial research memos against SEC evidence.

• 📄 Claim verification — every memo claim graded against SEC filings
• 🔍 Two-axis verdicts — grounding level + evidence conflict detection
• ⚖️ Explainable scoring — provenance-traced confidence with reason codes

Building on Graphite? Open an issue to get featured.

---

Benchmark

In EdgarOS benchmark runs built on Graphite primitives:

Metric	Score
Conservative Precision	90.5%
Weakest-Link Top-1 Accuracy	69.0%
False Contradiction Rate	0.0%
Neutral Precision	91.7%

^{100 synthetic memos, 320 claims, evaluated against gold-labeled verdicts. Full methodology and eval harness in the EdgarOS repo.}

---

Optional extras

Core (always included): networkx + pydantic

pip install -e ".[llm]"     # Gemini structured extraction
pip install -e ".[neo4j]"   # Neo4j graph storage
pip install -e ".[pdf]"     # PDF parsing
pip install -e ".[all]"     # Everything

---

License

Apache-2.0 — see LICENSE.