consistency-checker 0.4.1

Detects internal contradictions across a corpus of documents using a two-stage NLI + LLM pipeline.

contradictionchecker

Scans a corpus of documents for internal contradictions and divergent definitions of the same term. Designed for symmetric scans across a static corpus, not for guarding an existing knowledge base against new documents. The definition-inconsistency detector is on by default; the pairwise NLI + LLM detector is opt-in via --pairwise (ADR-0015).

v0.3 — FastAPI + HTMX web UI, three-document conditional contradictions (graph triangles), numeric short-circuit, and PDF/DOCX loaders. See CHANGELOG.md for the full feature list, futureplans.md for what's next.

What runs by default

The definition-inconsistency detector runs by default: it groups assertions by canonical term and asks the LLM judge whether same-term definitions diverge across the corpus. This is the detector that carries the headline value on every corpus shape we've measured.

The pairwise contradiction detector (NLI gate → LLM judge on candidate pairs) is off by default as of ADR-0015 — own-corpus eval on legal prose showed near-zero useful yield at high compute cost. Enable it per run with --pairwise, or set pairwise_enabled: true in config.yml. When pairwise is enabled, the two-stage flow is:

Stage	Component	Role
A	NLI checker (microsoft/deberta-v3-large-mnli family)	Cheap bidirectional contradiction score. Gates candidate pairs to Stage B.
B	LLM judge (Anthropic Claude or OpenAI, structured JSON output)	Verifies with rationale and evidence spans.

A single LLM-only check has been benchmarked at ~16% precision on pairwise contradiction detection in domain text (legal/financial). The NLI gate lifts precision to ~89% while cutting LLM cost roughly an order of magnitude — so when pairwise is the right detector (numeric-/spec-heavy corpora), the two-stage flow is the cheap way to run it.

See docs/decisions/0015-pairwise-opt-in.md for the rationale, and docs/ARCHITECTURE.md for the full module breakdown.

Install

From PyPI:

pipx install consistency-checker   # isolated CLI install (recommended)
pip install consistency-checker    # or into the current environment

Heads-up — heavy first install. The first install pulls torch, faiss, and unstructured, so it is multi-GB and can take several minutes. Model weights (sentence-transformers / DeBERTa) download on the first check / --pairwise run, not at install time.

From source:

git clone https://github.com/toddaerickson/contradictionchecker
cd contradictionchecker
uv sync

Corporate / sensitive-data users: read docs/corporate-setup.md first. This tool sends every document chunk to a third-party LLM API — confirm that's allowed by your data-classification policy before running it.

Quickstart

# 1. Scaffold a working directory (writes config.yml + a .env template here)
consistency-check init
# From source instead of a pipx install, use: cp config.example.yml config.yml

# 2. Put your API key in the .env that `init` created (NEVER in config.yml —
# config.yml is safe to commit; your key is not). The default provider is
# Moonshot/Kimi; edit config.yml's judge_provider for anthropic | openai.
# .env:  MOONSHOT_API_KEY=...       (or ANTHROPIC_API_KEY / OPENAI_API_KEY)

# 3a. Web UI flow (single-page UI, ADR-0017)
uv run consistency-check serve --open    # browser opens to http://127.0.0.1:8000
# Create a corpus and add files via [+ New corpus] in the sidebar → click
# [Run check] (toggle Deep for three-document conditional contradictions;
# live progress shows on the corpus row). Findings stream in the main pane:
# mark verdicts inline, filter with the chips, open the Assertions /
# Definitions / Stats drawers to drill in, or Export CSV.

# 3b. CLI-only flow
uv run consistency-check ingest path/to/corpus/
uv run consistency-check estimate-cost               # rough API-spend ceiling before you commit; per-call defaults now follow your configured judge_provider (Moonshot/Kimi projects sub-cent — ~$0.0001–$0.001 per call — vs Anthropic/OpenAI ~$0.003–$0.010)
uv run consistency-check check                       # add --pairwise for the NLI gate (off by default — see ADR-0015); --deep for triangle pass (requires --pairwise); --no-definitions to skip the definition stage; --max-cost <USD> aborts before judge bootstrap when the projection exceeds the ceiling (ADR-0016)
uv run consistency-check report                      # writes data/store/reports/cc_report_<ts>_<run_id>.md
uv run consistency-check export csv                  # writes data/store/reports/cc_assertions_<ts>.csv

The export command emits (doc_id, assertion_id, assertion_text) rows for downstream tooling. --out is optional for both report and export; omit it and the file lands under <data_dir>/reports/ with a unique descriptive name. The first check --pairwise run downloads a ~440 MB DeBERTa-base NLI model from Hugging Face (or ~1.5 GB if you opted up to DeBERTa-v3-large via nli_model in config); subsequent runs hit the cache. Default check runs (definition detector only) do not download or load the NLI model — same one-line download warning pattern as the OCR fallback (ADR-0014).

Vendoring HTMX

The web UI ships with a placeholder htmx.min.js. After cloning, run once:

uv run python scripts/vendor_htmx.py

to download HTMX v1.9.12 into consistency_checker/web/static/. Tests use FastAPI's TestClient which doesn't execute JS, so CI doesn't need the real script.

Other CLI commands

uv run consistency-check serve --host 127.0.0.1 --port 8000  # launch the web UI
uv run consistency-check store stats                          # row counts
uv run consistency-check store rebuild-index                  # regenerate FAISS from SQLite
uv run consistency-check --help                               # all commands

Benchmarks

benchmarks/contradoc_harness.py runs Stage A + Stage B against a normalised CONTRADOC dataset and reports precision / recall / F1. The dataset is not redistributed; see docs/benchmarks.md for the input format and runbook.

uv run python -m benchmarks.contradoc_harness \
    --input contradoc.jsonl --output metrics.json --sample 50

Development

uv sync
uv run pytest -m "not slow and not live"   # default CI gate
uv run pytest -m slow                      # downloads HF models (~800 MB - 1.5 GB)
uv run pytest -m live                      # hits Anthropic / OpenAI APIs
uv run ruff check .
uv run ruff format --check .
uv run mypy consistency_checker

See CONTRIBUTING.md for branching, PR conventions, and the dev loop.

Architecture decisions

Recorded as ADRs in docs/decisions/:

• 0001 — LLM judge provider: both Anthropic and OpenAI, behind a Judge Protocol.
• 0002 — Embedding model: sentence-transformers/all-mpnet-base-v2 default.
• 0003 — CONTRADOC integration timing: in MVP scope.
• 0004 — PDF/DOCX loader backend: unstructured for both.
• 0005 — Numeric short-circuit before the LLM judge: deterministic sign-flip detector skips the judge.
• 0006 — Three-document conditional contradictions: graph triangles on FAISS-similarity edges, opt-in via --deep.
• 0007 — Web UI: FastAPI + HTMX, server-rendered, cc_-prefixed templates.
• 0015 — Pairwise contradiction detector becomes opt-in: pairwise off by default; enable per run with --pairwise or pairwise_enabled: true.
• 0016 — Pre-flight cost ceiling for check: --max-cost <USD> / max_cost_usd aborts before judge bootstrap when the conservative projection exceeds the ceiling; estimate-cost defaults per-call costs from the configured judge_provider (Moonshot ~10–100× cheaper than Anthropic/OpenAI).

Supported formats

Extension	Loader	Notes
.txt, .md	built-in plaintext loader	char spans round-trip exactly
.pdf, .docx	unstructured (strategy="fast")	body-content elements only; sidecar element_spans in documents.metadata_json

Scanned-image PDFs are auto-escalated to unstructured's hi_res (OCR) strategy when fast extraction returns near-empty text. First OCR run downloads ~500 MB of layout + OCR models. Requires system Tesseract (apt install tesseract-ocr on Debian/Ubuntu, brew install tesseract on macOS).

Other formats can be added via the LOADERS registry in consistency_checker/corpus/loader.py.

Known limitations

Carried forward into the v0.4+ roadmap in futureplans.md:

• Chunk overlap > 0 is unimplemented.
• Three-document detection misses triangles whose edges fall below the FAISS gate threshold; v0.4 #6 adds an entity-NER cluster pass to catch these. Three-document detection requires --pairwise since it shares the NLI gate's output.
• First check --pairwise run downloads ~440 MB for the default NLI model (DeBERTa-v3-base). Switch to DeBERTa-v3-large via nli_model in config for higher recall at ~1.5 GB. Default check runs (pairwise off, see ADR-0015) skip the download entirely.
• OCR fallback is automatic for image-only PDFs (--no-ocr to disable); first use downloads ~500 MB and requires system Tesseract.
• data_dir/uploads/<upload_id>/ grows without bound; v0.4 will add a GC pass.
• The web UI is single-user, localhost-only, with no authentication or CSRF protection. serve refuses to bind to a non-loopback host unless you pass --unsafe-no-auth; doing so exposes an unauthenticated file-upload and corpus-mutation surface to anyone who can reach the host, so only use it on a trusted, isolated network.

License

Apache-2.0. See LICENSE.