Governed contract extraction
Contract obligations your systems can act on, each tied to the sentence that states it.
DocChunker reads contracts for the terms that cost money when missed: dispute windows, notice periods, liability caps, covenants. Every value is bound to its exact source in the PDF. Anything that cannot be verified is blocked, and only a named reviewer can promote a value into your systems of record.
“Any billing dispute must be submitted in writing within sixty (60) days of the invoice date”
Page 5, characters 4170–4261. Found in the PDF and it states the value, so a reviewer may approve it.
“In the event of a Force Majeure declared by Carrier, the standard dispute window of 60 days is reduced to 30 days”
This sentence is not in the contract. With no source to check, the value cannot be approved by anyone.
Computed from the PDF when the server starts, not typed in. See every binding.
The short answer
Measured on three fictional sample contracts (a carrier MSA, a HIPAA business associate agreement and a credit agreement), each with a hand-built answer key.
Found in the PDF, page and position derived, value confirmed in the sentence.
An invented clause, a misread number, a combined figure and a paraphrased quote.
The server refuses the approval. The rule does not depend on the screen.
Against 0.07 for pattern rules on the same 24 values (1 of 3 contracts run so far). Recorded run, 26 Sept 2026, 15:58 UTC.
Method and per-contract results: Evaluation. Raw data: evaluation, binding audit.
Why contract data needs a control layer
Deadlines hide in clauses
A carrier agreement gives you 60 days from the invoice date to dispute a bill; after that the invoice stands. Notice periods, breach-reporting windows and covenant tests work the same way. The money is lost quietly, on a date nobody was tracking.
An untraceable value is not evidence
A model can read a contract. What an auditor, regulator or counterparty asks is where a number came from. A value with no source sentence cannot be defended, so today it gets re-keyed by hand.
Review does not scale when it cannot be trusted
If reviewers cannot see the sentence behind each value, they re-read the whole contract. The extraction saved nothing, and the review left no record of what was checked.
How it works
The model proposes. Code checks. A person decides. Confidence scores order the queue; they never approve anything.
- 1
Extract
A small model reads the contract against a fixed schema and returns each value with the words it came from.
- 2
Bind
Code, not the model, finds those words in the PDF, derives page and character position, and checks the value is actually in them.
- 3
Gate
A value that fails binding cannot be approved or edited into the record. It can only be rejected or sent to a senior reviewer.
- 4
Decide
A named reviewer claims one value at a time, sees the sentence highlighted on the page, and approves, corrects, rejects or flags it.
- 5
Record
Each decision is a linked, hash-chained event. Anyone can re-verify the chain; only approved values reach downstream systems.
What happens to a wrong answer
These are the mistakes extraction models actually make, planted in the sample workspace. Each one is stopped by the binding rule before a reviewer can approve it, including the one where the value happens to be right.
Dispute Window: 30 days
“In the event of a Force Majeure declared by Carrier, the standard dispute window of 60 days is reduced to 30 days”
That sentence is not in the document.
Late Payment Rate: 15% per month
“Late payments shall accrue interest at the rate of 1.5% per month”
The sentence is real, but it does not state this value.
Breach Notification Window: 72 hours
“Business Associate must notify the hospital of any breach within 72 hours”
That sentence is not in the document.
Collateral Value: $53,000,000.00
“Real Property Collateral"), valued at $35,000,000.00”
The sentence is real, but it does not state this value.
Who it is for
Head of Legal or Contract Operations
A contract data record that shows its source for every term, and a review queue sized to what is actually uncertain.
Procurement and telecom expense owners
Dispute windows, activation deadlines and SLA credits as dated obligations, before the window closes.
CISO and AI governance
A model that cannot promote its own output, a deployment path that keeps text in-network, and an audit trail that can be re-verified.
Deployment and controls
What runs in this build, and what an enterprise deployment adds.
- Built
Source binding for every value
Each value must come with a quote that is found in the parsed PDF and contains the value. Page and character offsets are derived from where the quote sits, never taken from the model.
- Built
Approval gate enforced by the API
A value without a verified source cannot be approved, and an edit can only correct a value to what the source says. The server refuses the request; hiding a button is not the control.
- Built
One value per decision, claim ownership
No bulk approval. A reviewer claims a value before deciding it, and only the claim holder can approve, edit, reject or flag it.
- Built
Hash-chained audit trail
Every review decision is a SHA-256 linked event, and the whole chain can be re-verified on demand. The transition and its event are one write.
- Built
Highlighted source on the page
Selecting a value opens the PDF at the right page with the bound sentence highlighted in the text layer.
- Built
Measured accuracy
Field-level precision, recall and calibration against a hand-built answer key for each sample contract, recomputed on every read, with the pattern-rule baseline alongside.
- Built
Model health and spend ceiling
Pinned model versions with a fallback list, checked against the provider catalogue for retirement dates, and a durable daily ceiling on paid calls. Over the ceiling, extraction falls back to pattern rules and says so.
- Built
No document retention in the demo
Uploaded PDFs are parsed in memory for one request and discarded. Review actions live in the visitor’s own signed session for 24 hours.
- Roadmap
Single sign-on and roles
SAML/OIDC identity, reviewer and approver roles, and four-eyes approval on high-value fields. The demo gives each browser session its own reviewer identity instead.
- Roadmap
Scanned contracts
OCR, with values bound against the OCR text layer the same way. The demo accepts PDFs with selectable text and says so when it receives a scan.
- Roadmap
Private deployment
A small open-weight model inside the customer’s network, so contract text never leaves it. The demo calls a hosted model through OpenRouter (a small open-weight model first, with pinned fallbacks) and shows which one answered.
- Roadmap
Connectors, amendments, export
Ingest from SharePoint, S3 and CLM systems; resolve amendments against the base agreement; push approved values to ERP or expense systems by webhook.
Why now
- Gartner published its first Magic Quadrant for Intelligent Document Processing solutions in September 2025, naming ABBYY, Hyperscience, Infrrd, Tungsten Automation and UiPath among the Leaders. (Hyperscience, ABBYY)
- Amazon Textract sends low-confidence fields to human reviewers through Amazon Augmented AI. Google deprecated Document AI's Human-in-the-Loop feature on January 16, 2024. (AWS docs, Google Cloud docs)
- LLM-native extractors now cite their sources: LlamaExtract returns a citation and reasoning for each field (May 2025), and Unstructured offers a schema-based structured data extractor. (LlamaIndex, Unstructured docs)
Extraction is becoming a feature every platform ships. What a regulated buyer still has to build is the control around it: a value is admissible only if its source sentence is found and states it, a named person makes each decision, and the record can be re-verified by someone who does not trust the vendor. DocChunker is that layer, and it governs any extractor's output, including the pattern rules measured above.
Vendor names describe the market from each company's public materials, linked above. DocChunker is an independent product concept and is not affiliated with, endorsed by or compared in testing against any of them. The sample contracts and every party in them are fictional.
See it on a contract
Run a sample, or bring a PDF with selectable text. Nothing you upload is stored.