How Rivell reviews a contract.
Three engines share one knowledge base, and a person signs off between them.
Start here
Select any part of the system.
Dark boxes are the engines. Dashed boxes are where a person decides. The band below holds the four kinds of intelligence the engines share.
Input
Your knowledge
The playbooks, clause libraries, templates and past deals your team already keeps, in Word or Excel.
Engine
Knowledge engine
Reads each document with its headings, tables and rows intact. Tags every entry with its topic and the roles it states, such as preferred, fallback, walk-away or escalation. Flags entries that duplicate, contradict or replace what is already published.
You
You approve every entry
Nothing reaches a review until a person has checked it against its source and published it.
Store
Published knowledge
Versioned. Every review records the exact version it used, so a past review can always be explained.
Input
Incoming contract
A Word document on either side's paper. A file that already carries tracked changes is reviewed as if they were accepted. Settle them in Word before Rivell writes its redline.
Engine
Review engine
Reads the contract's structure, parties and your side. Gathers the positions that apply, judges each clause, checks across clauses, and turns the results into findings with a fixed severity.
You
You decide the findings
Accept, edit, keep as drafted, dismiss or escalate. Findings marked informational need no decision. You can also add a finding Rivell did not raise.
Engine
Redline engine
Turns each decision into an exact replace, insert or delete on the original document, written as a tracked change.
Output
Redlined Word file
Read back after it is written. Each change is checked against what you approved before the file is offered for download.
Intelligence
Rules
Fixed logic owns structure, the order of authority, severity and the final check of the file. Used by all three engines. The rules always give the same answer for the same input.
Intelligence
Jev, a System One decision model by TypeSafe
Answers fixed-choice questions about each entry and each clause, and says how likely each answer is. Used by the knowledge and review engines.
Intelligence
Language model
Turns free-form text into structured positions, drafts wording when nothing in your library fits, and gives a second opinion on close calls. Used by the knowledge and review engines.
Intelligence
Hybrid search
Searches by exact words and by meaning, then combines the results, so it finds the positions a clause touches even in different words. Meaning-based search needs a model you set up. Used by the review engine.
Follow one clause through the engine.
Scroll through the seven steps, or select one.
Kestrel Assurance · Master Services AgreementClause 12.2
Each party's aggregate liability under this Agreement shall not exceed three times the annual Fees.
Illustrative values for one question.
the Fees paid or payable in the 12 months before the claim
Approved organizational drafting · Clause library
Liability cap is above your limit
The clause allows three times the annual fees. Your playbook walks away above two times.
Any increase needs Marcus Feld and Priya Raghavan
12.2 Each party's aggregate liability under this Agreement shall not exceed three times the annual Fees the Fees paid or payable in the 12 months before the claim.
Rivell finds the clause in the document's structure and places it under a topic: liability cap. It already knows your side of the deal. You are the supplier.
It collects what applies, in order of authority: your playbook position first, then approved wording, then similar clauses from past deals. Search can surface a source. It can never promote one.
Jev compares the clause with the positions that apply, all at once. It does not write an essay. It picks an answer and says how likely each option is.
Rules look at how sure Jev is. A clear answer stands, so a clear breach becomes a finding. A close call goes to the language model for a second opinion. If the two disagree, the clause is marked for your look.
Rivell checks that your approved wording meets the position and closes this gap. Here it does. When nothing in your library fits, Rivell drafts new wording and labels it Proposed by Rivell.
You see the gap, the wording and who must approve. Accept, edit, keep the clause, or escalate it.
Your accepted change becomes a tracked change in the original Word file. Rivell reads the file back and confirms the change is exactly what you approved.
What each part does.
Rules
Fixed logic, no model
Owns the document structure, which source outranks which, how severe a finding is, and the final check of the Word file. Legal work needs the same answer for the same input, so severity never depends on how confident a model feels.
Jev
System One decision model by TypeSafe
Answers fixed-choice questions: yes or no, pick one, or rate. It answers all of a clause's questions at once and gives a likelihood for every answer. Its judgments are fast and consistent, and Rivell can see when it is unsure.
Language model
Your choice of provider
Turns free-form playbook text into structured positions, drafts wording when nothing in your library fits, gives second opinions on close calls, and reads the whole contract for issues your playbook does not cover. Some questions need language, not a choice.
Hybrid search
Exact words and meaning
Searches by exact words and by meaning at the same time, then merges the two result lists. Exact words miss a paraphrase, and meaning alone can miss a defined term, so together they find the position a clause touches. Meaning-based search runs once you choose a model for it.
You
The final word
You approve the knowledge, decide each finding that needs a call, and add what Rivell missed. Rivell assists. A lawyer signs off.
Why a decision model sits at the centre.
Most AI tools ask a language model to write an answer, then read a decision out of its prose. Rivell asks most of its questions of a System One decision model instead.
TypeSafe borrowed the name from System 1 in psychology, the fast and intuitive mode of thinking. The model answers many structured questions at once and gives a likelihood for every option, so rules can act on its doubt.
Is clause 12.2 acceptable under our liability cap position?
The clause appears to set a cap that is higher than many suppliers would accept, although this may be reasonable depending on the value of the contract and the wider commercial context. You may wish to consider...
Hard to act on. No clear answer, no measure of doubt.
Where does clause 12.2 fall against the liability cap position?
Ready for rules. A clear answer, and how sure it is. Illustrative values.
Authority comes before similarity.
Search finds clauses that look alike, but it cannot know which ones your team agreed to. Every source has a fixed rank, so a close match can inform a finding but never override your playbook.
- Past dealA clause you signed last yearContext only
- Clause libraryApproved liability cap wordingSupplies wording
- Your playbookLiability cap position, v4.2Decides
Built to be checked.
Every review can be explained
A review keeps the contract version and the knowledge version it used, so you can always see why it said what it said.
Changes are read back and checked
After it writes tracked changes, Rivell reads the Word file back and checks each one. A change it cannot place exactly is reported, never guessed.
- Contract
- v1, fixed for this review
- Knowledge
- Knowledge version 4
- 3 approved changes
- 3 applied
- 3 verified
What Rivell is built with.
Self-hosted. Contracts and knowledge stay on your own servers. Clause text goes only to the model providers you configure.
- Java 21 and Spring BootEngines and API
- PostgreSQL with pgvectorStorage and meaning-based search
- React and TypeScriptThe workbench
- docx4j and Apache POIReading and writing Word and Excel
- Jev by TypeSafeSystem One decisions
- A language model of your choiceReading and drafting text
See it on a real contract.
The demo follows a fictional software company through five contracts, from the first read to a checked redline. No sign-up needed.