Adversarial due diligence engine you own
Test your hypothesis on your
private haystack
Lodestone simulates an adversarial AI trial to test high-stakes hypotheses against your company's private data. Get a verdict in minutes with full evidence traceability. Deploy it completely offline or in the cloud.
00 · The status quo
Billion-dollar decisions still rely on
manual document review.
The needle you miss breaks the deal.
The problem
Teams of junior analysts spend weeks combing through thousands of contracts, filings, and reports looking for the one clause that kills a deal. They miss things. They know the other side might find what they didn't. Every overlooked red flag becomes a liability that compounds after close.
75% → 20%
Attorneys believed they had retrieved 75%+ of relevant documents; in reality they found about 20%.
Roitblat et al., Document Categorization in Legal Electronic Discovery
35 → 88
Average length of an M&A contract, in single-spaced pages, has more than doubled over twenty years.
Coates, J. C., Why Have M&A Contracts Grown? Evidence from Twenty Years of Deals
Not the solution
Keyword search… seriously? AI chatbots hallucinate, their context window is too small, and there is no provable way to trust the output. The other option is wrestling legacy software with a brutal learning curve and features that don't justify the pain.
01 · Interface
Hypothesis · preparing evidence
Target company revenue exceeds guidance once deferred contracts are recognized.
Run Overview
Normalize hypothesis: Rewritten as a falsifiable claim.
Retrieve evidence: 128 chunks across 4 files.
Extract claims: Extracting claims...
Filter relevance
Score coherence
Assemble feeds
Blue
round 1 · 2 claimsRound 1/2
Round 2/2
Report
Revenue upside is plausible if deferred contracts convert on the signed schedule.
Claim B1
Deferred revenue schedules support upside once backlog converts.
Claim B2
The side letter narrows termination triggers instead of widening them.
Red
round 1 · 2 claimsRound 1/2
Round 2/2
Report
Recognition is still conditional, and the unsigned appendix keeps the upside case attackable.
Claim R1
Recognition still depends on unsigned appendices and counterparty approval.
Claim R2
Board minutes imply unresolved indemnity exposure after close.
Judge
pendingDisposition
Judge is waiting for the debate rounds to finish.
Final Verdict
Awaiting Blue and Red round 2 before judgement.
02 · Why Lodestone
Three properties set Lodestone apart from off-the-shelf RAG: adversarial reasoning, deployment optionality, and scale that holds as your corpus grows.
Adversarial research
A Red agent and a Blue agent respectively attack and defend your hypothesis. A judge provides a verdict with precise document citations supporting each argument.
Offline or in the Cloud
Choose an air-gapped on-premise deployment if working with NDA-bound or classified documents. Choose a cloud-based one if you need flexibility.
Thousands of documents
Lodestone's architecture is built for scale. All the core features experience no slowdown as the document corpus grows.
03 · How it works
Trial
Ingest documents, make an hypothesis, let adversarial agents test it against your corpus
Ingest
Drop in contracts, filings, or reports. Lodestone parses, chunks, and embeds every document.
Hypothesis
State a claim about your documents and let Lodestone gather the documents that support or refute it.
Debate
Using the relevant documents, the Red Agent attacks the hypothesis while the Blue one defends it across multiple rounds.
Verdict
A Judge weighs both sides and delivers a structured ruling: confidence score, cited evidence, and dissenting arguments.
Chat with documents
Ask a question, let the agent search the corpus, then audit the cited sources.
Ask
Ask a question over the whole corpus or just a selection of files.
Agent searches
Agent searches through the documents, finding the most relevant and delivering a response with precise citations.
Audit
Check the relevant sources, by clicking on the agent's citations to open the original passages.
Data ownership
Choose where Lodestone runs. Every deployment offers the same features.
Air-gapped
Complete network isolation. Everything runs offline with local models.
- Local models only
- Classified environments
- Document secrecy
Hybrid
The middle ground. Run the core pipeline on your infrastructure while offloading specific tasks.
- Granular control
- Mix local & remote models
Cloud
Managed deployment on our infrastructure. We handle uptime, scaling, and updates.
- Managed infrastructure
- Auto scaling
- Low maintenance