Evidence
NUMBERS WITH RECEIPTS.
PUBLISHED, NOT PROMISED.
Every figure below is from a public source. If we can't cite it, we don't claim it.
Straight talk: these aren't our client logos. We're a young studio. The results below were published by the teams who pioneered each pattern — Anthropic, Harvey, Buildots, Klarna — not by us. We show them because they prove the category works, and because we build on the exact same architecture. When we have our own numbers to show, they'll be here too.
retrieval-failure reduction with Contextual Retrieval
Anthropic, Sept 2024
real-query coverage with hybrid search (vs 62% dense-only)
Superlinked benchmark
saved per contract review at A&O Shearman
Harvey case study
in court sanctions for AI-hallucination filings (Q1 2026)
1,348 documented cases
end-to-end success of a 10-step chain at 85% per-step (Lusser's Law)
Why we cap chains at ≤6 steps
Note on the Klarna number: their early 2024 framing of “AI replaced 700 humans” was walked back by mid-2025 as they rehired for complex cases. The $40M operational saving still stands. We scope agents to repeatable workflows and route edge cases to humans — that's the difference between a deployment that holds and one you walk back.
What we build, per industry
Four industries. One architecture.
For each one: what we'd build, and the public benchmark that proves the pattern holds. The before/after figures are the published source's — not ours.
92 minutes on 5 NDAs → 26 seconds. 94% accuracy vs 85% human.
Agents for contract review, due diligence across thousands of documents, and precedent search. RAG over your contract corpus — no hallucinations, every answer cited.
What we build
- Contract review (NDA, MSA, loan agreements)
- M&A due diligence over 1000+ contracts
- Precedent search with citation-grounding
- Draft generation for memos and briefs
M&A due diligence
Partner write-downs
Contract review accuracy
Sources: LawGeex · Thomson Reuters 2025