Investor brief

High-stakes work is being checked by tools that cannot be trusted to be right.

Perasys Labs builds local-first verification for work where a wrong answer is expensive and the material cannot leave the building. The first wedge is the one the market is already paying for in fines and sanctions: verifying that legal citations are real. The same engine, decompose a claim, check it against authority, write a record you can audit, is not specific to law.

This page describes the business. It is information, not an offer to sell or a solicitation to buy any security. Detailed financials are shared privately with qualified investors under NDA.

Thesis

Capability went up. Trust did not.

General-purpose AI can produce fluent, professional-looking work in seconds. In domains where being wrong has consequences, fluency is the problem, not the solution: confident output that looks correct is harder to catch than output that looks broken. The bottleneck has moved from producing work to verifying it.

Most of the market is answering this with bigger models and more cloud. We think that is the wrong axis. For privileged, regulated, or confidential work, the constraint is not model quality, it is verifiability and control: can you prove each claim is supported, and can you do it without sending the material to someone else's machine. Perasys builds for that constraint first. Capability is rented; trust is the product.

The problem, measured

The first market is not a forecast. It is a running tally with dollar figures attached.

Fabricated legal citations are the clearest, best-documented instance of the trust gap, because courts write down every failure. The evidence is independent of us and it is getting worse, not better.

58 to 88%
of answers from general-purpose LLMs hallucinate on specific, verifiable questions about real federal cases.
Stanford RegLab / HAI, 2024 1
17 to 33%
hallucination rate of the purpose-built legal tools from LexisNexis and Thomson Reuters, despite "hallucination-free" marketing.
Stanford RegLab, 2024 2
1,348
court and tribunal cases worldwide with documented AI-fabricated content by April 2026, up from 200 a year earlier.
Charlotin Database, HEC Paris 3
$110K
single-case sanction (Oregon, 2026), with new documented cases now appearing five to six times a day.
Court orders, 2026 3

The willingness to pay is not hypothetical. The cost of being wrong here is malpractice exposure, six-figure sanctions, bar discipline, dismissed cases, and reputational damage that has already reached Am Law 100 firms.4 When the downside is measured in sanctions and license suspensions, verification stops being a feature and becomes a line item.

Why incumbents do not close it

The companies best positioned to fix this have a structural reason not to.

THE GAP

Adding a model to a cloud database does not produce trust.

The dominant approach bolts a language model onto a proprietary corpus and ships it as a cloud service. Independent testing shows that even these purpose-built tools fabricate between 17 and 33 percent of the time. The architecture optimizes for answers, not for proof that the answer is supported.

And it requires the most sensitive work to be uploaded to a vendor, which is a non-starter for a large share of privileged matter.

OUR ANGLE

Verification as the primary object, on the operator's own machine.

CiteChain does not try to write the brief. It takes the brief and proves, claim by claim, whether each citation is real and supports what it is cited for, then writes a tamper-evident record of the check. It runs entirely on the user's computer, so the material never leaves the building.

That is a different product category from "AI that drafts." It is the layer that makes the drafting usable in court.

The wedge, tested against the hard questions

Legal citation verification is not chosen because it is large. It is chosen because it is urgent, budgeted, and forced.

Who has the pain
Litigators, solo and small firms, in-house counsel, and pro se filers, anyone who signs a filing and is personally accountable for its citations under Rule 11 and bar rules.
How often
Every filing that cites authority. The verification step is mandatory and repeated, not occasional.
What they do today
Manual cite-checking, or trust an AI tool that fabricates 17 to 33 percent of the time, or pay for incumbent research seats and verify anyway.
Why now
Regulation is forcing it. ABA Formal Opinion 512, a growing set of judicial standing orders, and new court-wide AI rules now require attorneys to attest they verified AI output. Verification is becoming a compliance obligation, not a preference.4
Why incumbents miss
Their architecture is cloud-and-corpus and their numbers prove the gap. The privileged-work segment also cannot upload, which a local-first tool serves and a cloud tool cannot.
What is actually defensible

The moat is not the model. It is the architecture and the record.

LOCAL-FIRST

The material never leaves the machine.

Privileged and confidential work is processed on the operator's own hardware. This is a capability cloud incumbents cannot match without abandoning their model, and it directly addresses the segment they cannot serve.

VERIFIABLE BY DESIGN

Every check writes an auditable record.

The output is not just an answer, it is a tamper-evident ledger of what was checked, against what authority, and what a human decided. That record is the artifact regulators and courts are starting to require.

SUBSTRATE-AGNOSTIC

The engine outlives any one model.

CiteChain is the first application on Logos, a local verification kernel. The model underneath is a swappable substrate. The durable asset is the verification layer, not a dependency on a single vendor's model.

Where this generalizes

The wedge is law. The architecture is not.

"Decompose a document into claims, check each claim against an authoritative source, and write an auditable record, locally" is a pattern, not a legal-specific feature. The same engine plausibly applies anywhere a claim must be verified against a source of truth under accountability. We treat these as hypotheses to validate after the wedge is won, not as committed markets or counted demand.

Hypothesis

Regulatory compliance

Checking filings, disclosures, and policies against the actual text of the controlling rule, on confidential internal material.

Hypothesis

Audit and assurance

Verifying that assertions in a report are supported by the underlying workpapers and source documents.

Hypothesis

Medical and clinical coding

Confirming that codes and claims map to documented evidence in the record, where errors carry liability.

Hypothesis

Financial disclosure

Checking statements against source data and prior filings before they are signed and filed.

Note on discipline: these are adjacent applications of one architecture, not evidence of demand in those markets. None is counted in any market sizing until a paying customer in that segment validates it. The expansion case is optionality earned by the wedge, not a TAM argument.

The numbers

The financial model is real and specific. It is shared privately, not posted publicly.

The full materials include the items below, with sourced benchmarks rather than aspirational figures. They are provided to qualified investors under NDA.

Available in the full materials
  • Unit economics and gross margin. Driven by a bring-your-own-substrate architecture.
  • Three-statement pro forma. Built on real benchmarks, with assumptions stated.
  • Go-to-market and pricing. The path into the legal wedge and early pipeline.
  • Use of funds and milestones. What the round buys and what it proves.
  • Round structure and terms. Reviewed with prospective investors directly.
  • Technical and security appendix. Architecture, the verification ledger, and the local-first posture.
Next step

If the thesis holds for you, request the full brief.

Tell us who you are and we will share the deck, the financial model, and a private CiteChain walkthrough. Materials are sent under NDA.