Pre-launch · Raising
In the last year the major eDiscovery platforms folded generative-AI review into their base pricing. The capability stopped being a differentiator. At the same time, courts began restricting which AI may touch confidential evidence at all — and the honest answer to that restriction is architectural, not contractual. That is the company.
The application
Most platforms give you a query box and a results list. Evidentiary.ai turns the corpus into pictures a trial team can reason about — and every picture drills down to the underlying documents.
The Concept Cluster Wheel organises the entire corpus into a radial hierarchy of themes — fraud allegations, compliance, damages — sized by how much of the record they cover. Click a segment to isolate its passages and jump straight to the documents behind it.
Key events, communications, meetings, and filings on one interactive timeline — flagged by severity, linked to custodians, and tied to the related documents. The narrative your opening statement needs, assembled from the evidence itself.
Metadata inconsistencies, off-hours bursts, communication drops, tone shifts, privilege gaps — anomaly signals from every analytic, aggregated into one triage queue with a custodian-by-type heatmap. Spoliation patterns stop hiding in the volume.
Search activity, sentiment alerts, and key events at a glance, with the full toolkit one click away: advanced search, email network graphs, keyword analytics, entity extraction, and natural-language inquiry with cited answers.
Screenshots are the live product rendering a fictional demonstration matter — not mock-ups.
What actually exists
The most common failure mode in this category is a demo with a services company behind it. What follows is measured from the build itself, and stated with its limits.
Identification through in-court presentation, on one case file and one audit ledger — not an integration of separate products.
Verified by automated HTTP probes against running services, not by a checklist. The remaining seven are front-end and library functions with no service endpoint to probe.
Across 26 architecture layers. Document processing, OCR, and rendering are built in-house in Rust rather than outsourced — the layer most review platforms buy or partner for.
Figures are drawn from the build's own manifest and its most recent automated validation run. The platform is pre-launch and in active development. We publish no accuracy or throughput benchmarks, because we have not yet run and validated them to a standard we would defend in a deposition. That work, and its current honest state, is covered in the pitch.
Why now
Each is a matter of public record. Together they describe a market that has commoditised the feature we are not selling, and opened a gap we are.
Between late 2025 and early 2026 the leading platforms moved generative-AI review into all-inclusive per-gigabyte pricing. Selling "AI review" as a premium feature is over. Whoever competes on that axis alone is competing on price.
The first significant ruling on generative-AI review held it to be a form of technology-assisted review under the existing rules, rather than a novel methodology deserving heightened scrutiny. The bar to adoption came down, not up — which accelerates the whole category.
Courts have begun restricting the use of mainstream AI tools on confidential material unless the provider is barred from retaining or training on what it sees. Cloud vendors answer that with a contract. Only local architecture answers it structurally.
The wedge
Relativity has publicly announced that its on-premises Server product will accept no new matters after January 1, 2028. The firms, agencies, and defense teams that chose on-premises deployment for a reason — classified work, sealed matters, confidentiality obligations that a contract does not satisfy — have a fixed deadline and no like-for-like destination. No major platform offers air-gapped generative-AI review today.
Their AI is architecturally bound to third-party cloud models. Serving this segment is not a feature they can ship; it is a rebuild of the thing their economics rest on.
We do not claim a better model. We claim the only place some evidence is permitted to be processed — on hardware the client controls, with no outbound inference.
Self-hosted models change the unit economics of AI review rather than the sticker price. What that permits commercially is covered in the pitch.
Sources for each of the above — rulings, vendor pricing announcements, and the underlying competitive research — are cited in full in the pitch materials.
Who is building it
The deployment constraint at the centre of this company is not a market theory. It is the daily operating reality of the largest civil litigation program in the United States — run, for over a decade, by the person building the platform.
Joshua Wood spent over two decades at the Department of Justice and over a decade as the Director of the Office of Litigation Support for the Civil Division — building and managing the most advanced litigation support environment in the federal government.
At DOJ, Josh oversaw a $90–$120 million annual operation spanning 500+ staff, a petabyte-scale data infrastructure, and twelve FISMA-authorized systems. He served as a Department recognized subject matter expert on AI in litigation, delivering training programs on text analytics and predictive coding across DOJ's litigating divisions.
He served as the "Tier 3" E-discovery technical lead, advising attorneys and agencies, and was the technical member of DOJ's eLitigation Steering Committee. Evidentiary.ai is the product of everything he learned building litigation technology at that level — now available to every litigation team that demands defensible, auditable AI.
Awards
Selected speaking
Defensibility
Where we are
Pre-revenue and pre-launch. A founding cohort of litigation teams is forming now. The platform runs; the commercial proof does not exist yet, and we would rather say so here than have you discover it in diligence.
The full discovery lifecycle, the strategy layer, the forensic engine, and self-hosted inference — running today on real public corpora.
Benchmark validation, the packaged validation artifact, security certification groundwork, and practitioner validation of the criminal-procedure workflows.
Published accuracy figures, customers, revenue, or security certifications. Anyone claiming otherwise at this stage is describing a roadmap.
Analysis · From the public record
In March 2025 the FBI reviewed roughly 100,000 Epstein-related records the way reviews were run in 1985 — by hand, around the clock, on end-of-life software — in a trove that would ultimately run past six million pages. Forty years of eDiscovery research predicted the result. This is a sourced analysis of how that review was run, and why it failed.
This review draws only on the public record — Congressional correspondence, internal FBI emails released to reporters, court filings, and the government's own releases. Four findings recur across those sources:
Linear, page-by-page human review at scale — the exact approach empirical research discredited four decades ago.
Review software bought in 2011 (~$11.9M, per procurement records) — then bought again in 2021–22 ($4M more) instead of being replaced.
No stated recall target, sampling protocol, or quality-control metric — no way to know what the review missed.
Machine translation, handwriting OCR, and entity resolution exist in unclassified form — and were absent from the process.
The idea that armies of human reviewers are the "careful" way to work through a document population was tested — and disproven — four decades ago, in a study built on litigation over a Bay Area Rapid Transit (BART) accident. Every rigorous test since has widened the gap. Any critical review of the FBI's process has to start here, because none of this was obscure: it is the settled, judicially recognized baseline of the field.
David Blair and M.E. Maron studied the discovery database from the BART litigation — roughly 40,000 documents, about 350,000 pages. The attorneys running the review were confident they had retrieved more than 75% of the relevant documents. Measured against the actual population, they had found about 20%. Human judgment about review completeness was off by a factor of nearly four — and the reviewers never knew.1
Share of relevant documents retrieved in the BART litigation review (Blair & Maron, 1985)
Grossman & Cormack, analyzing TREC 2009 Legal Track data in the Richmond Journal of Law & Technology, showed that technology-assisted review (TAR) can be more effective and more efficient than exhaustive human review, on both recall and precision.2
In Da Silva Moore v. Publicis Groupe (S.D.N.Y., Feb. 24, 2012), a federal court judicially accepted TAR for the first time, citing the superiority of computer-assisted review over the alternatives.3
Modern review layers large-language-model analysis on top of TAR: entity resolution across aliases, timeline construction, machine translation, and handwriting OCR — capabilities that once existed only inside classified intelligence systems and are now available in unclassified form to any agency or litigant that chooses to use them. None of it requires a clearance — only the decision to use it.
Finding 1 — The methodology was the one discredited in 1985. In March 2025, roughly 1,000 FBI Information Management Division personnel were put on 24-hour shifts to comb through approximately 100,000 Epstein-related records by hand, supplemented by hundreds of New York Field Office personnel — many without training in identifying statutorily protected victim information or handling FOIA material.4 That is linear manual review: the approach whose recall Blair & Maron measured at roughly 20%, and which both the research literature and the courts have treated as inferior to technology-assisted review since 2011–2012.
Finding 2 — The tooling predates the modern field. The software behind FBI eDiscovery review is the Clearwell platform, a product of the late 2000s that the Bureau adopted in 20115 — and whose product line, after passing from Clearwell Systems to Symantec (2011) to Veritas, has since reached end-of-life. Federal procurement records show the purchase: in September 2011 the FBI obligated roughly $11.9 million for "CLEARWELL SOFTWARE" across three reseller awards — $1,986,794 via ThunderCat Technology (PIID DJFA1D104061, Sept. 9), and $4,714,875 plus $5,177,960 via Software Information Resource Corp. (PIIDs DJFA1D1131101 and DJFA1D1131102, Sept. 29–30, under IDC DJFJFBI11311; the first order may duplicate the parent vehicle's reported value).8 What followed was maintenance, not modernization: a $30,956 upkeep order in March 2016 (PIID DJF161200D0004095), and then — a decade after the original buy — a $4.04 million order for a "60 TB license, Clearwell/Veritas software" for the BIDMAS program ($2,037,000 on Feb. 24, 2021, plus a $2,000,000 option exercised Feb. 16, 2022; PIID 15F06721F0000517).8 In other words, when the platform aged out, the Bureau's response was to buy more of it. The procurement record shows no FBI award for any modern review platform — Relativity, Everlaw, Casepoint, or Exterro — only small forensic-tool purchases (Nuix, Intella) in the $875–$15,000 range. A review run in 2025 on a platform bought in 2011 cannot apply fifteen years of accumulated method.
FBI obligations on Clearwell/Veritas review software by calendar year, from FPDS8 — the empty years are real
| Date | What the FBI bought | Vendor | PIID | Obligated |
|---|---|---|---|---|
| Sept 9, 2011 | "CLEARWELL SOFTWARE" | ThunderCat Technology | DJFA1D104061 | $1,986,794 |
| Sept 29, 2011 | "CLEARWELL SOFTWARE" (order under IDC DJFJFBI11311) | Software Information Resource Corp. | DJFA1D1131101 | $4,714,875 |
| Sept 30, 2011 | "CLEARWELL SOFTWARE" | Software Information Resource Corp. | DJFA1D1131102 | $5,177,960 |
| June 25, 2012 | "CLEARWELL TRAINING" | Symantec Corp. | DJFA2DOE0030 | $3,325 |
| Mar 7, 2016 | "CLEARWELL SYMC SOFTWARE MAINTENANCE" | Software Information Resource Corp. | DJF161200D0004095 | $30,956 |
| Feb 24, 2021 | "EDISCOVERY 60 TB LICENSE, CLEARWELL/VERITAS SOFTWARE FOR BIDMAS PROGRAM" | ThunderCat Technology | 15F06721F0000517 | $2,037,000 |
| Feb 16, 2022 | Option exercise on the same 60 TB Clearwell/Veritas license | ThunderCat Technology | 15F06721F0000517 P00003 | $2,000,000 |
| 2015–2022 | Small forensic tools only (Nuix licenses, Intella Pro) — no review platform | Nuix, Vound, resellers | various | $875–$15,000 ea. |
The FBI's eDiscovery purchase history, from the Federal Procurement Data System.8 Fourteen years of records, one product. No FBI award names Relativity, Everlaw, Casepoint, Exterro, or any other modern review platform — the highlighted rows show that in 2021–22, a decade after the original buy, the Bureau's response to an aging platform was to purchase more of it. (The Sept. 29 order may duplicate its parent vehicle's reported value; keyword-based FPDS search can also miss awards with unusual descriptions.)
Finding 3 — Nothing was measured. The central lesson of the BART study is that reviewer confidence is not a metric: the attorneys who had found 20% believed they had found 75%. Nothing in the public record describes a recall target, a sampling protocol, or any quality-control measurement for the FBI's review. The observable outcome matches the prediction: the government has since acknowledged fixing thousands of documents in the released files that may have contained victim information.7
Finding 4 — Available capabilities went unused. Machine translation, handwriting OCR, entity resolution, and image analysis exist in unclassified form and are in routine use across government and private litigation. Internal FBI emails from the 2025 review instead show personnel asking for capabilities their tooling simply lacked — including help identifying public officials and celebrities in images.6 The gap was not one of clearance or availability, but of process.
Postscript — January 2026: the corpus was millions of pages, and full of duplicates. The March 2025 hand review covered roughly 100,000 records. When the Epstein Files Transparency Act forced a full accounting, the real scale arrived: Deputy Attorney General Todd Blanche described a review of "over six million pieces of paper, thousands of videos, tens of thousands of images," and the Department published more than 3.5 million responsive pages — the final ~3 million of them beginning January 30, 2026.9,10 Days before that release, the U.S. Attorney for the Southern District of New York told a federal court that roughly one million newly identified FBI records were still surfacing — and that a "meaningful portion" may be duplicative of records already collected, yet "nonetheless still need to undergo a process of processing and deduplication."11 Deduplication at this scale is an automated, ingestion-time step in modern review platforms; done by hand, every duplicate gets reviewed twice. And the measurement failure the BART study predicts repeated at scale: despite hundreds of reviewers, the Department acknowledged that unredacted victim information survived into roughly 0.1% of released pages, and survivors' counsel identified at least 31 people victimized as children whose identities were not protected.10
The conversation
Competitive analysis with citations, the patent portfolio, unit economics, the go-to-market, and a live walkthrough of the platform on a real corpus — shared directly, under NDA where appropriate. Tell us who you are and we'll set it up.