We built the engine for doing to litigation what quant algorithms did to stock trading.
We manufacture deal flow at the trivial cost of running software.
A Blecher.ai project · Built by JDBlecher
Returns are uncorrelated to equities, rates, and real estate. Whether the economy is expanding or contracting, insurers still deny claims, businesses still breach contracts, and importers still misdeclare. And every claim gets its fair day in court.
A [$25,000 denied insurance claim] cannot justify a human lawyer billing $500–$1,000 an hour. We can bring those claims at near-zero marginal cost.
A [$549M customs-fraud recovery] justifies any lawyer. No one has found it. We can.
Valid commercial debt and healthcare receivables sit on ledgers because the cost to liquidate them exceeds their worth. We acquire these distressed claims at 5–15¢ on the dollar; our engine drives the marginal cost of prosecution to near zero, capturing a structural spread traditional firms cannot reach. The capital is trapped by human friction.
These are not small claims priced out of collection. They are eight- and nine-figure frauds sitting in the public record, undetected wrongdoing ripe for private collection under the False Claims Act, where a person may bring a civil action. Forge reads dozens of public registries at machine scale, resolves every entity to its ultimate parent, and surfaces the documentary contradiction that becomes a filed claim, originated at zero acquisition cost.
This is a high-yield, uncorrelated asset class.
This is a software company that finds the claims everyone else leaves trapped and uncovers the massive qui-tam claims no one else could see.
We originate and prosecute claims at software cost.
I. Originated claims;
II. Verified debt;
III. Efficient prosecution;
IV. Low-cost execution.
Forge is a deterministic verification engine built in Rust. It ingests dozens of public registries at once: vessel manifests, antidumping and countervailing-duty orders, CBP evasion determinations, FCC, UL and FDA certifications, federal contract and procurement records, GSA listings, SEC filings. It resolves every entity to its ultimate parent, evaluates it against forty regulatory regimes across sixteen jurisdictions, and harvests every government-facing assertion each entity has made.
Where an entity asserts one thing while a second document establishes another, Forge emits a contradiction carrying both documents, the ownership chain, the cited statute and its verbatim text, signed with Ed25519. Every input is stamped and scored for public-disclosure exposure and first-to-file position before anything is filed. While FastBurn underwrites claims sourced from aggregators, Forge originates claims that were never for sale, at zero acquisition cost. Reachable as a platform or by API, and runnable on a schedule or at boot, so verification runs continuously rather than once.
Executes specialized legal prosecution to maximize win rate. 95% cost compression. Local counsel reviews and signs every filing.
On the Stanford 100-query legal-AI benchmark, Veracity-Engine returned 100 correct answers and zero hallucinations (a perfect score [96/100 pre-PAD, 100/100 post-PAD]). By comparison: Lexis+ AI hallucinates 17%, Westlaw Precision AI 33%, GPT-4 43%, and Thomson Reuters' Ask Practical Law AI 17%. Practical Law's number is the denominator trick — it refuses 62 of 100 legal questions and hallucinates roughly one in six of the rest.
What that looks like in practice: Westlaw fabricated a jurisdictional rule for FRBP 4007 that Kontrick v. Ryan, 540 U.S. 443 (2004), directly rebuts. Lexis+ AI told users that Casey's undue-burden test still governs abortion regulation — three years after Dobbs overturned it. Thomson Reuters fabricated a Justice Ginsburg dissent in Obergefell with the fabrication about copyright (wrong topic, wrong opinion, wrong author). vLex's citator marked a unanimous Marshall SCOTUS decision as having received "negative treatment" from a federal district court — structurally impossible.
Inference Relay is our proprietary compute routing infrastructure (patent pending, commercially licensed at inference-relay.com). It reduces token costs by approximately 50× by routing heavy execution through native CLI and local processing rather than metered API billing. This is a structural moat that no competitor can replicate without building equivalent infrastructure from scratch.
AI-powered litigation at scale would be economically prohibitive using standard API pricing. High-context legal document processing (OCR extraction, multi-step drafting, citation verification) costs hundreds of dollars per day in testing alone and would scale to tens of thousands monthly in production. These costs would destroy the fund's unit economics before a single claim is filed.
$6.8B in FY2025 False Claims Act recoveries, a record. 1,297 qui tam filings, also a record, with 401 new government investigations. The DOJ and DHS Trade Fraud Task Force passed $1 billion in under a year. Perfectus Aluminum settled for $549.5 million, the largest civil customs settlement in the statute's history. Ceratizit settled for $54.4 million, with $9.75 million to a single relator. More targets than capacity, and the Department has said so.
In April 2026 the Civil Division announced FOCUS, formally inviting data miners, already more than 45% of qui tam complaints since FY2024, to bring their signals. In mid-July it published the entry mechanism, a white-paper questionnaire, and made its Global Trade and Commerce Enforcement Section permanent. A federal agency published, in advance, the specification a screening engine must satisfy, then opened a door.
Section 3730(b)(1) is the entire standing test: a person may bring a civil action. Relators have been a competitor (Island Industries v. Sigma, an $8 million verdict trebled to almost $26 million), a trade association (Perfectus, a 17.5% whistleblower share potentially exceeding $96 million), and a purpose-formed investigative firm with no market position (Customs Fraud Investigations v. Victaulic, 839 F.3d 242). The question is not who you are. It is whether your analysis survives the public-disclosure bar.
Statistical anomaly and documentary contradiction are not the same asset. Courts dismiss the first and sustain the second. Forge does not produce anomalies. It produces contradictions: an entity asserts one thing in a government-facing document while a second document establishes another, both attached, the resolution path recorded, the statute cited. A complaint built on two documents and a traced ownership chain is a different pleading from a complaint built on an outlier.
Trebled duty damages are arithmetic, not argument, because origin and classification mechanically determine the rate. Per-entry-summary penalties run $14,308 to $28,619 per violation and compound past the damages on high-volume filers. Reckless disregard suffices; there is no specific-intent element. Materiality avoids the Escobar fight. Relator share is 15 to 25% where the government intervenes, 25 to 30% where it declines, with fees paid by the defendant.
Every input Forge ingests is stamped and classified by public-disclosure channel, so a finding carries a computed exposure score before it is ever filed. That converts the central risk of data-miner qui tam from a litigation gamble into a pre-filing calculation, and builds the original-source record contemporaneously rather than reconstructing it under attack.
United States v. Heppner drew the line.
On February 17, 2026, Judge Rakoff (S.D.N.Y.) held in Heppner that communications with an AI platform lose attorney-client privilege the moment the platform's terms permit the operator to access user data. Capability of access, not actual disclosure, defeats confidentiality.
Kovel preserves one path: AI as counsel's privileged agent, but only when the operator structurally cannot access client communications. The test a court will apply is the platform's own terms.
The same week, Judge Patti (E.D. Mich.) reached a different result in Warner v. Gilbarco — but only on the work-product doctrine, which waives only on disclosure to an adversary. Heppner's privilege analysis stands. Prudent firms plan for the more demanding standard.
“Knowledge Source and Web Browsing subprocessors are not HIPAA compliant.”— §4.4
“Customer data may not process in Your selected data processing region.”— §4.4 / §4.5
“[User] authoriz[es] us to access and share Your Data with the third-party provider.”— §2.1
“the Syncly DMS service may be hosted in a single or multi-tenant environment in our discretion.”— §3.3
Lexis+ AI, Westlaw Precision AI, and most legal AI products (vLex, Harvey, CoCounsel, etc.) run on third-party foundation models (OpenAI's GPT family, Anthropic Claude, others). The customer contract is with the legal AI vendor, not with the foundation-model operator whose servers actually run the inference. Zero-data-retention commitments in the vendor contract do not reach the upstream provider. Under Heppner's capability analysis, the customer has no privity with the entity that holds the access capability.
“[T]he written privacy policy to which users of Claude consent provides that Anthropic … reserves the right to disclose such data to a host of 'third parties,' including 'governmental regulatory authorities.'”— United States v. Heppner, No. 25 Cr. 503 (JSR), slip op. at 6 (S.D.N.Y. Feb. 17, 2026)
Sovereign Shield is the architecture Heppner contemplates. Unlike Harvey's subprocessor chain, CoCounsel's authorized third-party sharing, or the foundation-model gap in Lexis and Westlaw, Sovereign Shield gives the platform operator no path to access client communications. The privilege guarantee is enforced by hardware, not by promise.
Sovereign Shield gives the firm a choice: inference in the firm's own browser using its own API key (BYOK), or inside hardware-isolated Trusted Execution Environments managed for the firm (Enterprise / Professional). All three tiers share one structural guarantee: the platform operator cannot access client communications.
Operator access is mathematically impossible, not contractually restricted. Subprocessors: zero. Retention: hardware-enforced zero, with cryptographic attestation. Privilege survives by construction.
Agentic Associate executes project-scale work end to end; Plan Mode: define the workflow once, approve, walk away.
Proactive Authority Detection: every citation, quote, and holding auto-verified against external databases before it reaches you. 100/100 on Stanford [96 pre-PAD, 100 post-PAD].
inference-relay routes compute through existing subscriptions, roughly 50× cheaper than metered API.
Entire discovery drops handled end to end: bulk productions, depositions, timelines, red-flag surfacing.
Comparison current as of May 2026. Vendor terms can and do update. The mechanism we use here — verbatim citation to specific clause numbers, dated — is the same mechanism a court will use in a privilege dispute.
Where the engine deploys.
Originated, not purchased. Duty and tariff evasion surfaced by Forge's contradiction engine across public trade, certification, and procurement records; prosecuted under the False Claims Act with whistleblower counsel. Treble duty damages plus per-entry penalties; a 15–30% statutory relator share with fees paid by the defendant. Long-duration, zero acquisition cost.
Every claim comes with a signed Bill of Lading and Rate Confirmation. Immaculate paperwork. Defendants are typically operating businesses that settle quickly. Fee recovery may be available under the Carmack Amendment.
Insurers use algorithms to mass-deny commercial claims for administrative reasons. Providers sell denied portfolios at steep discounts. ERISA and state insurance codes provide the cause of action. Bulk settlement is the norm.
Unpaid invoices governed by aggressive, one-sided vendor terms. Mandatory fee-shifting clauses, late penalties, and explicit choice-of-law provisions are standard.
Subcontractors with timely mechanic's liens hold debt secured by real estate. The property owner or title insurance forces settlement to clear title. Collection risk is structurally mitigated by the real estate securing the debt, provided statutory notice requirements are met.
Insurance companies hold thousands of low-value subrogation claims ($5K–$15K) they never pursue. We acquire the recovery right and deploy Veracity-Engine at a fraction of the traditional cost.
Capital is spread across hundreds or thousands of claims, not concentrated. The big cases are upside. The small cases are the fuel.
Systems engineer and former M&A lawyer at a top NYC firm. UC Berkeley School of Law JD, 99th percentile New York Bar Exam, 99th percentile MPRE, and a perfect score on the New York Law Exam (NYLE), with the highest-scoring exam essay in four of his first-year law school classes and a BS in Biology earned in two years, 98th percentile nationally and 3rd in class on the National Biology Major Field Test. Sole architect of Veracity-Engine (legal AI that verifies every citation against its source), Plaintiff Zero (the litigation platform for plaintiff and mass-tort firms), inference-relay (compute routing that cuts AI cost roughly 50×), and Forge (the supply-chain compliance and claim-origination engine). Co-founder and lead scientist of Blecher Group; co-designer of BrashZero (a caseless firearm operating system). Special Counsel for CFIUS, ITAR, and dual-use export controls, architecting the compliance framework for a cross-border defense-manufacturing supply chain between the United States and an allied Pacific Rim precision manufacturer, and serving as primary legal contact for technical-data exchanges with DARPA and international integrators.
Serial entrepreneur behind the highest multiple private exit in Pandora Jewelry’s history. Former Kirkland & Ellis M&A lawyer and Stikeman Elliott securities lawyer; JD, Osgoode Hall, 1st place in US Securities Regulation and Business Associations. Trained in engineering at Boston University and physics at York University, he has designed over twelve firearms since he was fourteen, including the X-Gun (a recoilless firearm purpose-built for unmanned aerial systems). Founder of Blecher LLC (defense R&D and weapon systems), Flying Gun (ultralight weapon systems for unmanned aircraft), Blecher Precision (billet CNC machining and prototyping), Corso Kinetic (edge compute for racing motorcycles), and Reckon (an AI fire-control optic and interceptor that breaks drone swarms, at cuas.app). His work spans Kalman/IMM sensor fusion (tracking moving targets from noisy sensors), DSP, control theory, and applied cryptography across production systems.
Introductions come first. Tell us who sent you and what brought you here. We will follow up to schedule a call.
This form initiates a conversation. It is not a subscription agreement and does not establish any investment relationship. Any offering will be made only by means of a confidential private placement memorandum and only to qualified investors.
Judge Rakoff's February 2026 S.D.N.Y. ruling on whether AI-assisted legal work is privileged. Why the platform's underlying architecture, not its privacy policy, decides whether attorney-client privilege survives.
How Heppner's structural-confidentiality test maps onto vendor architecture. The published terms of leading legal AI platforms — Harvey, CoCounsel — show subprocessor chains and data-sharing authorizations that fail the test on their own face, before evidentiary discovery starts.
Veracity-Engine on the Stanford Legal AI hallucination dataset. 100 queries, 0 hallucinations, 0 refusals. Leading competitors hallucinate 17–33%. The structural reasons retrieval-only legal AI keeps failing and why verification architecture carries the load.
For CTOs, CISOs, and security auditors. How the library is an enterprise procurement shortcut, bringing Shadow AI into compliance by routing through already-vetted organizational subscriptions. Three deployment architectures, none of which creates a new data processor.
How a mocked hypothetical in the birthright citizenship case quietly maps the constitutional basis for global prosecutorial reach. Citizenship confers allegiance; allegiance determines prosecutorial reach; defining who owes allegiance defines who can be prosecuted.
The disruption thesis was that AI would let small builders out-compete incumbents. Project Glasswing shows the opposite: architecture decides winners. Frontier capability plus centralized billing plus partner-gated distribution equals incumbents winning by default.
How Florida's 2023 constitutional carry amendment unintentionally created new criminal exposure for defendants whose prior cases ended in adjudication withheld. A doctrinal walkthrough of how the statutory framework collides with an older sentencing concept, and the compliance trap it created for the substantial population it sweeps in.
The Dual Envelope architecture, now a memory-safe Rust daemon. Developer ships intelligence with full operational visibility; user prompts and completions never leave the user's subscription boundary. Gross margin on inference-heavy products moves from ~15% to ~98.9%. Unlimited concurrent sessions for multi-agent topologies impossible at metered rates. Patent Pending.
Regulatory & Privilege Sovereignty for law, healthcare, and finance. Routing attorney work product or PHI through a third-party developer's API key can waive attorney-client privilege or collapse a HIPAA BAA; the Dual Envelope architecture keeps the developer outside the data-processor stack entirely. The user's existing BAA or DPA remains the sole governing framework. IR2 also responds to Anthropic's June 15, 2026 Agent SDK billing split: the daemon drives interactive Claude Code, which stays subsidized; agent calls never touch the metered pool.
Legacy. The original npm-library distribution's compliance argument, post-April-4-2026, for routing inference through users' existing subscriptions. Superseded for new deployments by the IR2 opinion.