typesafe jev legal intakeWhy TypeSafe Jev and System One Models Beat LLMs for High-Volume Legal Intake

When the first commercial automobiles rolled onto public streets at the turn of the twentieth century, vehicle builders did not design modern transit systems from first principles. They mounted single-cylinder gasoline engines onto traditional wooden carriages, preserving high buggy perches, leaf springs, and decorative whip sockets.

For the past four years, legal technology has been trapped in the exact same horseless carriage pattern.

When foundation models proved capable of generating human-grade prose, the legal tech sector assumed that artificial intelligence in law practice had to take the form of a conversational chatbot. Legal software vendors took large language models optimized for human conversational preferences via Reinforcement Learning from Human Feedback (RLHF) and bolted them directly onto legal document extraction, case review, and client intake workflows.

The operational results have been disastrous for high-volume litigation firms:

• Intake paralegals wait 20 to 45 seconds for a model to generate text tokens sequentially.
• Backend ingestion workers crash because an LLM emitted invalid JSON or escaped an apostrophe inside an indemnity clause.
• Models hallucinate statutory deadlines or dispute forums with absolute linguistic confidence.
• Inference bills for a single 5,000-claimant mass arbitration campaign surge past $15,000.

Today, TypeSafe AI founder Diogo Almeida (co-author of the foundational OpenAI research behind ChatGPT, with over 63,000 scholarly citations) published Introducing System One Models & Jev. When one of the principal researchers who invented modern chat models wears an "Anti Language Language Model" shirt, legal engineers must pay attention.

The thesis behind System One models directly confirms the architectural doctrine we established while engineering legal intake systems at OBE: litigation software does not need an articulate chatbot. Litigation software needs fast, typed, machine-native decisions.

Here is what TypeSafe's Jev accomplishes, the profound legal impact it creates for mass arbitration and mass tort intake, and the operational capabilities it unlocks across OBE's high-volume legal factory.

---

What System One Models & Jev Actually Are

To understand why Jev represents a fundamental leap for litigation technologists, you have to look at what it intentionally discards: string generation.

Conventional frontier LLMs are autoregressive token predictors. To determine whether a disputed solar agreement contains an enforceable arbitration clause, an LLM must predict one character token at a time: {", a, r, b, i, t, r, a, t, i, o, n, ":, true. Every single token requires a full forward pass through hundreds of billions of parameters.

Jev breaks this paradigm entirely. Instead of training models to generate polite conversation, TypeSafe built an architecture trained with Reinforcement Learning for Calibrated Decisions (RLCD). Unstructured legal records go in. typed, probabilistic decisions come out in parallel.

Architectural Dimension	Conventional Frontier LLMs	TypeSafe System One (Jev)
Optimization Target	Human preference & conversational fluency (RLHF / RLVR)	Calibrated probabilistic decisions on software tasks (RLCD)
Primary Input	Conversational chat strings	Unstructured text and structured program state
Primary Output	Freeform text strings requiring regex/JSON parsing	Mathematically guaranteed type-safe values (booleans, enums, numbers)
Sampling Mechanism	Sequential token-by-token generation	Parallel multi-decision sampling in a single pass
End-to-End Latency	3,000ms to 45,000ms+ per document call	70ms to 500ms (40x to 200x speedup)
Inference Cost	$0.20 to $10.00 / MTok input. up to $30 / MTok output	$0.042 / MTok input ($42 per billion tokens). outputs free
Type Safety	Probabilistic schema matching. frequent JSON syntax errors	Mathematically zero schema or type errors
Uncertainty Calibration	Uncalibrated overconfidence. high risk of hallucination	Honest epistemic probability scores returned on every field

By giving up the ability to write essays, Jev gains sub-second execution, near-zero cost, mathematical type safety, and honest confidence calibration.

For OBE, those characteristics solve the most severe architectural bottlenecks in mass arbitration and complex statutory claims.

---

Cut Inference Latency and Eliminate Parse Errors

Autoregressive LLMs take up to 45 seconds per document and frequently output invalid JSON. System One models return typed evaluations in under 500 milliseconds with zero schema errors.

Schedule a Technical Architecture Audit →

---

1. AI in the Inner Loop: Why Legal SaaS Still Feels Like 2019

During the technical launch exchange on X, AI researcher Rohan Paul noted the critical threshold:

"150ms can truly change the equation. So now it can put model decisions inside software paths that were previously too latency-sensitive for an LLM."

Diogo Almeida responded directly:

"Exactly! AI in the inner loop should allow software to be more powerful. I often ask 'why is SaaS basically the same as it was in 2019?'"

In legal technology, that question exposes the core failure of legacy case management systems.

Traditional legal software platforms like Filevine, Clio, and Lead Docket treat artificial intelligence as an outer-loop novelty. It exists as an isolated sidebar chat widget or an asynchronous background job where paralegals upload a PDF, navigate away, and wait several minutes for an extraction queue to finish. The core experience of managing litigation files remains completely unchanged from 2019.

At OBE, sub-150ms latency allows us to pull intelligence directly into the inner loop of legal operations:

• Inside Form Keystrokes: When a homeowner uploads a solar installation contract or loan disclosure on Step 1 of an intake questionnaire, Jev evaluates 20 statutory criteria in parallel before the user finishes typing their address on Step 2.
• Inside Database Mutations: Rather than queuing files for offline processing, Convex mutations evaluate forum jurisdiction, fee-shifting applicability, and statutory multipliers before the record commits to disk.
• Inside the Active Review Workstation: Intake staff toggling between matters experience zero loading spinners. Entity highlights and coordinate-mapped proofs render instantly.

Moving artificial intelligence from an outer-loop batch job into the inner loop of user interaction eliminates lead abandonment and paralegal wait times.

---

2. Smart If-Statements: The Return of Neuro-Symbolic AI in Law

On X, technologist Paul Mit captured the fundamental transition away from conversational bloat:

"Hot take: most software doesn't need another LLM generating text. Now tech world need to teach them to make decisions: Is this urgent? Is this spam? Does this need a human? Should we send this notification? That's the interesting idea behind TypeSafe basically: smart if statements."

Almeida affirmed this view:

"Absolutely yes to 'smart if statements', or as the old school would put it 'neuro-symbolic AI'!"

Legal intake is the ultimate neuro-symbolic domain.

Statutory litigation is strictly symbolic. Deadlines under statutes of limitations are exact calendar dates. Poverty fee-waiver thresholds under American Arbitration Association (AAA) or JAMS rules are specific percentage multiples of the federal poverty line. Statutory damages under the Truth in Lending Act (15 U.S.C. Section 1635) or state elder abuse statutes follow rigid mathematical formulas.

Intake evidence, however, is fuzzy neural text. It arrives as scanned installation agreements, skewed PDF addenda, ambiguous dispute clauses, and contradictory sales representative notes.

Pure symbolic software (rigid forms, regular expressions) shatters when confronted with contract variation. Pure neural language models hallucinate, drop formatting, and fail to provide auditable guarantees required by state bar ethics rules.

System One models unite these paradigms. Jev handles neural pattern matching across messy contract language, and outputs discrete, typed values that ordinary code can branch on:

// Neuro-symbolic legal triage inside OBE
const decision = await jev.evaluate(contractDocument, {
  arbitrationForum: ["JAMS", "AAA", "CPR", "NONE"],
  hasClassActionWaiver: "boolean",
  hasIndividualFeeShifting: "boolean",
  povertyFeeWaiverEligible: "boolean",
  elderFraudMultiplierApplies: "boolean",
}).

if (decision.arbitrationForum === "JAMS" && decision.hasIndividualFeeShifting) {
  await stageArbitrationDemand(claimantId, "JAMS_CONSUMER_EXPEDITED").
} else if (decision.povertyFeeWaiverEligible) {
  await attachAffidavitOfIndigence(claimantId).
}

This is not prompt engineering. It is deterministic, auditable software engineering powered by neural semantic understanding.

---

3. Calibration Was Table Stakes for ML: Ethical Compliance Under ABA Opinion 512

In the launch dialogue, developer Santiago raised the critical question regarding decision reliability:

"I really like that the model returns a confidence score. That alone can reduce a ton of hallucinations because we can decide whether to present a response to a user or pass it to a HIL [Human-in-the-Loop]. What I'm curious to understand is whether those confidence scores are actually accurate..."

Almeida reminded the community of how production machine learning operated before the generative hype cycle:

"Fun fact: back in the pre-GenAI days, calibration was just table stakes for ML (how else could it be used in systems???)"

In legal practice, calibration is not an academic nicety. It is an ethical requirement.

Under the American Bar Association's landmark ethics guidance, ABA Formal Opinion 512 on Generative AI, attorneys bear strict professional responsibility for technological competence (Rule 1.1), managerial supervision (Rule 5.1), and supervisory responsibility over nonlawyer assistance (Rule 5.3). Lawyers cannot satisfy their duty of candor to the tribunal (Rule 3.3) by delegating case evaluations to uncalibrated models that claim absolute certainty while hallucinating facts.

In high-volume mass arbitration, filing hundreds of demands containing unverified claims or hallucinated contract terms exposes a law firm to Federal Rule of Civil Procedure 11 sanctions, fee-shifting penalties, and summary dismissal by arbitration providers.

Because Jev is trained using Reinforcement Learning for Calibrated Decisions, its probabilities reflect measured epistemic accuracy. When Jev returns a 99% confidence score, it is correct 99% of the time. When it encounters genuine contract ambiguity, its confidence drops accordingly.

Our foundational operating rule at OBE is: AI will process the data, lawyers will decide what matters.

Calibrated confidence scores allow us to build a mathematically governed Human-in-the-Loop workflow:

{
  "arbitrationForum": { "value": "JAMS", "confidence": 0.994 },
  "hasFeeShiftingClause": { "value": true, "confidence": 0.982 },
  "statutoryRightOfRescissionDisclosed": { "value": false, "confidence": 0.961 },
  "forgerySuspected": { "value": true, "confidence": 0.618 }
}

• Autonomous Routing (>95% confidence): Undisputed arbitration clauses, clean disclosures, and clear party records advance directly into demand staging without manual intervention.
• Surgical Exception Triage (60% to 85% confidence): When an electronic signature timestamp shows potential tampering and confidence drops to 61.8%, the system flags that single field in the intake specialist's priority work inbox.

In our 50/50 split-screen workstation, the PDF canvas automatically pans and zooms directly to the suspicious signature block with coordinate-level SVG bounding boxes. The intake specialist resolves the uncertainty in ten seconds. Review shifts from an exhausting manual audit of every field into focused exception handling.

---

Benchmark Your Document Qualification Pipeline

Evaluating thousands of contracts in mass arbitration requires sub-second execution and honest calibration scores. See how our neural triage pipeline handles disputed agreements.

Book a Technical Consultation with Tim Ottowitz →

---

4. Zero-Repair Ingestion Pipelines in Convex

In traditional legal tech stacks, defensive engineering consumes up to 40% of custom backend code.

When engineers build on chat-based LLM APIs, every document extraction call requires an entire defense apparatus:

• Retrying requests when models output conversational greetings or markdown formatting fences.
• Installing runtime patch tools like json-repair to fix unescaped double quotes inside contract indemnity text.
• Handling Zod schema validation failures when an LLM returns a string description instead of an enum variant.
• Managing dead-letter queues to keep corrupted payloads from breaking reactive Convex database mutations.

With Jev, schema matching is mathematically guaranteed. When OBE requests a typed dispute evaluation:

type DisputeQualificationRecord = {
  arbitrationForum: "JAMS" | "AAA" | "CPR" | "NONE".
  hasMandatoryClassWaiver: boolean.
  hasFeeShiftingClause: boolean.
  statutoryRightOfRescissionDisclosed: boolean.
  elderFraudStatuteApplicable: boolean.
  povertyFeeWaiverEligible: boolean.
  disputedContractDate: string.
}.

Jev samples directly into the discrete output space. The model cannot hallucinate unexpected keys, cannot emit malformed JSON, and cannot produce invalid types.

Ingestion workers write directly into Convex database tables with zero translation layers, zero regex repair routines, and zero queue poison risks.

---

5. The Mass Arbitration Economics Shift: 400x Cost Drops

Consumer mass arbitration campaigns live and die on unit economics and filing pressure.

Under AAA Consumer Arbitration Rules and JAMS Consumer Arbitration Minimum Standards, corporate respondents must pay the majority of arbitration administration and arbitrator fees ($1,500 to $3,500+ per matter) just to initiate proceedings. When a law firm represents 2,500 verified claimants, the corporate respondent faces $3,750,000 to $8,750,000 in immediate, non-refundable filing fees before a single hearing occurs.

To establish that advantage, the plaintiff firm must verify four legal criteria across every single contract:

1. Does the agreement contain an enforceable arbitration clause?
2. Does the agreement include a fee-shifting clause obligating the respondent to pay claimant attorney fees upon prevailance?
3. Does the claimant qualify for a 100% filing fee waiver under Department of Health and Human Services (HHS) poverty guidelines?
4. Do statutory elder abuse multipliers apply (for example, California Civil Code Section 3345 or state deceptive trade practice statutes) to multiply damages?

Evaluating thousands of multi-page agreements using frontier LLMs creates a prohibitive cost bottleneck:

• 5,000 claimants × 35 pages = 175,000 pages of text (~120 million tokens).
• At standard frontier pricing ($3.00 to $10.00 per million input tokens plus costly output tokens), qualifying the portfolio costs $1,500 to $5,000+ in raw API spend. With multi-pass extraction and verification, costs exceed $15,000.

At Jev's pricing of $0.042 per million input tokens ($42 per billion tokens) with free output tokens, qualifying the exact same 5,000-claimant portfolio costs under $15.

Portfolio Caseload	Pages Evaluated	Frontier LLM Stack Cost	TypeSafe Jev Ingestion Cost	Direct Savings
500 Claimants	17,500 pages	$450.00	$1.25	99.7%
2,500 Claimants	87,500 pages	$2,250.00	$6.25	99.7%
5,000 Claimants	175,000 pages	$4,500.00	$12.50	99.7%
15,000 Claimants	525,000 pages	$13,500.00	$37.50	99.7%

This cost collapse restructures plaintiff firm economics. When qualifying a claim file costs cents rather than dollars, boutique law firms can represent consumer classes with lower individual recoveries that were previously cost-prohibitive to litigate.

---

6. Composable Legal Decision DAGs (The Death of the Prompt Blob)

Early legal AI implementations relied on massive prompt blobs: five pages of text begging an LLM to read a contract, extract thirty entities, evaluate four statutory claims, and format everything as JSON.

These prompt blobs are impossible to maintain. Updating a prompt to improve interest rate extraction breaks arbitration clause detection. Adding instructions for signature verification causes the model to hallucinate cancellation dates.

TypeSafe's System One architecture enables OBE to replace prompt blobs with discrete, typed decision nodes arranged in a directed acyclic graph (DAG):

┌─────────────────────────────────────────────────────────────┐
│                 Raw PDF Intake Ingestion                    │
└──────────────────────────────┬──────────────────────────────┘
                               ▼
┌─────────────────────────────────────────────────────────────┐
│          Node 1: Document Classification (Jev)              │
│       Detects: Loan Agreement, Permit, Disclosure           │
└──────────────────────────────┬──────────────────────────────┘
                               ▼
┌─────────────────────────────────────────────────────────────┐
│          Node 2: Arbitration Forum & Rules (Jev)            │
│       Evaluates: JAMS vs AAA, Class Waiver, Fee-Shift       │
└──────────────────────────────┬──────────────────────────────┘
                               ▼
┌─────────────────────────────────────────────────────────────┐
│          Node 3: Statutory Violation Engine (Jev)           │
│       Screens: TILA APR Caps, Right of Rescission, Elder    │
└──────────────────────────────┬──────────────────────────────┘
                               ▼
┌─────────────────────────────────────────────────────────────┐
│          Node 4: Sovereign Convex Case Record               │
│       Clean, typed data locked with bounding box proofs     │
└─────────────────────────────────────────────────────────────┘

Because each node runs in 80ms and output tokens are free, every legal decision is isolated, testable, and version-controlled. If a firm adjusts intake criteria for an upcoming campaign, our engineers modify a single discrete node without risking regressions across the rest of the pipeline.

---

The Machine-Native Legal Factory

The legal industry does not need artificial general intelligence to transform litigation operations. It needs reliable, machine-native intelligence infrastructure that connects cleanly to code.

For four years, legal tech forced conversational language models into transactional data pipelines. By discarding string generation and focusing entirely on parallel, calibrated decisions, TypeSafe AI's Jev demonstrates the real future of legal technology: sub-second execution, zero syntax failures, calibrated human oversight, and near-zero unit costs.

At OBE, we build software that allows high-volume law firms to operate at machine speed while keeping licensed attorneys in full control of their case records. System One models provide the exact decision engine that vision demands.

To see how OBE automates case qualification, evidence collection, and split-screen review for high-volume litigation, schedule a private architecture walkthrough with our team.

---

Find the Right Intake Software and Developers for Your Firm

Ready to slash document processing costs while maintaining strict type safety?

👉 Schedule a Consultation with Tim Ottowitz to Review Your Case Type
We will inspect your current extraction stack, benchmark inference costs across your dockets, and demonstrate how calibrated decision models eliminate parsing errors.

Book a demo →