jhana

Science &
Technology

VerificationPreliminary results

We decouple taste (penumbra) from reading (core), and certify whether the minimum evidence was considered. That takes the liability out of full automation, and gives models a signal they can hill-climb.

Motivation

Consider due diligence tasks on closed data rooms. Or audit, forensics, scrutiny, eDiscovery, etc. There is no open world or external research, but we must carefully consider very large N document sets of 1000s of pages and, at least, millions of tokens.†

Problem

Legal tasks are open texture and sans oracle: reasonable experts disagree based on the same evidence. Only 44% of USSC judgments are unanimous! So there is no answer key to grade an agent against.

Bound

A high score cannot hide skipped evidence. A 99% score can hide only required evidence within 1% of the weight, even for an agent trained to maximize the score. The bound is proved, and tightCˆw(T)≥1−ε ⟹C∗(T)≥1−μw(ε) (tight)
μw(ε): largest fraction of required units with total weight ≤ε
.

Verifier scores match human lawyers better than naively reading more or LLM querygen

Lawyer-marked evidence read 0%50%100%0%50%100% Reading volume τb = 0.368 0%50%100% Verifier score τb = 0.548

Agents use 31⁠–⁠92% fewer tokens while mean coverage improved by >21.5 necessary, missed pages

naive verifier n tokensper point .49.26.34.13.15 .68GPT 5.6 Luna23−31%.52GPT 5.4 Mini22−53%.46Gemini 3.5 Flash24−32%.42Gemini 3.1 Flash Lite24−41%.35Gemini 3.1 Pro24−92%

Convergence to coverage 1.0 if used as a remediation skill

Mean 35% → 98%.

024680.51 remediation/repair round stop at τ = .95 35% 98%

Not yet saturated by top models, well-differentiated signal

This verifier can be hill-climbed to finetune agents and harnesses for the law.

.30.35.40.45.50 naive control .345 verifier score GPT 5.6 Sol (high)GPT 5.6 SolGPT 5.6 TerraGPT 5.6 LunaOpus 5Sonnet 5Haiku 4.5DeepSeek V4 ProDeepSeek V4 FlashKimi K3Grok 4.6 .498.443.432.405.463.410.368.405.381.445.415

†40 contract tasks & 11,356-page eval: AUC .963 vs .900 BM25.

Real work data for evaluations

Real tasks from 16K+ lawyers and 650+ judges, and 300K+ disputes followed from filing through disposal. Build and evaluate agents against what actually happened in court, so they avoid liability, win more, and cost less.

2,440,503 requests by location

Through September 2026
15,121accounts
63,483agent tasks
6,999,784document pages processed
2,506towns and cities with active users
9,527minutes of audio recorded
774,566documents uploaded

Model‑ & harness‑agnostic e2e orchestration

Powerful AI is flooding us with claims and disputes.

  • Claim my lost-bag compensation.

    Grok Bot
  • Dispute my credit report.

    Dots
  • Add me to the mass arbitration.

    Muse
  • Push back on my roof claim.

    Instinct
  • Get a small-claims notice issued.

    Muse
  • File a complaint with the regulator.

    AI personal injury firm
jhana

A bank’s FCRA disputes

Nov 3Dispute forecastCorrected

Sent to credit bureaus
640 accounts at Status 71, 30–59 days past due
Payment ledger
Paid on time, held in suspense after loan transfer
Correction
AUD to all 3 bureaus, same day
Forecast if uncorrected
70 disputes, 4 suits
Received
6 disputes, 0 suits

This weekClocks0 late

  1. Bureau disputes
    6 h of 30 d
  2. Direct disputes
    Day 1 of 30
  3. CFPB complaints
    Day 2 of 15
  4. Arbitration
    Answer filed, day 1
  5. Federal suits
    Day 3 of 21

FCRA § 1681s‑2(b)Demand letterRouted

“$4,500, or we sue.”

Records read
63 of 63 necessary
Finding
Reported accurately

Expected cost, by route

  1. Pay as asked$4,500
  2. Settle$1,200
  3. Answer with the record$310

1:26‑cv‑04127Federal suitClosed

Claim
Investigation was not reasonable
Answer
Filed day 3 of 21
Investigation file
37 of 37 necessary records read
Motion
Summary judgment
Outcome
Judgment for the bank

NovemberHuman approval, by authorityApproved

  1. ApproverClaimsLiability
  2. Panel firmAbove $50,000, or class9Panel firm
  3. In-house counsel$5,000 to $50,00096Bank
  4. jhana flex staffingMinor claims, under $5,0001,412jhana

2023–25Eval on real work dataScored

Work data
4,812 closed claims, filing through disposal
Verifier
Coverage of required evidence, 0.98
Lost on skipped evidence
131 claims, 127 caught on replay
Swarmmany agents work every step at once
End to endeach claim runs one line from arrival to close, like a factory floor
Sovereignthe whole stack can run on a local model in your cloud or on your premises, with no data leaving

Runs on

  • Any model

    A lab API, or a local model

    OpenAIAnthropicGoogle GeminixAIDeepSeekKimi
  • Open weights

    Hill-climbed against the verifier to the Pareto frontier

  • Any harness

    Ours, or yours

  1. Capture & Connectors

    Call logs, ticket history, notices, data enhancement, and the CFPB clock

  2. Data Forensics

    Receipt vs. ledger, missing recordings, conflicting dates, discovery, productions, and deemed admissions

  3. Routing

    Arbitration clause, motion to compel, Rule 68, settlement authority, and trial

  1. Filings/Comms

    Answer and removal clocks, answers, early motions, e‑filing, service, and docketing

  2. Hearings & IRL Events

    AI transcripts and next dates

  3. Disposal & Follow-ups

    Release, dismissal with prejudice, tradeline correction, enforcement, closure, and root cause

Pattern informs prediction A shared payment-posting issue or cancellation flow

Start with a conversation.

Talk to our team

AI audit

AI safety &
human autonomy

Explainability, audit trail, and human liability.

Tokenomics
& ROI

(Re)Staffing, time & token attribution, and business development.

Sovereignty &
client privilege

Vendor exposure, data privacy, and client disclosures.

Forward-deployed technology

L0

Suit

Formalization & verification. Context & completeness.

L1

Para Agent

Model-agnostic or local harness with firm guardrails.

L2

Spatial

Role-based IT & cloud substrate for humans & AIs.

L3

Temporal

Assembly-line automation: triggers, schedules, and goals.