Technology
& practice.

Suit · Para Agent
Spatial · Temporal

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

L0Suit	Formalization & verification. Context & completeness.
L1Para Agent	Model-agnostic or local harness with firm guardrails.
L2Spatial	Role-based IT & cloud substrate for humans & AIs.
L3Temporal	Assembly-line automation—triggers, schedules & goals.

Architecture

L0 Suit

A structured record, linked to its sources.

Formalization and verification

L1 Para Agent

Research, drafting, and review

Models, tools, and attributed cost

People

Judgment, relationships, and sign-off

L2 Spatial

Roles and permissions for people and agents

L3 Temporal

Triggers, schedules, and goals coordinate tasks and people over time.

L0 Suit formalizes and verifies the source-linked record. L1 Para Agent works over that record; people retain judgment and sign-off. L2 Spatial controls access for both people and agents. L3 Temporal is the outer orchestration layer, coordinating tasks and people through triggers, schedules, and goals.

Token costs
& evaluations.

Karnataka crosses 300K case files processed by AI.

Court filing token costs
down to $0.44 per case.

Suit makes every model
cheaper/more complete.

We run 5M+ pages/week at the highest stakes.

This is powered by our Lab, researching across formalization, verification, and tokenomics for the law. Rupee economics pushed us to make inference leaner. Suit is our verifier which certifies coverage in subjective audit/diligence tasks. It improves the performance of every model while reducing costs. And per-case costs have fallen from ~$110 in the o1-era to between $0.44 and $3.8 today.

Verifiers &
coverage certificates.

Completeness of consultation.

Over a fixed document set, $\hat{q}$ estimates the minimum required reading for a query. T is what was read.

\hat{C} (T) = \frac{| \hat{q} \cap T |}{| \hat{q} |}

The certificate says where, not only how far. It also returns the uncovered set: $\hat{q} ∖ T$ .

A tight bound on certificate error.

$C_{q}$ measures completeness against the true, unobserved minimum reading set. For every reading T:

| \hat{C} (T) - C_{q} (T) | \leq \max (μ, σ)

$μ$: Miss rate. The share of required units the estimate leaves out.
$σ$: Spurious rate. The share of the estimate’s units that are not required.

Read the whole reference.

T \supseteq \hat{q} \Rightarrow C_{q} (T) \geq 1 - μ

The uncovered set is the repair procedure.

Choose the k highest-weight uncovered units.
Prompt the policy with those units and their claims.
Add the new reading. Repeat until the certificate reaches the target.

What the reference missed, its feedback cannot repair.

Read the formalization — definitions, bounds, and repair

Completeness

Every check resolved before output

Verification

Legal rules as executable verifiers

Audit trail

Every decision inspectable, on file

38 NLS-trained evaluators prefer jhana over GPT and human gold-standard.

Elo rating: 1173#1 overall
vs GPT-4o: 73.1%win probability
vs Human SOTA: 58.2%vs SCC editors

Capability	jhana	Harvey	Services	Labor KPO
Real-world retrieval + filings	✓	✗	?	✓
Lawyer verification	✓	✗	✗	✓
Automation + custom workflows	✓	?	?	✗
Formalize + verify, not RAG	✓	✗	✗	✗
BYOK / 100% local edge	✓	✗	✗	—

Our lawyers in the loop are alumni of