Google.org Impact Challenge · AI for Government Innovation

The Triad Justice System

Three independent pillars evaluate anonymized court cases in parallel. When they agree, confidence increases. When they diverge, the system flags cases for review — surfacing bias, error, and inconsistency that no current tool can detect.

Courts are among the most consequential public services, yet rulings are influenced by legally irrelevant factors — judge assignment, calendar position, cognitive fatigue. In the US, sentencing for identical offenses varies over 50% by judge. In Japan, civil-law opacity prevents citizens from assessing fairness. No existing system independently audits judicial reasoning in real time.

I

Human Judge

The existing judge's ruling as baseline. Retains full binding legal authority, constitutional grounding, and final decision-making power. The Triad never overrides — it audits in shadow mode.

II

Blockchain Citizen Jury

7–9 randomly selected citizens vote through the Kleros decentralized justice protocol. Anonymous, tamper-proof, publicly verifiable. Operational in Argentina's Mendoza Supreme Court since 2024.

III

Adversarial AI Agents

Two generative AI agents argue opposing positions. A third synthesizer stress-tests both and produces a reasoned conclusion. Multi-agent debate format proven for legal fact-finding.

System architecture

A shadow audit running in parallel — zero disruption to courts, no changes to judicial authority.

Input
Anonymized civil & family cases · US and Japan courts
↓
Pillar I
Human judge ruling
Pillar II
Kleros blockchain jury
Pillar III
Adversarial AI agents
↓
Convergence engine
Compare outputs · detect divergence · categorize: bias / error / complexity
↓
Agreement
All three align → confidence increases
Divergence
Pillars disagree → flagged for review

Dual-jurisdiction pilot

Identical methodology across common law (US) and civil law (Japan). The first quantitative cross-jurisdiction judicial consistency comparison.

USUnited States — California

Legal systemCommon law · adversarial
Target courtLA Superior Court
RegulationCA Rule 10.430 · Learned Hand
Year 1 target200–500 cases

JPJapan — Okinawa & Tokyo

Legal systemCivil law · inquisitorial
Target courtFamily courts · Okinawa & Tokyo
Academic anchorsOIST · Univ. Tokyo · Keio
Year 2 target200–500 cases

36-month roadmap

MONTHS 1–6 · PROTOTYPE
Build adversarial AI prototype on Vertex AI
Adapt Kleros for US legal context. Develop anonymization pipeline. California court outreach. Google.org Accelerator Demo Day December 2026.
MONTHS 7–12 · US PILOT
100+ anonymized cases through three pillars
Court partnership signed. Blockchain jury operational. First interim report published open-access. Beneficiary advisory panel established.
MONTHS 13–24 · US COMPLETE · JAPAN PREP
200–500 cases · first peer-reviewed paper
Japan legal framework mapping. Kleros adapted for Japanese context. Open-source codebase published.
MONTHS 25–36 · JAPAN PILOT · COMPARATIVE
First cross-jurisdiction judicial consistency data
Japan pilot complete. Cross-jurisdiction comparative analysis published. Replication toolkit released. 3+ additional jurisdictions engaged.

Project leadership

PROJECT LEAD

Dr. Jovan Rebolledo-Mendez

Founder, Exponential Japan KK. World's first AI-for-Good contest (2017), first Future of Democracy Conference (2022). Faculty AI at Univ. Tokyo, lecturer at Keio, Staff Scientist at OIST. 25+ years AI R&D. Direct domain expertise in cross-border judicial systems.

BLOCKCHAIN JUSTICE ADVISOR

Federico Ast, PhD

Co-founder & CEO, Kleros — the world's first blockchain justice protocol with formal supreme court pilot (Mendoza, Argentina 2024). Advisory Board, Stanford Journal of Blockchain Law & Policy. Letter of Collaboration filed with Google.org application.

Open source. Open access. Open by design.

All code released under MIT / Apache 2.0. All findings in open-access journals. All AI reasoning chains and blockchain jury records publicly auditable. Complete replication toolkit enables any jurisdiction to deploy independently.

Partner with Exponential Japan →