JUDGE

Every Output Verified.
Every Decision Defended.

The dedicated verification agent that governs every AI outputinside Mentis OS — eliminating hallucinations, enforcing policy,and producing audit-ready intelligence.

96.8%
Avg Confidence
<200ms
Verification
100%
Traceable
Zero
Blind Trust

The Verification Layer

What Is
JUDGE?

JUDGE is a dedicated, independent verification agent that sits between every AI output and the end user. It does not generate intelligence — it governs it.

Where JUDGE Sits in the Mentis OS Stack
DATA SOURCESAGENT LAYERJUDGEOUTPUTDocuments & SOPs14,200 filesDatabases & APIsSQL, CRM, ERPIoT & TelemetryReal-time streamsExternal SignalsMarket, regulatoryAegisResearch AgentSophiaKnowledge AgentAnalyticaData AgentCustom AgentUser-definedSource VerificationHallucination CheckPolicy ComplianceConfidence ScoreVERDICTVERIFIEDDelivered to userREJECTEDSent for re-reasoningESCALATEDHuman reviewAudit TrailImmutable logMENTIS OS — GOVERNED EXECUTION PIPELINE

Independent Agent

JUDGE is not the same model that generated the output. It is a separate, dedicated verification agent with its own reasoning — eliminating self-validation bias.

Multi-Dimensional Checks

Every output is evaluated across source fidelity, factual consistency, policy compliance, and confidence scoring — not just a single pass/fail gate.

Audit-Ready by Default

Every verification produces an immutable trace — what was checked, what was found, what verdict was reached. Always reconstructible.

Verification Pipeline

Five Gates.
Zero Blind Trust.

Every AI output passes through five independent verification stages before reaching the user. Each gate can reject, escalate, or approve.

JUDGE Verification Pipeline — Real-Time
01
Source Cross-Reference32ms
02
Hallucination Detection58ms
03
Policy & Compliance Gate41ms
04
Confidence Scoring29ms
Verdict & Trace Generation18ms
PASSREJECTESCALATE
Total verification: 178ms5 gates · 0 bypassed
01

Source Cross-Reference

Every claim in the output is traced back to the source documents or data that informed it. If a claim cannot be linked to an approved source, it is flagged. No unsupported assertions pass through.

02

Hallucination Detection

JUDGE compares the generated output against retrieved evidence using semantic consistency checks. Fabricated facts, conflated entities, invented statistics, and logical inconsistencies are caught and rejected.

03

Policy & Compliance Gate

Outputs are validated against enterprise-defined rules — data classification, disclosure restrictions, role-based access policies, and regulatory constraints. JUDGE enforces what the enterprise allows, not what the model can produce.

04

Confidence Scoring

A composite confidence score is computed across source strength, reasoning coherence, and verification pass rates. Low-confidence outputs are either rejected or escalated to human review — never silently delivered.

Verdict & Trace

JUDGE issues a final verdict — PASS, REJECT, or ESCALATE. Every verdict generates an immutable audit record containing the full chain of evidence, checks performed, and reasoning. Reconstructible months or years later.

Live Verification

See JUDGE
Catch a Hallucination

RAW AGENT OUTPUT — PRE-JUDGEUNVERIFIED
QUERY

"What is our data retention policy for EU customers?"

AGENT RESPONSE

EU customer personal data is retained for 48 months post-contract termination with automated deletion workflows. The policy was updated in January 2025 following board approval.

Retention period: 48 months — Source says 36 months
Update date: January 2025 — Source says March 2024
Automated deletion workflows — confirmed
JUDGE VERDICT: REJECTED

2 factual claims contradict source documents. Output sent back for re-reasoning with corrected evidence.

VERIFIED OUTPUT — POST-JUDGE✓ VERIFIED
QUERY

"What is our data retention policy for EU customers?"

JUDGE-VERIFIED RESPONSE

EU customer personal data is retained for 36 months post-contract termination with automated deletion workflows. The policy was updated in March 2024 following board approval.

Retention period verified Legal/GDPR/retention.pdf §4.2
Update date verified Board-Minutes-Q1-2024.pdf
Automated deletion verified IT-Ops/deletion-workflow.yaml
✓ JUDGE VERIFIED3 SOURCES98.2% CONFIDENCE
AUDIT-ID: JDG-EU-RET-4417 · 178ms verification · 2nd pass
Failure Categories

What JUDGE
Catches & Prevents

Factual Hallucinations

Fabricated statistics, invented dates, conflated entities, or claims that do not exist in any source document.

"Revenue grew 34% in Q3" — No source supports this figure

Source Misattribution

Correct facts attributed to wrong documents, or citations that reference non-existent sections.

"Per Board-Minutes §7.3" — §7.3 does not exist in referenced file

Logical Inconsistencies

Contradictory claims within the same response, or conclusions that do not follow from presented evidence.

Claims "risk is low" while citing critical severity indicators

Policy Violations

Outputs that disclose restricted data, exceed role permissions, or violate enterprise-defined content policies.

Returned PII data to user without clearance for that domain

Stale Information

Answers based on superseded documents or outdated data versions when newer authoritative versions exist.

Referenced v2.1 policy — v3.0 supersedes since Feb 2025

Low-Confidence Speculation

Assertions delivered with high confidence when source evidence is weak, ambiguous, or insufficient to support the claim.

Stated "definitive" conclusion from 1 partial source match
Immutable Record

Every Verdict.
Permanently Recorded.

The audit trail is not a log file. It is a cryptographically verifiable chain of evidence — reconstructible by auditors, regulators, and forensic teams months or years later.

JUDGE AUDIT RECORD — JDG-EU-RET-4417
IMMUTABLE
TIMESTAMP
2026-02-15T14:23:47Z
USER
sarah.mitchell@corp
AGENT
Sophia (Knowledge)
VERDICT
PASS (2nd attempt)
Verification Chain
Gate 1 — Source cross-referencePASS32ms
Gate 2 — Hallucination detectionPASS58ms
Gate 3 — Policy compliancePASS41ms
Gate 4 — Confidence scoring98.2%29ms
Gate 5 — Final verdictVERIFIED18ms
Sources Verified
Legal/GDPR/retention.pdf §4.2Board-Minutes-Q1-2024.pdfIT-Ops/deletion-workflow.yaml
Hash: 7f3a9c...e4d2b1CipherVault encrypted · PQ-256
The Difference

AI Without JUDGE
Is a Liability

Dimension
Without JUDGE
With JUDGE
Hallucination Control
Hope for the best
Detected & rejected
Source Attribution
Optional, unverified
Mandatory, cross-checked
Policy Enforcement
Model discretion
Enterprise rules enforced
Confidence Awareness
Always confident
Scored & thresholded
Audit Trail
Partial logs
Immutable, complete
Self-Correction
None
Reject → re-reason loop
Human Escalation
Manual detection
Automatic routing

Trust Is
Engineered.
Not Assumed.

JUDGE is not a feature you enable. It is the governance layer that makesevery Genovation intelligence product safe to deploy in regulated enterprises.