Kritarch Lite — AI Jury

Elapsed: 0:00

Hard questions deserve more than one perspective. Kritarch Lite puts three AI agents on your problem — each analyzes independently, critiques the others, and they converge on a verdict. You get structured reasoning, not a single guess.

Jump to debate Jump to glossary

Command center

Agent roster

EvidenceRiskSpeed

Three specialists debate every prompt. Together they deliver a verdict that is safer, sharper, and ready for action.

Juror 1

Cautious Analyst

Risk & compliance control

Operational

Stabilizes the debate with rigorous checks and uncertainty flags.

Validates assumptions and flags unknowns.
Surfaces edge cases and regulatory risk triggers.
Recommends guardrails and fallback paths.

Bias: Risk-averseLens: Evidence-firstOutput: Safety notes

Evidence

Risk

Speed

Juror 2

Devil's Advocate

Counterfactual pressure tests

Operational

Challenges consensus to prevent blind spots and groupthink.

Introduces counterexamples and hard constraints.
Stress-tests weak logic and shaky evidence.
Reframes the prompt from adversarial angles.

Bias: ContrarianLens: Red-teamOutput: Gaps found

Evidence

Risk

Speed

Juror 3

Pragmatic Expert

Decision-ready synthesis

Operational

Turns the debate into an actionable plan with clear tradeoffs.

Balances risks with execution realities.
Prioritizes next steps and owners.
Delivers concise, decision-ready guidance.

Bias: Execution-focusedLens: Tradeoff-awareOutput: Action plan

Evidence

Risk

Speed

Jury chamber

The jury issues the final verdict

Chief Justice

After the three agents debate, the Chief Justice synthesizes their positions into one decision-ready verdict. If there is disagreement, it is called out explicitly — no silent averages.

Consensus synthesis with dissent flags.
Confidence-weighted verdict framing.
Actionable next steps with tradeoffs.

Juror AJuror BJuror CChief JusticeVerdict

Verdict payload

Verdict:Final decision with rationale.
Key evidence:Most persuasive facts or signals.
Risk flags:Open questions and failure modes.
Next actions:Practical steps to move forward.

Jury charter

Balanced: evidence, risk, and actionability.

Transparent: dissent and low-confidence points are highlighted.

Decision-ready: outputs are scoped for real-world approval.

Use cases

High-stakes domains

Finance
Credit, risk, and compliance decisions.
Healthcare
Guideline comparisons and triage framing.
Legal
Policy compliance and argument mapping.

Finance lending example: Go / No-Go lending decisions that enforce government SOPs and company SOPs.

Human team analogy

Teams delegate research to members and seniors, regroup to compare findings, and stakeholders make the final call. Kritarch Lite mirrors that flow with AI agents — faster, broader in scope, 24/7 — while keeping a human in the loop.

ModelAlternate baseline: gpt-5-mini

Status: idle

Baselines

Round 1

Round 2

Round 3

Verdict

Complete

Ready when you are — enter a question to start the debate.Elapsed: 0:00

Debate rounds

Compare all jurors within each round to trace convergence and dissent.

Auto-following live round

Jump to roundRound 1 Round 2 Round 3

Round statusRound 1 PendingRound 2 PendingRound 3 Pending

Juror A

Cautious Analyst

Awaiting stance

Awaiting position summary…

Juror B

Devil's Advocate

Awaiting stance

Awaiting position summary…

Juror C

Pragmatic Expert

Awaiting stance

Awaiting position summary…

Juror A

Cautious Analyst

Awaiting position...

Streaming response...

Juror B

Devil's Advocate

Awaiting position...

Streaming response...

Juror C

Pragmatic Expert

Awaiting position...

Streaming response...

Why this wins

Multi-agent consensus keeps dissent visible and delivers decision-ready outputs.

Jury Verdict

Verdict will appear after Round 3.

Baselines vs Jury

Compare single-model baselines to the multi-agent verdict.

Side-by-side

Baseline (selected model)

Primary

Selected-model baseline will appear once the debate starts.

Baseline (alternate model)

Alternate

Alternate baseline will appear once the debate starts.

AI Jury

Multi-agent

Verdict will appear after Round 3.

Glossary

Key terms

Baseline: Two single-model baselines: one using your selected model (fairness) and one using the alternate model (cost/quality contrast).
Jurors: Three agents with distinct reasoning styles.
Chief Justice: Aggregates juror outputs into the final verdict.
Rounds: Pipeline steps: baselines, positions, critique, rebuttal (deep deliberation only), revision, verdict.
Critique: Round 2 where jurors challenge each other's positions.
Position: Juror stance: support, oppose, or nuanced.
Critique assessment: Overall verdict on a juror's position: strong, moderate, weak.
Critique severity: Per-issue severity tags: minor, moderate, major.
Revision: Round 3 updates after critiques (sometimes skipped).
Verdict: The Chief Justice's consensus synthesis.
Evaluation: Independent scoring that compares the jury output against the selected-model baseline.
Safety guardrails: Preflight prompt-injection heuristics + moderation checks; unsafe requests are blocked (fail-closed if moderation is unavailable).
Prompt injection: Attempts to override instructions or reveal hidden prompts; detected and blocked.
Redaction: If unsafe output is detected, responses are masked before streaming to the UI.
Cost estimates: Costs depend on the selected model. gpt-5.2 (~$1.75 in / $14 out per 1M tokens) vs gpt-5-mini (~$0.25 in / $2 out). Jurors/Verdict/Evaluator use the selected model; the alternate baseline uses the other. Estimates are approximate and can be overridden.
Coordination: Agreement-based control that can deepen or skip rounds.
Agreement score: How aligned the jurors are; higher scores can fast-track the flow.
Average confidence: Mean juror confidence used alongside agreement to skip rounds.
Disagreement focus: Key conflicts extracted to guide the critique prompts.
Skip logic: Three modes: fast-track (high agreement), standard, or deep deliberation (low agreement adds rebuttal round).
Deep deliberation: Activated when agreement < 40%. Adds a rebuttal round between critique and revision so jurors respond to challenges before revising.
Rebuttal: Extra round in deep deliberation where jurors concede valid critique points, defend strong arguments, and refine their positions.