Council

Xpersona
one task · multiple frontier models · one verified answer

Ask several frontier models the same task. Compare the evidence side by side. Judge them against criteria you control. Continue with the winner. Xpersona Council turns your multi-model catalog into a real decision and execution layer.

MODEL FAMILIES3
ELIGIBLE CANDIDATES16
JUDGINGBlind + rubric
ACCESSCredits + packages
CORE MESSAGE
One task. Multiple models. One verified answer.
JUDGING MODES
AccuracyCodingReasoningCustom rubric

How Council works

Submit once. Let multiple models answer. Judge. Synthesize. Continue.

1 · Submit one task

One task. Multiple frontier models. Ask GPT, Claude, Gemini, and Xpersona models the same question.

2 · Select candidates

Choose 2–4 real models from the live catalog. Each runs through genuine Xpersona inference.

3 · Blind judging

An independent judge scores anonymized responses against your rubric before any model identity is revealed.

4 · Compare + synthesize

Ranked results, cost/latency/tokens, and an optional synthesis that preserves disagreement.

5 · Continue with the winner

Open a normal Xpersona chat seeded with the winning response and provenance.

A real comparison

Representative form — actual costs, latency, and winners vary by task and run.

Task: "Draft a clear, correct SQL query for monthly active users."3 candidates · blind judging · synthesis off

Response A

-- selects each distinct user with any event in the month
SELECT month, COUNT(DISTINCT user_id) AS mau
FROM events
WHERE occurred_at >= '2026-01-01'
  AND occurred_at <  '2026-02-01'
GROUP BY month;

Response B

SELECT
  DATE_TRUNC('month', occurred_at) AS month,
  COUNT(DISTINCT user_id)          AS mau
FROM events
GROUP BY 1
ORDER BY 1;
WinnerResponse B — verified after reveal
Judge cost · tokens · latency$0.009 · 1,184 tokens · 2.1s
This is an illustrative example of the comparison format, not a fixed outcome. Council reports the observed cost, latency, and tokens for each actual run.

Built for evidence, not vibes

Private by default

Proof receipts stay private until you explicitly publish them under an unguessable slug.

Genuine blind mode

Judge sees Response A/B/C/D, never which model produced which answer.

One verified verdict

Validated structured scores, evidence, and a single winner — or an honest tie.

xpersona-auto explains itself

A rules-based auto-router returns the selected model and the reason behind it.

Candidates from the live catalog

Council populates selections from the same model catalog that powers /v1/models.

OPENAI · ANTHROPIC · GOOGLE · XPERSONA
Xpersona Frieren 1Xpersona A_BETTER_YOUClaude Fable 5Claude Opus LatestClaude Sonnet LatestOpenAI GPT LatestGPT-5.6GPT-5.6 SolGPT-5.6 TerraGPT-5.5GPT-5.4GPT-5.4 Mini

Ask several frontier models. Compare the evidence. Continue with the winner.

Council consumes your prepaid credits or monthly package capacity per model call.

Open the workspace