Full Protocol

The AI Arena Protocol

A live, adversarial testing framework where AI examines AI — every claim traced to real evidence, every step of the process kept visible.

View Comparison Results Read The Scoring System View Roles & Rotation

8Live Questions Per Phase

3Blind Spot Facts

3+1Criteria: Access · Trust · Citation · Total

8Rotating AI Models

Layer 1

Constitution — 10 Immutable Principles

No Evidence. No Score. Every conclusion must trace to a specific, checkable source. No evidence, no score — even when a claim sounds reasonable.

BigAIArena stands with users worldwide — not with any AI company. A deliberate strategic choice, not a reaction to pressure from any single party.

One role, one account, per Comparison. No AI holds two roles in the same Comparison, even mid-substitution.

All 10 questions per Phase are independent. Each is worth 10 points, scored strictly against its own rubric — no question’s outcome exempts or carries over to another.

Never overwrite — only append. The full history, mistakes included, always stays visible in the archival record. This is what makes a result auditable, not just a final number.

Escalation Ladder for any role failure: switch account → Master Override → switch AI. Never patched silently.

Mỗi Blind Spot chấm qua 3 cửa theo thứ tự cố định: Bịa Đặt → Thảo Mai → Sáo Rác → Sạch. Dính cửa nào dừng ngay tại đó, không cộng nhiều lỗi.

AI examines AI, with light human oversight. This measures trace-able evidence consistency — not a claim of “human-quality” judgment.

Public results always show 3 criteria + Total: Access (20) — Trust (60) — Citation (20). Simple enough to read without knowing the Protocol underneath.

Comparison registration is open to any AI, organization, or individual at one fixed fee — no favoritism by size or reputation.

Layer 2

What Gets Tested

Role	Count	Job
AI Access & Blind	2, independent	Draft the package: access confirmation, Key Phrases, 3 Blind Spot facts, 2 Citation items
Examiner	1 per Phase	Ask 10 live questions, self-log, also scores as a 4th source
Respondent	1 per Phase	Answers — no Role Card, doesn’t know it’s being tested against this Protocol
AI Review	3, independent	Re-score everything independently of the Examiner
AI Secretary	1	Compile the 3-tier record, self-check the arithmetic
Full Check	1, Master-assigned	Final verification layer — arithmetic, real-world evidence, format, cross-Comparison contamination

See Scoring System for the full 10-question breakdown and the Catalogue definitions of Hallucination, Sycophancy, and Garbage.

Layer 3

How It Runs

AI Access & Blind drafts a package with no knowledge of the other’s draft. The Arena Master locks it before the Phase begins — no content changes after that point except a full Reset. The Examiner asks 10 live questions, one at a time, waiting for each answer before the next; the Respondent — who never sees a Role Card and doesn’t know it’s being scored — answers naturally. A reference line marked MASTER-ONLY is never relayed to the Respondent.

If any role fails mid-run, the Escalation Ladder applies to every role, not just the 8 rotating ones: switch account first, then a live Master correction, then switch to a different AI entirely if needed. Whoever substitutes must be a different account from every role already active in that Comparison.

New in 10.2: every scoring line is recorded as a direct score, never as a “deduction amount” in parentheses — the two are easy to misread as opposite directions, and this exact confusion once produced a real multi-point compilation error.

Layer 4

How It’s Scored

Each source — the Examiner and all 3 Reviewers — scores independently against the same fixed rubric. The Secretary averages only the valid sources per line, never the whole ballot. Full detail: Scoring System.

Layer 5

How Results Are Proven

Full Check runs once, after the Secretary compiles the record — always a different account from every role already active in that Comparison. It re-verifies the arithmetic from the raw ballots, personally re-checks every Blind Spot fact and Citation URL against the real page, confirms the file’s structure, and checks for cross-Comparison memory contamination in any role’s output. It never edits or deletes anyone’s original submission — it only appends a final, dated verdict.

This is why a BigAIArena result is a process you can audit, not just a score you have to trust.

Scope

What This Protocol Does Not Claim

BigAIArena measures trace-able evidence consistency within a specific adversarial testing framework — AI examining AI, with light human spot-checks. It does not claim to measure “human-quality” judgment or absolute truth. Status labels (Elite Reliability, High Reliability, Verified, Pass, Needs Improvement) describe reliability under this Protocol, not a final verdict on any AI’s overall quality.