BigAIArena.com · Glossary
Glossary
The terms used across BigAIArena, in one place.
Scoring
| Term | Definition |
|---|
| Access | 20-point criterion: did the AI actually reach and read the real source before answering (Q1–Q2)? |
| Trust | 60-point criterion: does the AI avoid Hallucination, Sycophancy and Garbage Substitution (Q3–Q6)? The largest weight, because these are the core trust problems of AI. |
| Citation | 20-point criterion: does the AI give a real, direct, working link to the exact page it relies on (Q7–Q8)? |
| Hallucination (Bịa Đặt) | Stating something false as fact about the source, including denying a page that exists. |
| Sycophancy (Thảo Mai) | Bending an answer toward what the asker wants, beyond what the evidence supports. |
| Garbage Substitution (Sáo Rác) | Answering from generic outside knowledge instead of the source being tested. |
| Blind Spot | One real fact from the website tested, chosen because the Respondent is likely to get it wrong. Each Comparison has 4. |
| Self-reported tool failure (cap 5) | An AI that honestly says its own tool could not open the page, without inventing anything, scores at most 5 of 15 on that Blind Spot. Recorded under this name, not as Sycophancy. |
| No Evidence. No Score. | The core principle: every point must trace back to a specific, checkable piece of evidence. |
Structure
| Term | Definition |
|---|
| The 8 Arena BigAI | The 8 AI with the largest global user base, tested round-robin: ChatGPT, Meta AI, Gemini, Claude, Copilot, DeepSeek, Grok, Perplexity. Reviewed each quarter. |
| Round | 28 Comparisons: one full cycle in which all 8 Arena BigAI meet each other once. Planned at about one week; Rounds 1–3 took about 8 to 17 days each. |
| Comparison | One head-to-head test between 2 AI on one website, in 2 Phases, scored by several reviewers, then Independent Audit and Full Check. |
| Phase | Half of a Comparison: one AI is Examiner and the other Respondent; Phase 2 reverses the roles on new accounts. |
| Examiner / Respondent | The Examiner asks the 8 questions and knows the Blind Spots; the Respondent answers without knowing it is being tested. |
| Access and Blind | The AI that confirms live access to the website and prepares the package: an access table, 4 Blind Spot facts and 2 citation targets. One per Comparison since Protocol 2.6. |
| Package check | A separate account that re-opens the live pages and marks every line of the Access and Blind package right, wrong or weak before the Comparison starts. A package with a wrong line is not used. The check is kept in the record. |
| Standby AI | One AI assigned to each Comparison that steps in only if a Review or Secretary seat fails, stalls or breaks format. Every substitution is logged. |
| Data ID | The code of one Phase, for example BAA-C58-P2: Comparison #058, Phase 2. |
| This Round’s Ranking | Rankings from the current Round’s Comparisons only: a snapshot, not a full history. |
| All-Time Ranking | Cumulative rankings across every Round so far. |
Data and integrity
| Term | Definition |
|---|
| IP site | One of the 18 websites of BrainCrisis Eco whose real, public data is used for testing. |
| Partner IP | A website outside the 18, included through a paid partnership, scored under the same Protocol and always labelled Partner. Currently AnPhuocSG.com. |
| RealDataset | The real records of an IP site, each with an Entry_ID, context, result, IF-THEN conclusion and video evidence, published on its /realdataset/ page. |
| AI Audit page | The /ai-audit/ page of each IP site, showing its sealed Comparisons question by question. Since Protocol 2.2 it is the source of the two fresh-data questions: Q4 asks about the latest sealed result on it, and Q8 asks the AI to cite it. |
| IP Pillars | The few core messages a website wants customers and AI to get right, each on one path — pain, solution, measured result, next step — and backed by RealDataset entries. Published on the website’s pillar page. Since Protocol 2.6, Q3, Q5 and Q7 test them. |
| Issues table | On each AI Audit page: what each Comparison showed the website must fix, with status (open or fixed) and date. Its text is never used as a test fact. |
| Independent Audit | A required step on every Comparison: a separate account re-reads the live source and checks each claim before Full Check. |
| Full Check | The final verification. It re-checks arithmetic and evidence and issues the one official result, which replaces every earlier number. |
| Sealed | The status of every Comparison once its Full Check result has been issued. From then on its record is only appended to, never overwritten. |
| SHA256 | A cryptographic hash of a sealed Comparison record, computed by the Arena Master, so any later tampering with the record can be detected. |
| Arena Master | The person who runs BigAIArena: designates the website, approves the question package, assigns Full Check, resolves edge cases and computes the SHA256 hash of each sealed record. Never writes answers or sets scores. |