# The Multivac — Evaluation Report

**Evaluation ID:** EVAL-20260402-181028
**Date:** Apr 02, 2026
**Category:** reasoning
**Question ID:** REASON-026

---

## Question

A teacher gives a test. Students who scored in the top 10% get praised. Students who scored in the bottom 10% get extra tutoring. On the next test, the top scorers decline slightly and the bottom scorers improve. The teacher concludes: 'Praise is counterproductive, but tutoring works.' (1) What's actually happening? (2) Design a study that separates regression to the mean from real effects. (3) Give three real-world examples where this fallacy leads to bad policy decisions.

---

## Winner

**Claude Opus 4.6** (openrouter)
- Winner Score: 9.61
- Matrix Average: 9.09
- Total Judgments: 90

---

## Rankings

| Rank | Model | Provider | Avg Score | Judgments |
|------|-------|----------|-----------|----------|
| 1 | Claude Opus 4.6 | openrouter | 9.61 | 9 |
| 2 | GPT-5.4 | openrouter | 9.31 | 9 |
| 3 | MiMo-V2-Flash | Xiaomi | 9.19 | 9 |
| 4 | GPT-OSS-120B | OpenAI | 9.16 | 9 |
| 5 | MiniMax M2.5 | openrouter | 9.15 | 9 |
| 6 | Grok 4.20 | openrouter | 9.13 | 9 |
| 7 | Gemini 2.5 Flash | openrouter | 9.11 | 9 |
| 8 | Claude Sonnet 4.6 | openrouter | 9.04 | 9 |
| 9 | DeepSeek V4 | openrouter | 8.97 | 9 |
| 10 | Gemini 3.1 Pro | openrouter | 8.18 | 9 |

---

## 10×10 Judgment Matrix

Rows = Judge, Columns = Respondent. Self-judgments excluded (—).

| Judge ↓ / Resp → | GPT-OSS-120B | Gemini 3.1 Pro | DeepSeek V4 | Claude Opus | GPT-5.4 | Grok 4.20 | Claude Sonnet | Gemini 2.5 | MiMo-V2-Flash | MiniMax M2.5 |
|---|---|---|---|---|---|---|---|---|---|---|
| GPT-OSS-120B | — | 7.5 | 8.7 | 8.7 | 8.4 | 8.4 | 8.4 | 8.7 | 8.7 | 8.7 |
| Gemini 3.1 Pro | 9.4 | — | 9.7 | 10.0 | 10.0 | 9.4 | 8.4 | 10.0 | 10.0 | 9.0 |
| DeepSeek V4 | 9.4 | 8.6 | — | 10.0 | 9.4 | 9.4 | 9.7 | 9.4 | 9.4 | 9.4 |
| Claude Opus | 9.4 | 8.3 | 9.0 | — | 9.4 | 9.2 | 9.2 | 9.2 | 9.2 | 9.0 |
| GPT-5.4 | 8.2 | 7.3 | 8.6 | 9.8 | — | 8.8 | 8.1 | 9.1 | 8.8 | 8.7 |
| Grok 4.20 | 8.8 | 8.1 | 8.8 | 9.0 | 8.8 | — | 9.0 | 8.8 | 8.8 | 8.8 |
| Claude Sonnet | 9.6 | 8.3 | 8.3 | 10.0 | 9.6 | 9.4 | — | 8.8 | 9.0 | 9.0 |
| Gemini 2.5 | 10.0 | 9.4 | 10.0 | 10.0 | 9.4 | 9.4 | 10.0 | — | 10.0 | 10.0 |
| MiMo-V2-Flash | 9.2 | 8.6 | 9.0 | 10.0 | 9.8 | 9.2 | 10.0 | 9.2 | — | 9.8 |
| MiniMax M2.5 | 8.4 | 7.7 | 8.8 | 9.0 | 8.8 | 8.8 | 8.6 | 8.8 | 8.8 | — |

---

## Methodology

- **10×10 Blind Peer Matrix:** All models answer the same question, then all models judge all responses.
- **5 Criteria:** Correctness, completeness, clarity, depth, usefulness (each scored 1–10).
- **Self-judgments excluded:** Models do not judge their own responses.
- **Weighted Score:** Composite of all 5 criteria.

---

## Citation

The Multivac (2026). Blind Peer Evaluation: REASON-026. app.themultivac.com

## License

Open data. Free to use, share, and build upon. Please cite The Multivac when using this data.

Download raw JSON: https://app.themultivac.com/api/evaluations/EVAL-20260402-181028/results
Full dataset: https://app.themultivac.com/dashboard/export