# The Multivac — Evaluation Report

**Evaluation ID:** EVAL-20260402-155654
**Date:** Feb 18, 2026
**Category:** reasoning
**Question ID:** REASON-006

---

## Question

Schedule a one-day conference with these constraints:

TALKS: A (90min), B (60min), C (45min), D (30min), E (30min), F (45min)
ROOMS: Main Hall (capacity 500), Room 2 (capacity 100), Room 3 (capacity 50)
TIME: 9:00 AM - 5:00 PM, with mandatory lunch break 12:00-1:00 PM

CONSTRAINTS:
1. Talk A must be in Main Hall (expected attendance: 400)
2. Talk B and C cannot overlap (same speaker)
3. Talk D must be before Talk E (E builds on D's content)
4. Talk F requires Room 2's AV equipment
5. No room can have more than 3 talks total
6. At least one talk must be running at all times (except lunch)
7. Talk A cannot start before 10:00 AM (speaker arriving late)
8. Talk E must end by 3:00 PM (speaker leaving early)

Find a valid schedule or prove none exists.

---

## Winner

**GPT-5.4** (openrouter)
- Winner Score: 8.32
- Matrix Average: 5.13
- Total Judgments: 72

---

## Rankings

| Rank | Model | Provider | Avg Score | Judgments |
|------|-------|----------|-----------|----------|
| 1 | GPT-5.4 | openrouter | 8.32 | 8 |
| 2 | Grok 4.20 | openrouter | 7.38 | 8 |
| 3 | Claude Sonnet 4.6 | openrouter | 6.92 | 8 |
| 4 | Claude Opus 4.6 | openrouter | 5.83 | 9 |
| 5 | DeepSeek V4 | openrouter | 5.71 | 9 |
| 6 | MiMo-V2-Flash | Xiaomi | 5.35 | 9 |
| 7 | GPT-OSS-120B | OpenAI | 5.35 | 2 |
| 8 | Gemini 2.5 Flash | openrouter | 4.62 | 9 |
| 9 | Gemini 3.1 Pro | openrouter | 1.51 | 8 |
| 10 | MiniMax M2.5 | openrouter | 0.35 | 2 |

---

## 10×10 Judgment Matrix

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

| Judge ↓ / Resp → | DeepSeek V4 | GPT-OSS-120B | Gemini 3.1 Pro | Claude Opus | GPT-5.4 | Grok 4.20 | Claude Sonnet | MiMo-V2-Flash | Gemini 2.5 | MiniMax M2.5 |
|---|---|---|---|---|---|---|---|---|---|---|
| DeepSeek V4 | — | · | 2.9 | 9.6 | 9.6 | 9.3 | 8.6 | 9.3 | 8.1 | · |
| GPT-OSS-120B | 4.4 | — | 1.9 | 3.0 | · | 7.2 | 9.1 | 3.5 | 3.4 | · |
| Gemini 3.1 Pro | 3.4 | 0.7 | — | 3.9 | 9.4 | 3.8 | 4.0 | 3.8 | 1.8 | 0.5 |
| Claude Opus | 6.5 | · | 0.2 | — | 4.0 | 5.8 | 7.4 | 5.2 | 4.5 | · |
| GPT-5.4 | 2.6 | · | 0.7 | 3.8 | — | 6.0 | 8.8 | 3.8 | 2.6 | 0.2 |
| Grok 4.20 | 6.0 | 10.0 | 1.9 | 4.5 | 8.3 | — | 4.5 | 4.4 | 5.5 | · |
| Claude Sonnet | 8.0 | · | 1.0 | 5.0 | 7.0 | 8.3 | — | 7.3 | 4.8 | · |
| MiMo-V2-Flash | 7.8 | · | 2.2 | 8.2 | 8.7 | 9.0 | 6.0 | — | 6.0 | · |
| Gemini 2.5 | 6.8 | · | · | 7.6 | 10.0 | 9.8 | 7.0 | 4.4 | — | · |
| MiniMax M2.5 | 5.8 | · | 1.4 | 7.0 | 9.6 | · | · | 6.5 | 4.8 | — |

---

## 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-006. 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-20260402-155654/results
Full dataset: https://app.themultivac.com/dashboard/export