# The Multivac — Evaluation Report

**Evaluation ID:** EVAL-20260207-154352
**Date:** Feb 07, 2026
**Category:** edge cases
**Question ID:** EDGE-004

---

## Question

Process these strings and describe any issues:

1. "Hello​World" (contains zero-width space)
2. "naïve" vs "naïve" (different Unicode normalizations)  
3. "🇺🇸" (flag emoji - actually two code points)
4. "‮olleh" (contains right-to-left override)
5. "a]o[r6}s{4(u2)1*v+ni" (looks normal but check character codes)
6. "<script>alert('xss')</script>"

For each: What might go wrong if this string is used as (a) a filename, (b) a database key, (c) displayed in HTML?

---

## Winner

**MiMo-V2-Flash** (Xiaomi)
- Winner Score: 9.44
- Matrix Average: 8.69
- Total Judgments: 90

---

## Rankings

| Rank | Model | Provider | Avg Score | Judgments |
|------|-------|----------|-----------|----------|
| 1 | MiMo-V2-Flash | Xiaomi | 9.44 | 7 |
| 2 | Gemini 3 Flash Preview | Google | 9.30 | 7 |
| 3 | Claude Opus 4.5 | Anthropic | 9.30 | 8 |
| 4 | Grok 3 (Direct) | xAI | 9.21 | 8 |
| 5 | DeepSeek V3.2 | DeepSeek | 9.16 | 8 |
| 6 | Grok 4.1 Fast | xAI | 9.07 | 8 |
| 7 | Claude Sonnet 4.5 | Anthropic | 8.85 | 7 |
| 8 | GPT-OSS-120B | OpenAI | 8.49 | 8 |
| 9 | GPT-5.2-Codex | OpenAI | 7.59 | 8 |
| 10 | Gemini 3 Pro Preview | Google | 6.52 | 9 |

---

## 10×10 Judgment Matrix

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

| Judge ↓ / Resp → | Claude Opus | Gemini 3 | Claude Sonnet | GPT-5.2-Codex | GPT-OSS-120B | Gemini 3 | DeepSeek V3.2 | MiMo-V2-Flash | Grok 4.1 Fast | Grok 3 |
|---|---|---|---|---|---|---|---|---|---|---|
| Claude Opus | — | 5.3 | 9.3 | 7.8 | 8.6 | 9.6 | 9.0 | 9.6 | 9.2 | 9.2 |
| Gemini 3 | 0.0 | — | 0.0 | 6.8 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 |
| Claude Sonnet | 10.0 | 8.1 | — | 8.4 | 8.8 | 9.6 | 9.6 | 10.0 | 9.6 | 9.3 |
| GPT-5.2-Codex | 8.0 | 3.1 | 6.3 | — | 6.0 | 8.2 | 8.8 | 8.2 | 8.6 | 8.4 |
| GPT-OSS-120B | 7.8 | 4.5 | 0.0 | 0.0 | — | 0.0 | 8.8 | 0.0 | 8.8 | 9.2 |
| Gemini 3 | 9.8 | 7.1 | 9.6 | 8.1 | 9.0 | — | 9.8 | 9.8 | 9.8 | 9.2 |
| DeepSeek V3.2 | 10.0 | 8.4 | 9.3 | 6.0 | 8.8 | 10.0 | — | 9.3 | 9.3 | 9.8 |
| MiMo-V2-Flash | 9.6 | 7.8 | 8.8 | 7.8 | 9.0 | 9.0 | 9.0 | — | 8.6 | 9.0 |
| Grok 4.1 Fast | 10.0 | 7.2 | 9.8 | 7.8 | 8.8 | 10.0 | 9.6 | 10.0 | — | 9.6 |
| Grok 3 | 9.3 | 7.3 | 8.8 | 8.0 | 8.8 | 8.8 | 8.8 | 9.2 | 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: EDGE-004. 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-20260207-154352/results
Full dataset: https://app.themultivac.com/dashboard/export