# The Multivac — Evaluation Report

**Evaluation ID:** EVAL-20260402-121519
**Date:** Jan 19, 2026
**Category:** code
**Question ID:** CODE-002

---

## Question

Write a Python function that parses deeply nested JSON with the following requirements:
1. Handle missing keys gracefully (return None, don't crash)
2. Support a path syntax like "user.profile.settings.theme"
3. Handle arrays with index syntax like "users[0].name"
4. Return a typed result with proper error messages for debugging
5. Handle circular reference detection

Include type hints and comprehensive docstrings.

---

## Winner

**GPT-5.4** (openrouter)
- Winner Score: 9.13
- Matrix Average: 7.22
- Total Judgments: 88

---

## Rankings

| Rank | Model | Provider | Avg Score | Judgments |
|------|-------|----------|-----------|----------|
| 1 | GPT-5.4 | openrouter | 9.13 | 9 |
| 2 | GPT-OSS-120B | OpenAI | 8.86 | 9 |
| 3 | MiMo-V2-Flash | Xiaomi | 8.52 | 9 |
| 4 | DeepSeek V4 | openrouter | 8.16 | 9 |
| 5 | Grok 4.20 | openrouter | 7.87 | 9 |
| 6 | Gemini 3 Flash Preview | Google | 7.58 | 9 |
| 7 | Claude Opus 4.6 | openrouter | 7.26 | 9 |
| 8 | Claude Sonnet 4.6 | openrouter | 6.06 | 9 |
| 9 | MiniMax M2.5 | openrouter | 6.01 | 9 |
| 10 | Gemini 3.1 Pro | openrouter | 2.75 | 7 |

---

## 10×10 Judgment Matrix

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

| Judge ↓ / Resp → | GPT-5.4 | Claude Opus | Gemini 3.1 Pro | Claude Sonnet | Grok 4.20 | DeepSeek V4 | GPT-OSS-120B | Gemini 3 | MiniMax M2.5 | MiMo-V2-Flash |
|---|---|---|---|---|---|---|---|---|---|---|
| GPT-5.4 | — | 6.4 | 1.9 | 3.3 | 6.5 | 6.8 | 8.0 | 6.3 | 2.9 | 7.5 |
| Claude Opus | 9.2 | — | 1.6 | 6.3 | 6.0 | 7.4 | 8.6 | 7.8 | 5.7 | 8.2 |
| Gemini 3.1 Pro | 9.4 | 5.7 | — | 4.8 | 7.2 | 7.6 | 7.7 | 6.8 | 4.3 | 8.2 |
| Claude Sonnet | 9.0 | 8.0 | 1.9 | — | 6.8 | 8.2 | 8.9 | 7.5 | 6.4 | 8.2 |
| Grok 4.20 | 8.8 | 8.7 | 3.6 | 6.0 | — | 8.4 | 8.8 | 7.0 | 6.4 | 6.8 |
| DeepSeek V4 | 9.6 | 9.0 | · | 8.8 | 9.6 | — | 9.6 | 9.2 | 8.8 | 9.8 |
| GPT-OSS-120B | 8.8 | 3.4 | · | 3.8 | 8.3 | 8.8 | — | 7.8 | 3.6 | 8.6 |
| Gemini 3 | 9.8 | 9.6 | 2.9 | 9.3 | 9.6 | 9.8 | 9.8 | — | 9.0 | 9.8 |
| MiniMax M2.5 | 8.8 | 7.6 | 0.6 | 5.8 | 8.3 | 8.4 | 9.4 | 7.2 | — | 9.6 |
| MiMo-V2-Flash | 8.8 | 7.0 | 6.8 | 6.4 | 8.6 | 8.0 | 9.0 | 8.6 | 7.0 | — |

---

## 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: CODE-002. 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-20260402-121519/results
Full dataset: https://app.themultivac.com/dashboard/export