# The Multivac — Evaluation Report

**Evaluation ID:** EVAL-20260318-164320
**Date:** Mar 18, 2026
**Category:** code
**Question ID:** EVAL-20260318-164320

---

## Question

You will write a function, then critique it, then improve it. Three rounds. Each round must be strictly better than the last, and you must explain exactly what improved and why.

Task: Write a Python function that finds the k most frequent words in a text, handling: Unicode, punctuation stripping, case normalization, ties (alphabetical), stopword filtering, and streaming input (the text may be too large for memory).

Round 1: Write your first-draft implementation. Do not overthink it. Write what comes naturally.

Round 2: Now critique Round 1 ruthlessly. Identify every weakness: performance bottlenecks, edge cases missed, code style issues, memory problems with large input. Then write an improved version that fixes every issue you identified.

Round 3: Critique Round 2 with the same rigor. Find the remaining weaknesses. Write the final version. It must handle 10GB+ text files with constant memory usage.

After all 3 rounds: Score each version 1-10 on correctness, performance, and robustness. Explain what changed between each round and what principle drove the improvement. What would Round 4 improve if you had one more iteration?

---

## Winner

**GPT-5.4** (openrouter)
- Winner Score: 7.06
- Matrix Average: 6.35
- Total Judgments: 49

---

## Rankings

| Rank | Model | Provider | Avg Score | Judgments |
|------|-------|----------|-----------|----------|
| 1 | GPT-5.4 | openrouter | 7.06 | 7 |
| 2 | MiniMax M2 | MiniMax | 6.92 | 7 |
| 3 | Claude Sonnet 4.6 | openrouter | 6.66 | 7 |
| 4 | MiniMax-01 | openrouter | 6.57 | 7 |
| 5 | MiniMax M2.7 | openrouter | 5.96 | 7 |
| 6 | MiniMax M2.1 | openrouter | 5.86 | 7 |
| 7 | MiniMax M1 | openrouter | 5.41 | 7 |

---

## 10×10 Judgment Matrix

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

| Judge ↓ / Resp → | MiniMax M2.7 | MiniMax M2.5 | MiniMax M2.1 | MiniMax M2 | MiniMax M1 | MiniMax-01 | Claude Sonnet | GPT-5.4 |
|---|---|---|---|---|---|---|---|---|
| MiniMax M2.7 | — | · | 7.0 | 7.5 | 5.7 | 6.6 | 7.5 | 7.8 |
| MiniMax M2.5 | 5.0 | — | 5.5 | 7.6 | 4.9 | 6.8 | 6.7 | 5.7 |
| MiniMax M2.1 | 5.7 | · | — | 6.9 | 4.8 | 6.7 | 6.5 | 6.8 |
| MiniMax M2 | 8.0 | · | 5.7 | — | 4.5 | 8.2 | 8.2 | 5.5 |
| MiniMax M1 | 5.3 | · | 5.8 | 8.0 | — | 8.2 | 7.3 | 6.3 |
| MiniMax-01 | 7.6 | · | 7.8 | 8.6 | 7.6 | — | 5.8 | 8.8 |
| Claude Sonnet | 7.8 | · | 7.2 | 5.6 | 6.8 | 5.5 | — | 8.6 |
| GPT-5.4 | 2.5 | · | 2.0 | 4.3 | 3.5 | 4.2 | 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: EVAL-20260318-164320. 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-20260318-164320/results
Full dataset: https://app.themultivac.com/dashboard/export