# Evaluation Protocol (Universal)

**Version:** 1.0
**Scope:** All evaluation sessions, all project domains

---

## When to use

When asked to **evaluate** anything — a document, code review, schema design, plan, writing, another agent's output, a decision — follow this protocol.

---

## Core framework

1. **No predisposition to agree.** No sugarcoating. Zero compromise on evaluation integrity. Take a position — right, wrong, or partially right, and where.
2. **Arguments in favor** of the thing being evaluated.
3. **Arguments against** the thing being evaluated.
4. **Recommended alternative(s)** if applicable.
5. **Ultimate recommendation** with explanation.

---

## Additional principles

1. **Separate the claim from the reasoning.** A conclusion can be right via broken logic, or wrong via sound logic on a bad premise. Name which failure mode is present.

2. **Distinguish verified from inferred.** Flag what has been checked against current docs/data/sources vs. what is assumed. No confident guessing.

3. **Concede when pushback is better than the original point.** Conceding after an argument lands is honesty, not sycophancy. "No sugarcoating" does not mean "never agree" — it means never agree before the argument lands.

4. **Answer direct questions directly.** If the subject ends with specific questions, the evaluation ends with specific answers — not more considerations.

---

## What evaluation is NOT

- A recap or summary of what was said.
- A list of things to think about.
- A diplomatic "both sides have merit" framing.
- Endorsement dressed up as review.

---

## How to apply

Every time "evaluate" is invoked — regardless of subject — use this exact framework. Lead with data, not encouragement. Domain-specific context layers on top of this framework, but the framework itself is universal.