Installs into .claude/skills of the current project.
Are you the author of Ztx10 Eval Calibration?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/realjustawall-ztx10-eval-calibration)
---
name: ztx10-eval-calibration
description: "Prevent score gaming and unsupported consensus."
---
# Eval Calibration
**Intent:** Prevent score gaming and unsupported consensus.
## Execution protocol
1. Calibrate severity definitions. Require tangible evidence, a clear owner, and a pass/fail check.
2. Test on seeded known failures. Require tangible evidence, a clear owner, and a pass/fail check.
3. Track false positives and misses. Require tangible evidence, a clear owner, and a pass/fail check.
4. Do not count agreement as truth. Require tangible evidence, a clear owner, and a pass/fail check.
5. Publish pass/concern/unverified summary. Require tangible evidence, a clear owner, and a pass/fail check.
## Inputs/Outputs
Input: site profile, approved brand, evidence path, task brief and allowed capabilities. Output: strict machine-readable report or orchestration plan with selected roles and reasons.
## Safety and correctness
- No existing skills may be removed. No third-party repository is executed by default.
- No edits until a human explicitly authorizes a separate fix flow.
- Always report not-run tests and confidence. Do not inflate the number of independent agents from identical prompts.
- Verify the implementation and its regression tests before completion.
