Code-native resource
Agent evaluation canvas
A printable canvas for tasks, trials, graders, traces, harness conditions, failures, and review decisions.
Purpose
What this structure helps make visible
Keep evaluation design and its operational assumptions visible in one reviewable artifact.
Use sequence
Move through the artifact.
- Define the task and environment.
- Describe trial variation and grading logic.
- Record trace needs, harness assumptions, failure conditions, and release decisions.
Printable working view
Agent evaluation canvas
Task
Name observable behavior and boundaries.
Trial
Describe meaningful variation.
Grader
State decision logic and limitations.
Trace
Retain only diagnosis-relevant evidence.
Harness
Record environmental assumptions.
Decision
Connect evidence to the release gate.
Provenance
Source basis
External source links open in a new tab.
- Demystifying evals for AI agents
Task, trial, grader, trace, and harness vocabulary
Rights and provenance
The resource remains traceable.
- Origin
- code-native-original
- Holder
- Lucifer Rodstark, Ph.D.
- License
- CC BY 4.0
- Reviewed
- 2026-07-17
This license covers LucyAI's original code-native resource expression only; source facts, names, marks, and third-party materials are not relicensed.
Credit LucyAI and link to the original resource page. The original resource structure may be shared and adapted under CC BY 4.0. Cited sources remain under their own terms, and no AI-generated image is included in this license.