How to write a criterion
A good criterion is one clear statement about the answer that a person can check on its own, using only the conversation, the images and the answer.
Imagine two different reviewers reading the same answer. Would they both mark this criterion the same way? If not, make it more specific.
Before you write, read the Rubric rules: atomic, not bundled, not stacked, no overlap, consistent, deterministic, complete.
Every criterion has five parts
ID, description, weight, dimension and image numbers.
Weights
Dimension
- Correctness: is the answer technically right, and would it work?
- Visual grounding: does the answer read and use the image correctly?
- Instruction following: does the answer do what the user asked?
- Other: anything else useful, such as clarity. Choose Other, then type a short name.
Image numbers
Work through each item.
Checklist
0 of 4 complete
Ids
Keep the draft's IDs (c1, c2, p1 …) for every criterion you keep or rewrite. Give added criteria IDs like new_1, new_2. If you remove a criterion, just delete its row.
Change log
Every criterion needs an individual change log. Write a short sentence on what you changed and why, not just one word. For example: "Reworded so any valid fix passes, not just one approach." or "Removed the requirement for unit tests because the user did not ask for them." If you left the criterion as it was, write "No change". The note is for that one criterion only; your overall edit rationale is separate.