Built-in Criteria

Rewardkit ships with built-in criteria for common rubrics. Use them from any Python file in your tests directory:

import rewardkit as rk

rk.file_exists("output.txt")
rk.command_succeeds("python main.py", weight=2.0)

All criteria accept optional weight (default 1.0) and isolated (default false) parameters in addition to the ones listed below.

File criteria

Criterion	Parameters	Description
`file_exists`	`path`	File exists in workspace
`file_not_exists`	`path`	File does not exist
`file_contains`	`path, text`	File contains a substring
`file_contains_regex`	`path, pattern`	File content matches a regex pattern
`file_matches`	`path, expected`	File content equals expected text (whitespace-stripped)
`files_equal`	`path1, path2`	Two files have identical content
`diff_ratio`	`path, expected`	Similarity ratio between file content and expected text (returns 0.0–1.0)

Command criteria

Criterion	Parameters	Description
`command_succeeds`	`cmd, cwd?, timeout?`	Command exits with code 0
`command_output_contains`	`cmd, text, cwd?, timeout?`	Command stdout contains text
`command_output_matches`	`cmd, expected, cwd?, timeout?`	Command stdout equals expected (stripped)
`command_output_matches_regex`	`cmd, pattern, cwd?, timeout?`	Command stdout matches a regex

Default timeout is 30 seconds. The cwd parameter is relative to the workspace.

Data format criteria

Criterion	Parameters	Description
`json_key_equals`	`path, key, expected`	Top-level JSON key equals a value
`json_path_equals`	`path, json_path, expected`	Dot-separated path into JSON equals a value
`csv_cell_equals`	`path, row, col, expected`	CSV cell at row/col equals a value
`xlsx_cell_equals`	`path, cell, expected, sheet?`	Excel cell equals a value
`sqlite_query_equals`	`db_path, query, expected`	SQL query result equals a value

xlsx_cell_equals requires the office extra: uv add harbor-rewardkit[office]

For csv_cell_equals, row numbering depends on the column type. When col is an integer, a raw CSV reader is used and row 0 is the header row. When col is a string (column name), row 0 is the first data row after the header.

HTTP criteria

Criterion	Parameters	Description
`http_status_equals`	`url, status?, timeout?`	HTTP response has the expected status code (default 200)
`http_response_contains`	`url, text, timeout?`	HTTP response body contains text

Image criteria

Criterion	Parameters	Description
`image_similarity`	`path1, path2`	Pixel-level similarity ratio (returns 0.0–1.0)
`image_size_equals`	`path, width, height`	Image has the expected dimensions

Image criteria require the image extra: uv add harbor-rewardkit[image]

Trajectory criteria

These criteria inspect the agent's ATIF trajectory file (default path: /logs/trajectory.json).

Criterion	Parameters	Description
`trajectory_tool_used`	`tool_name, min_count?, path?`	Agent used a specific tool at least `min_count` times (default 1)
`trajectory_tool_not_used`	`tool_name, path?`	Agent did not use a specific tool
`trajectory_turn_count`	`max_turns, path?`	Penalizes exceeding a turn budget — returns 1.0 at `max_turns`, linearly decays to 0.0 at double

Optional extras

Extra	Criteria	Install
`office`	`xlsx_cell_equals`	`uv add harbor-rewardkit[office]`
`image`	`image_similarity`, `image_size_equals`	`uv add harbor-rewardkit[image]`
`all`	All of the above	`uv add harbor-rewardkit[all]`