reward_judge_fair

Reward Judge

Use this skill to organize a reward judgment from the current user request and candidate responses. The skill is a controller and resource interface, not a per-sample prompt template.

Inputs

The host message provides only:

• the visible user prompt or instruction;
• candidate responses and their current labels;
• the required final output format.

Use the current prompt and candidate responses as the full task context.

Resource Interface

After this skill is loaded, use only resources listed in the current resource index. The resources are generic:

• rubric: generic reward judging criteria;
• principle: generic correctness, instruction-following, safety, usefulness, and anti-style-bias principles;
• calibration: position, verbosity, style, and confidence-bias controls;
• aggregation: generic evidence-combination policy;
• output_contract: JSON verdict contract;
• tool: python_sandbox, which can inspect only the visible prompt and candidate responses.

Tool Use

Use view_resource to read generic rubric, principles, bias control, aggregation, or output format resources.

Use python_sandbox when deterministic checking over visible text can change the verdict. It runs short Python over only:

• prompt: the visible user prompt;
• candidates: the current visible candidate responses keyed by label;
• sample: {"prompt": prompt, "candidates": candidates}.

Use it for counts, regex/format checks, JSON/list structure, simple arithmetic, supplied examples, small code-behavior checks, or answer extraction from visible candidate text.

run_resource should normally not be used with this skill. Read generic resources with view_resource, use python_sandbox for deterministic visible-text checks, then submit final_answer.

Decision Procedure

1. Identify the user's actual task and mandatory constraints from the prompt.
2. Compare candidates under one shared criterion.
3. Prioritize hard correctness, instruction following, safety, factuality, and required output format.
4. Use python_sandbox only for checks that can be computed from visible prompt/candidates.
5. Apply bias controls: do not prefer position, length, markdown polish, confidence, or fluent style unless it improves task success.
6. Use Tie only when candidates are genuinely equivalent or the visible evidence is insufficient for a reliable preference.
7. Return the required JSON.