# Criterion Synthesis

> Synthesize a task-specific executable scientific rubric from an instruction-derived skeleton, literature grounding, and task-data profile. Use in AutoSciRub when Codex needs criteria with linked goals, data sources, required analyses, metrics, comparisons, expected artifacts, provenance, and satisfaction conditions.

- Skill: `zjunlp/criterion-synthesis` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add zjunlp/criterion-synthesis`
- Raw SKILL.md: https://api.skillmd.com/api/skills/zjunlp/criterion-synthesis/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Research & Search
- Author: zjunlp (https://skillmd.com/u/zjunlp)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/zjunlp/criterion-synthesis

---


# Criterion Synthesis

## Purpose

Combine instruction goals, scientific practice, and task-data feasibility into an executable rubric. This implements `phi_syn(G_i, K_i, P_i)` from AutoSciRub.

The rubric is the contract for execution, verification, and revision.

All `.autoscirub/` paths below refer to the state directory resolved by the controller. When invoked alone, use the user’s override, then `AUTOSCIRUB_STATE_DIR`, then project config `state_dir`, then `.autoscirub/`.

## Inputs

Read:

1. `.autoscirub/rubric_skeleton.json`
2. `.autoscirub/literature_grounding.json`
3. `.autoscirub/task_data_profile.json`
4. optional task config/profile

## Output

Write `.autoscirub/executable_rubric.json`:

```json
{
  "schema_version": "1.0",
  "goals": [
    {
      "goal_id": "G1",
      "title": "...",
      "requirement": "..."
    }
  ],
  "criteria": [
    {
      "criterion_id": "C1",
      "goal_ids": ["G1"],
      "requirement": "Concrete requirement to satisfy.",
      "data_sources": ["relative/path or dataset field"],
      "required_analysis": ["experiment, computation, comparison, proof step, or diagnostic"],
      "metrics": ["metric, statistic, threshold, or qualitative check"],
      "comparisons": ["baseline, control, condition, region, time, model, or ablation"],
      "expected_artifacts": [
        {
          "type": "figure | table | number | file | code | text_analysis",
          "name": "semantic artifact name",
          "expected_path": "optional relative path",
          "must_show": "what the artifact must demonstrate"
        }
      ],
      "satisfaction_condition": "Observable condition for marking the criterion satisfied.",
      "priority": "high | medium | low",
      "provenance": {
        "instruction": ["goal or instruction evidence"],
        "literature": ["source id or summarized guidance"],
        "data_profile": ["supporting files or feasibility notes"]
      }
    }
  ],
  "claims_to_avoid": ["unsupported claim pattern"]
}
```

## Procedure

1. Preserve every instruction-derived goal. If required inputs are missing, retain the goal and record the blocker; do not silently remove the user’s requirement.
2. Select feasible analyses that best satisfy each goal using literature-grounded practice.
3. Convert each requirement into a criterion with observable evidence and a satisfaction condition.
4. Link each criterion to goals, data sources, artifacts, and provenance.
5. Record unsupported or risky claims that the execution agent should avoid.

## Rules

- Prefer criteria that can be checked against code, results, figures, tables, files, or report text.
- Use direct task data when available; only allow surrogate analyses when the data profile justifies them.
- Avoid generic criteria such as "analyze thoroughly" or "discuss results".
- Do not introduce requirements for unavailable labels, tools, or private data. If the user explicitly requires an unavailable input, retain that requirement as blocked and explain what is needed.
- Define satisfaction by correctness of the method and evidence, not agreement with a desired result. Negative or inconclusive findings can satisfy an appropriately framed research question.
- Keep the rubric compact enough to execute, without a fixed minimum number of criteria.