ccpe-system/workbench/analysis/review-agent-regression-202.../rubrics/review-agent-regression-rub...

2.6 KiB

Review Agent Regression Rubric

Purpose

Use this rubric to compare review outputs across prompt variants and model environments.

Do not score only surface fluency. The goal is to detect model-kernel preservation or degradation.

Scoring

Use 1-5:

1 = failed / absent
2 = weak
3 = acceptable
4 = strong
5 = excellent

Criteria

1. Model Fidelity

Does the output preserve the agent's core model rather than behaving like a generic reviewer?

2. Method Fidelity

Does it preserve the original method kernel?

For 巨人认知, check whether 思想考古 is present:

surface phenomenon -> tool/model layer -> hidden assumption -> philosophical bedrock -> value premise

For 认知显影, check whether the original显影-style review kernel is present rather than generic objection.

3. Deep-Structure Performance

Does the output identify deep structural problems instead of only local edits?

4. Hidden Assumption Detection

Does it identify assumptions that the source text relies on but does not state?

5. Philosophical Bedrock Excavation

Does it reach value premises, worldview, category framing, or governing metaphors when relevant?

6. Context Fit

Does it stay close to the article's actual material and intent?

7. Concept Overfitting Risk

Does it force favorite concepts, metaphors, or labels where they are not needed?

Reverse scoring note:

5 = low overfitting risk
1 = severe overfitting

8. Output Actionability

Are the suggestions concrete enough to revise the article?

9. Naming / Protocol Discipline

Does the output preserve required names, layers, and output protocol?

Examples:

巨人认知 must use GL0-GL4, not L0-L4.
Reconstructed labels must be marked as reconstructed.

10. Platform Stability

Does the prompt produce stable behavior in this model environment?

Signs of instability:

format collapse
generic reviewer drift
hallucinated source claims
loss of original tone
overlong boilerplate

Comparison Table Template

| Agent | Prompt Variant | Model Env | Article | Model Fidelity | Method Fidelity | Deep Structure | Hidden Assumptions | Bedrock | Context Fit | Low Overfit | Actionability | Naming Discipline | Stability | Notes |
| ----- | -------------- | --------- | ------- | -------------- | --------------- | -------------- | ------------------ | ------- | ----------- | ----------- | ------------- | ----------------- | --------- | ----- |

Final Judgment Labels

Use one:

clear winner
conditional winner
environment-specific winner
inconclusive
regression detected