10 KiB
10 KiB
巨人认知回归对比报告
agent_family: giant-cognition
test_date: 2026-06-02
scope: Gemini 侧 12 个结果 + ChatGPT/Codex 侧 12 个结果
status: regression-comparison
promotion_decision: none
1. 结论摘要
本轮巨人认知回归测试的主结论：
clear winner: original-kernel-minimal-lite
conditional winner: ccpe-system-lite for actionability and protocol discipline
regression detected: full Lite rewrite has mild GL3 template-overfitting risk
original-kernel-minimal-lite 是当前最稳的生产候选方向。它保留原版的思想考古、GL0-GL4、意图/反思双循环和批判性语气，同时修复了平台边界与隐藏推理披露问题。它在 Gemini 和 ChatGPT/Codex 两侧都更接近原始评审内核。
当前 ccpe-system-lite 没有失败。它的协议纪律、结构完整性、行动建议和 GL3 隐喻器官生成都很强。但它也暴露出一个迁移风险：模型有时会把“隐喻器官”变成模板驱动的补件，出现概念增殖，而不是始终执行原版的深层假设考古。
原版 original-ccpe-2 仍然有效，尤其在 Gemini 环境里表现稳定。但它缺少最小 Lite 包装里的平台边界、隐藏推理修复和输出验证纪律，不适合作为跨平台最终形态直接使用。
2. 测试材料
Prompt variants:
original-ccpe-2
ccpe-system-lite
original-kernel-minimal-lite
Article corpus:
article-01-strong-metaphor: 2026-05-07-reconstruction-of-the-aesthetic-contract.md
article-02-business-analysis: 2025-12-28-the-roaring-above.md
article-03-logical-argument: 2026-01-07-anchoring-the-void.md
article-04-value-philosophy: 2026-05-09-the-dawn-of-silicon-based-life.md
Model environments:
Gemini: user-provided manual results
ChatGPT/Codex-side: 12 fresh background Codex threads, one condition per thread
3. Aggregate Scoring
Scores are qualitative 1-5 rubric estimates based on the 24 saved outputs. They are not statistical measurements.
| Prompt Variant | Model Env | Model Fidelity | Method Fidelity | Deep Structure | Hidden Assumptions | Bedrock | Context Fit | Low Overfit | Actionability | Naming Discipline | Stability | Judgment |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| original-ccpe-2 | Gemini | 4.5 | 4.3 | 4.2 | 4.0 | 4.0 | 4.4 | 4.2 | 4.0 | 4.8 | 4.5 | stable baseline |
| ccpe-system-lite | Gemini | 4.4 | 4.0 | 4.1 | 3.8 | 3.8 | 4.2 | 3.5 | 4.5 | 5.0 | 4.4 | actionability winner, overfit risk |
| original-kernel-minimal-lite | Gemini | 4.7 | 4.6 | 4.6 | 4.5 | 4.6 | 4.4 | 4.0 | 4.2 | 5.0 | 4.6 | environment winner |
| original-ccpe-2 | ChatGPT/Codex | 4.4 | 4.2 | 4.3 | 4.1 | 4.0 | 4.5 | 4.5 | 4.0 | 4.8 | 4.3 | stable baseline |
| ccpe-system-lite | ChatGPT/Codex | 4.2 | 4.0 | 4.1 | 3.9 | 3.8 | 4.4 | 3.2 | 4.7 | 5.0 | 4.5 | actionability winner, overfit risk |
| original-kernel-minimal-lite | ChatGPT/Codex | 4.7 | 4.6 | 4.7 | 4.6 | 4.6 | 4.5 | 4.1 | 4.3 | 5.0 | 4.6 | environment winner |
4. Per-Article Pattern
| Article | Gemini Best | ChatGPT/Codex Best | Notes |
|---|---|---|---|
| article-01-strong-metaphor | original-kernel-minimal-lite | original-kernel-minimal-lite | Kernel version best preserves the aesthetic-contract bedrock while avoiding pure metaphor optimization. Current Lite is more actionable but adds stronger template artifacts. |
| article-02-business-analysis | original-kernel-minimal-lite | original-kernel-minimal-lite | Kernel version digs into state/individual time-scale conflict and governance-contract assumptions. Current Lite gives useful “检修井/泄洪区” organs but risks over-mechanizing the water metaphor. |
| article-03-logical-argument | original-kernel-minimal-lite | original-kernel-minimal-lite | Kernel version best excavates constructivism, truth criteria, and observer self-position. Current Lite is operationally neat but more device-oriented. |
| article-04-value-philosophy | original-kernel-minimal-lite | original-kernel-minimal-lite | Kernel version best surfaces teleology, Conatus, non-ergodic ethics, and value boundary issues. Current Lite improves mechanism closure but tends toward engineered metaphor patches. |
5. Variant Diagnosis
5.1 original-ccpe-2
Strengths:
- Preserves original voice and report protocol.
- Strong Gemini compatibility.
- Produces real GL3 assessments rather than generic article reviews.
- Low concept-overfitting risk because it does not add many new wrapper rules.
Weaknesses:
- No explicit platform/retrieval boundary.
- Original hidden-reasoning language is not repaired.
- Output validation discipline is weaker.
- Actionability is sometimes less structured than the migrated variants.
Regression reading:
No core failure. The original prompt remains a strong regression reference, not a safe final cross-platform artifact.
5.2 ccpe-system-lite
Strengths:
- Best naming/protocol discipline.
- Always preserves GL0-GL4 rather than drifting into L0-L4.
- Strongest actionability and next-step structure.
- Handles ChatGPT/Codex environment well.
- GL3 “隐喻器官” rule creates useful repair suggestions in several cases.
Weaknesses:
- Mild template-overfitting risk.
- GL3 sometimes becomes “generate a new mechanism metaphor” rather than “excavate hidden assumptions”.
- More likely to add new labels such as 审美阀门, 检修井, 校准仪, 痛觉阈门 even when the original article already has enough internal model material.
- Longer outputs may feel more engineered than the original review kernel.
Regression reading:
No catastrophic regression, but a real style/method drift exists. The full Lite rewrite improves usability but partially shifts the kernel from 思想考古 toward mechanism-completion.
5.3 original-kernel-minimal-lite
Strengths:
- Best balance of original kernel fidelity and cross-platform safety.
- Preserves original report style and GL0-GL4 scan.
- Explicitly repairs hidden chain-of-thought disclosure.
- Explicitly marks platform/retrieval boundaries.
- Strongest hidden-assumption and philosophical-bedrock performance across both environments.
- Lower overfitting risk than current Lite while still benefiting from minimal validation rules.
Weaknesses:
- Less operationally standardized than current Lite.
- Does not always generate the extra “隐喻器官” field that current Lite makes explicit.
- Some ChatGPT/Codex outputs still introduce new model names, e.g. 非遍历伦理学, which are useful but should be marked as analysis/reconstruction if promoted into prompt rules.
Regression reading:
This variant best supports the handoff hypothesis: mature review agents should first preserve original CCPE 2.0 kernel, then add minimal Lite metadata, safety, platform boundary, reasoning-disclosure repair, and validation discipline.
6. Gemini vs ChatGPT/Codex
Gemini:
- Shorter, sharper, closer to the old Web-style expert-agent feel.
- Very stable on original and original-kernel variants.
- Current Lite follows protocol well, but the added “隐喻器官” behavior becomes more visible.
ChatGPT/Codex:
- Longer, more explicit, more systematic.
- Current Lite produces the most complete output, but also the clearest template pressure.
- original-kernel-minimal-lite keeps enough structure without over-expanding.
- Fresh-thread execution worked: 12 result files were created under results/chatgpt/.
Environment-specific conclusion:
Gemini can safely run original-kernel-minimal-lite as the default production candidate.
ChatGPT/Codex should also prefer original-kernel-minimal-lite unless the user specifically wants stronger action-planning and explicit “隐喻器官” suggestions.
7. Answer to Handoff Questions for Giant Cognition
- Which prompt variant performs best?
original-kernel-minimal-lite.
- Which prompt variant transfers best to ChatGPT/Codex?
original-kernel-minimal-lite. Current Lite is stable, but has more template-overfitting risk.
- Which prompt variant performs best on Gemini?
original-kernel-minimal-lite, with original-ccpe-2 as a strong baseline.
- Does original-kernel-minimal-lite outperform full Lite rewrite?
Yes, for model-kernel fidelity, hidden-assumption detection, and philosophical-bedrock excavation.
No, if the only criterion is actionability and explicit repair formatting.
- Is CCPE System migration preserving the old CCPE 2.0 review kernel?
Partially. The current Lite preserves names and structure but introduces method drift. The minimal-kernel route preserves the kernel better.
- Should mature review agents use separate production prompts per model environment?
Not yet necessary for Giant Cognition. One original-kernel-minimal-lite prompt appears portable enough across Gemini and ChatGPT/Codex. Keep environment-specific notes only if later tests show model-specific failures.
- What should change in CCPE Forge migration rules?
For mature single-agent review prompts:
1. First create an original-kernel-minimal-lite variant.
2. Run regression against original and full Lite rewrite.
3. Treat current Lite template additions as optional hardening, not default rewrite structure.
4. Require GL3 思想考古 to preserve hidden-assumption / philosophical-bedrock excavation before adding metaphor-mechanism repair rules.
5. Mark newly generated labels or organs as analysis/reconstruction unless they become user-confirmed canonical terms.
8. Recommendation
Do not replace agents/lite/giant-cognition.prompt.md yet.
Recommended next action:
1. Keep all three prompt variants in the regression directory.
2. Use original-kernel-minimal-lite as the candidate direction for any next Giant Cognition prompt revision.
3. Wait until Cognitive Imaging regression is complete.
4. Then write the combined report and decide whether CCPE Forge should formally adopt “original-kernel-minimal-lite first” as a migration rule.
Human decision required:
Whether to revise the canonical Giant Cognition Lite after both agent-family reports are complete.