8.5 KiB
8.5 KiB
CCPE-Lite Migration Strategy Summary
created: 2026-06-03
suite: workbench/analysis/review-agent-regression-2026-06-02
scope: Giant Cognition + Cognitive Imaging regression synthesis
status: final summary before canonical prompt changes
policy_decision: minimal-kernel first, refined-lite later
1. 背景
本轮回归测试比较了两个成熟 CCPE 2.0 单 Agent：
Giant Cognition / 巨人认知
Cognitive Imaging / 认知显影
每个 Agent 均比较三种 prompt 形态：
original-ccpe-2
ccpe-system-lite
original-kernel-minimal-lite
每种 prompt 在两个模型环境中测试：
Gemini
ChatGPT/Codex
每个环境使用四篇文章，覆盖：
strong metaphor
business analysis
logical argument
value philosophy
2. 两个个案的迁移历史差异
2.1 认知显影
认知显影 Lite 的形成过程实质经历了多轮 A/B：
1. 初版迁移。
2. 发现原 CCPE System 对 CCPE 2.0 工作内核保留不足。
3. 先修复 CCPE System，再重新生成 Lite。
4. 新 Lite 在 Gemini 侧已超过原版。
5. 继续补充纪律规则并再次调整。
结论：
认知显影是 refined-lite 成功样本。
当前 canonical Lite 应保留。
2.2 巨人认知
巨人认知 Lite 的形成过程较短：
1. 旧 CCPE 2.0 原版。
2. 第二轮生成当前 Lite。
3. 发现当前 Lite 稳定、可用，但仍不如原版工作内核。
4. 暂停精修，进入回归测试。
结论：
巨人认知应回到 original-kernel-minimal-lite 作为下一步生产候选。
当前 Lite 不删除，保留为 refinement candidate。
3. 回归测试总判断
3.1 直接结论
Giant Cognition:
clear winner: original-kernel-minimal-lite
current ccpe-system-lite: useful but has GL3 template-overfitting risk
Cognitive Imaging:
production winner: ccpe-system-lite
kernel-fidelity winner: original-kernel-minimal-lite
current ccpe-system-lite: refined success, keep canonical
3.2 策略结论
本轮测试不支持“所有成熟 Agent 都应直接重写成 full Lite”。
更可靠的迁移策略是：
minimal-kernel first,
refined-lite later.
也就是：
1. 先从原 CCPE 2.0 prompt 生成 original-kernel-minimal-lite。
2. 保留原始工作内核、输出行为、关键术语和方法压力。
3. 只添加必要的 Lite metadata、平台边界、CoT 修复、输出验证纪律。
4. 做低成本回归测试。
5. 通过则作为短期生产版本。
6. 只有高价值、高频、长期使用的 Agent，才进入多轮 refined-lite 精修。
4. 为什么 minimal-kernel first 是更高 ROI 路线
4.1 人力成本
认知显影的精修路线证明 refined-lite 可以超过原版，但人工成本高：
multiple A/B rounds
manual Gemini operation
side-by-side result inspection
prompt rule diagnosis
CCPE System rule adjustment
retest
这类流程适合少数核心模型，不适合短期批量迁移。
4.2 迁移风险
full Lite rewrite 的主要风险不是格式失败，而是工作内核漂移：
- 名称保留了，但方法压力变了。
- 输出更完整了，但原始批判力度被驯化。
- 新模板更稳定了，但可能产生概念增殖。
- 结构更规范了，但旧 Agent 的 productive strangeness 被削弱。
巨人认知显示出典型风险：
GL3 隐喻器官规则提高了 actionability，
但也可能从思想考古转向机制补件生成。
4.3 快速迁移收益
original-kernel-minimal-lite 的优势：
- 迁移速度快。
- 保留原版 working kernel。
- 人工评测负担低。
- 可立即作为回归参照。
- 比原版更安全：修复 CoT、平台边界、输出纪律。
- 比 full rewrite 更少引入模板漂移。
5. 新迁移路线
5.1 Fast Migration Lane
适用对象：
成熟 CCPE 2.0 单 Agent
Web / GPT / Gemini / Claude 风格专家 prompt
用户已经证明其有效
当前目标是短期迁移与批量升级
流程：
1. Audit original artifact.
2. Classify as CCPE-Lite first unless scenario requires heavier layers.
3. Extract and preserve original working kernel.
4. Generate original-kernel-minimal-lite.
5. Add only minimal migration repairs:
- platform boundary
- hidden reasoning disclosure repair
- output validation
- source/retrieval policy if needed
- version/status metadata
6. Run a small regression sample.
7. Promote as temporary production Lite if it beats or matches original.
8. Preserve original and current rewrite as regression references.
Stop condition：
If original-kernel-minimal-lite performs acceptably,
do not continue full refined-lite optimization during the batch migration phase.
5.2 Refinement Lane
适用对象：
high-value agents
high-frequency use
model drift detected
kernel preserved but production stability insufficient
user explicitly wants deeper optimization
流程：
1. Start from original-kernel-minimal-lite, not from scratch.
2. Identify concrete regression or improvement target.
3. Add small discipline rules, not wholesale rewrite.
4. Run focused A/B tests.
5. Compare:
- original-ccpe-2
- original-kernel-minimal-lite
- refined-lite candidate
6. Promote only if refined-lite improves production stability without losing kernel force.
Stop condition：
If refined-lite loses kernel force, pause and keep original-kernel-minimal-lite.
6. Agent-Level Decisions
6.1 巨人认知
Recommended action:
Change production candidate direction to original-kernel-minimal-lite.
Do not delete current ccpe-system-lite.
Keep current Lite as paused refinement candidate.
Do not create Skill / Agent Spec / Runtime.
Do not split Giant Cognition into sub-model cards.
Policy note:
Giant Cognition is a strong example of why full Lite rewrite should not be the default first move.
6.2 认知显影
Recommended action:
Keep current canonical ccpe-system-lite.
Use original-kernel-minimal-lite as regression reference.
Do not roll back.
Do not create Skill / Agent Spec / Runtime.
Policy note:
Cognitive Imaging is a strong example of when refined-lite is worth the cost.
7. Policy Updates Required
Update CCPE System documentation with:
1. Minimal-Kernel First Rule.
2. Fast Migration Lane vs Refinement Lane.
3. Kernel Force vs Production Stability split scoring.
4. A/B Budget Rule.
5. Regression Reference Preservation Rule.
6. No automatic full Lite rewrite rule.
7. No automatic Agent / Skill / Runtime expansion rule.
Recommended files:
ccpe-protocol/ccpe-migration-policy.md
ccpe-protocol/ccpe-quality-rubric.md
.codex/skills/ccpe-forge/SKILL.md
.codex/skills/ccpe-forge/references/refactor-mode.md
Optional later files:
.codex/skills/ccpe-forge/templates/ccpe-lite.prompt.md
ccpe-protocol/ccpe-layer-spec.md
8. Updated Evaluation Criteria
For mature Lite migrations, evaluate both:
Kernel Force:
Does it preserve the original method pressure, voice, conceptual edge,
output behavior, and distinctive cognitive operation?
Production Stability:
Does it run reliably across target model environments,
with clear boundaries, safe reasoning policy, and usable output?
Do not treat a higher production-stability score as sufficient if kernel force collapses.
Do not treat a stronger kernel-force score as sufficient if platform safety is broken.
9. Canonical Promotion Rule
Before replacing a canonical Lite prompt:
1. Keep original-ccpe-2 as regression reference.
2. Keep original-kernel-minimal-lite as kernel reference.
3. Compare full Lite rewrite only when needed.
4. Record the promotion reason.
5. If model card/index already exist, update usage mapping only when canonical production path changes.
6. Do not update model definition unless the cognitive model itself changed.
10. Final Recommendation
Adopt this migration policy:
Default:
original-kernel-minimal-lite first
Promote refined-lite only when:
high-value agent
clear improvement target
successful A/B evidence
no kernel-force regression
Preserve:
original artifact
minimal-kernel candidate
refined candidate if produced
Avoid:
full rewrite as default
four-layer expansion as default
repeated A/B for every old Agent
This gives the CCPE System a practical migration path:
fast enough for batch work,
faithful enough to preserve old CCPE 2.0 power,
structured enough to support future refined Lite upgrades.