Results, skill evolution, and interpretation

The strongest result is the performance of the trained policy without external skills. This separates final model capability from the additional inference-time effect of the selected skill, while leaving the contribution of individual training mechanisms unresolved.

Four paired test conditions

Condition	Strict reward	Success	Mean model steps	Coverage
Base, no skill	0.5422	51.75%	6.2125	400/400
Base + S₀	0.5663	54.50%	6.6125	400/400
Iter80, no skill	0.8501	84.25%	5.7525	400/400
Iter80 + Sₜ	0.8671	86.25%	6.0100	400/400

Four test conditions

Paired success difference	Percentage points	95% interval
Iter80 free − Base free	+32.50	[+27.50, +37.50]
Base + S₀ − Base free	+2.75	[−1.50, +7.25]
Iter80 + Sₜ − Iter80 free	+2.00	[−0.50, +4.75]

Intervals are task-paired percentile bootstrap estimates with 10,000 resamples. They describe test-task uncertainty conditional on this training run, not variability across training seeds. Both skill-only intervals span zero. No claim of a statistically established final-skill gain is made.

Bare Base → Iter80 changes 139 tasks from failure to success and nine from success to failure; 198 succeed in both conditions and 54 fail in both. Adding Sₜ to Iter80 fixes 19 tasks and regresses 11, net eight successes. Aggregate improvement does not mean every task improves.

Validation and the evolving bank

Checkpoint	Bare strict	Random-mask strict	Bare success	Selected chunks	Candidates
Iter20	0.765952	0.791661	74.75%	7	6
Iter40	0.813119	0.840690	80.00%	5	6
Iter60	0.860119	0.879774	84.75%	8	6
Iter80	0.869417	0.876024	86.00%	8	4

Validation and paired test effects

The randomized gate evaluates a pool of current and candidate versions, not the selected full bank. Its mean cannot substitute for final full-skill evaluation. The bank's initial ten chunks shrink to seven and five, then grow to eight. This reflects both retirement and new/revised procedures.

Round	All-wrong groups	Failed retry / analyzed cards	Eligible cards	Candidates
R0	104	87	61	6
R1	57	56	22	6
R2	54	53	20	6
R3	39	39	11	4

Skill evolution and failure funnel

Subtracting analyzed cards from all-wrong groups does not yield the number solved by full skills: technical retry errors also occur. Recorded internalization-deficit counts are 9, 1, 1, and 0. Technical faults, all-wrong policy groups, remaining diagnostic failures, and eligible skill proposals are distinct stages.

Contribution history

The heatmap aligns logical chunk identities; a dot denotes a retained winner, and an empty cell means not evaluated in that gate. Rewrites retain logical identity. Losing versions remain in the downloadable coefficient table. The home page's skill explorer includes the verbatim Chinese text of every selected bank.

Additional guidance also has costs

Reward components

Not every component improves when Sₜ is added. Iter80 + Sₜ uses 2,404 total model steps with 276 protocol errors, compared with 2,301 steps and 144 errors for Iter80 free. Error counts divided by total steps are 11.48% and 6.26%; these differ from averaging each task's error rate. API generation averages 191.02 versus 172.93 tokens per turn. More prompt tokens and measured latency also accompany the full skill, but latency is service- and hardware-dependent.

Ranking appendix and missing controls

An additional Base-model comparison at fixed k=4 gives top4 success 51.75%, bottom4 55.00%, and one random4 draw 54.25%. Only random seed 20260907 was completed; four planned random draws were not run. These factors came from a separate local Base gate, not the historical recovered S₀ coefficients. This does not establish held-out ranking validity or a reverse-ranking advantage.

The experiment has no plain-GRPO, fixed-S₀, no-mask, no-online-gate, q-sweep, or multiple-training-seed controls. It demonstrates substantial improvement over the initial model under this training system. It does not establish that co-evolution outperforms ordinary RL, or identify which component causes the observed gains. Repeated validation use, noisy main-effect coefficients, possible chunk interactions, and historical S₀ provenance remain material limitations.

Continue with behavior changes, optimization diagnostics, or the full original experiment report.

实验结果、技能演化与解释

最强的结果是训练后模型在无外部技能条件下的表现。这区分最终模型能力与技能的推理时附加作用，但没有分离训练各机制的因果贡献。

四个配对 test 条件

条件	Strict reward	成功率	平均模型步数	覆盖
Base，无技能	0.5422	51.75%	6.2125	400/400
Base + S₀	0.5663	54.50%	6.6125	400/400
Iter80，无技能	0.8501	84.25%	5.7525	400/400
Iter80 + Sₜ	0.8671	86.25%	6.0100	400/400

四组测试结果

配对成功率差值	百分点	95% 区间
Iter80 free − Base free	+32.50	[+27.50, +37.50]
Base + S₀ − Base free	+2.75	[−1.50, +7.25]
Iter80 + Sₜ − Iter80 free	+2.00	[−0.50, +4.75]

区间按任务配对做 10,000 次 percentile bootstrap，描述本次训练条件下的测试任务不确定性，不代表跨训练种子变异。两个仅加技能的区间均跨零，不宣称最终技能收益已获统计确证。

Bare Base → Iter80 有 139 个任务从失败变成功、九个从成功变失败；198 个均成功、54 个均失败。Iter80 加 Sₜ 修复 19 个任务、退化 11 个，净增八个成功。总体改善不意味着每个任务改善。

验证与技能库变化

Checkpoint	Bare strict	Random-mask strict	Bare success	选中 chunks	候选
Iter20	0.765952	0.791661	74.75%	7	6
Iter40	0.813119	0.840690	80.00%	5	6
Iter60	0.860119	0.879774	84.75%	8	6
Iter80	0.869417	0.876024	86.00%	8	4

验证与测试配对效果

随机 gate 评估 current 与候选版本池，而非选中后的完整技能，其均值不能替代 full-skill evaluation。技能库由初始十块收缩至七块、五块，再扩至八块，是淘汰与新增/改写共同作用的结果。

轮次	All-wrong groups	失败重试 / analyzed cards	Eligible cards	候选
R0	104	87	61	6
R1	57	56	22	6
R2	54	53	20	6
R3	39	39	11	4

技能演化与失败漏斗

All-wrong 数减 analyzed cards 数并不等于完整技能修复数，因为重试也会发生技术故障。实际 internalization-deficit 数为 9、1、1、0。技术故障、全错策略组、剩余诊断失败、符合提案条件的 cards 是不同阶段。

贡献历史

热图按 logical chunk identity 对齐，圆点表示保留的 winner，空白表示本轮未评估；改写保留 logical identity。失败版本保留在可下载系数表。首页技能浏览器可展开每版 selected bank 的原始中文正文。

附加指导也有成本

奖励分量

加入 Sₜ 后并非所有分量改善。Iter80 + Sₜ 共 2,404 模型步、276 协议错误；Iter80 free 为 2,301 步、144 错误。总错误 / 总步数为 11.48% 与 6.26%，不同于逐任务错误比例再平均。API 平均每步生成分别为 191.02 与 172.93 token。Full skill 也带来更多 prompt token 和观测时延，但时延取决于硬件与服务条件。

排序附录及缺失对照

Base 模型固定 k=4 的额外比较中，top4 成功率为 51.75%，bottom4 为 55.00%，一次 random4 为 54.25%。只完成了 random seed 20260907，其余四个计划随机组没有运行。排序来自另一个本地 Base gate，不能与历史恢复 S₀ 系数混用。结果未确立 held-out ranking 有效性，也不能证明反向排序更好。

本次没有 plain GRPO、固定 S₀、无 mask、无 online gate、q sweep 或多训练种子对照。结果展示该训练系统下模型相对初始 checkpoint 的明显提升，不能确立共同演化优于普通 RL，也不能确定收益由哪个组件产生。验证集复用、主效应系数噪声、chunk 交互及历史 S₀ 来源仍是实质限制。

继续阅读行为变化、优化诊断或完整原始实验报告。