把最终补丁反馈，分配到编辑轮次

paper-rl/paper/sections/5_evaluation.tex

\paragraph{Policy effectiveness: pending.}

The base-model development pilot generated eight trajectories on each of the original 40 tasks. Of these 320 planned trajectories, 244 met the training query rules and all 244 intervention trials returned usable reports. Here, usability follows the training adapter's fixed-list utility contract. The exclusions were 41 format-error trajectories, 34 without an eligible block, and one without an original utility after the agent deleted installation files. This last trajectory is retained as a model failure; it is not resampled.

The contrast is nonzero for 92 queries (37.7\%), including 17 from failed original trajectories: 12 positive and five negative, spanning 14 tasks and six repositories. These are 17 of the 111 queried failed trajectories, within 150 original trajectories with a known failed verdict. The twelve positive contrasts split evenly between a positive fail-to-pass (F2P) component and the regression indicator alone (\cref{app:signal-origin}). These observations establish information availability on failures, rather than a benefit from learning with that information. \Cref{tab:pilot} preserves the original pilot denominator and separately exposes the 39-task stratum admitted to training.

The cap activates for 83 of the 92 nonzero contrasts (90.2\%), placing most observed nonzero cases in the sign-and-support regime of \cref{eq:coefficients}. Among these 92 cases, the median selected-token fraction $m/L$ is 2.10\% and median shrinkage $s$ is 0.181; \cref{app:geometry} gives their distributions. Before format-error and original-utility query exclusions, 282 trajectories have a constructed intervention artifact; of these, 47 become an empty patch. The median fraction of final changed paths touched by the selected block is 0.5. One of the 92 nonzero cases instead selects paths absent from the final patch: the block repairs earlier agent damage. This illustrates why last-touch eligibility is defined relative to the pre-block tree rather than only to the final net diff. These structural measurements do not certify semantic attribution accuracy.

All 120 prescribed cross-machine factual replays match their originals at the level of reported test labels. The initial 244 intervention trials take 237.3 seconds on average, with a 416.9-second empirical 95th percentile. A separate deployment repeat of the same 244 artifacts on the designated verifier host, at concurrency 24 with images retained between trials, yields identical per-test reports for all cases and takes 115.1 seconds on average, with a 189.9-second 95th percentile. Service time includes setup, verification, and cleanup, and excludes semaphore waiting. Host placement, concurrency, and cleanup differ together; the comparison does not isolate a cache effect. Production queueing and total training cost remain unmeasured.

All 120 prescribed cross-machine factual replays match their originals at the level of reported test labels. The initial 244 intervention trials take 237.3 seconds on average, with a 416.9-second empirical 95th percentile. A separate deployment repeat of the same 244 artifacts on the designated verifier host, at concurrency 24 with images retained between trials, yields identical per-test reports for all cases and takes 115.1 seconds on average, with a 189.9-second 95th percentile. Service time includes setup, verification, and cleanup, and excludes semaphore waiting. Host placement, concurrency, and cleanup differ together; the comparison does not isolate a cache effect. Production queueing and total training cost remain unmeasured.

SHA-256 90a6a8574b27466b6e876253200ff9e7762f5747fc8c62cfe4cb0e71484bc811

paper-rl/PAPER_PLAN.md

| M/B在同一轨迹上具有同带符号/L1原始辅助系数质量 | 公式代数、CPU实际loss接入检查 | 技术支持；不含梯度范数/裁剪等价 | §3 |

| 每条准入轨迹至多一个额外verifier query、零策略续写 | 已实现查询路径；实际端到端成本仍在测 | 查询界受支持；可接受成本及相对成本尚未证实 | §3/§5 |

主终点80。M−B是主要比较；M−O、M−E次之。独立seed和task配对统计放附录，主表不写CI，但不删实验范围/局限。给出真实成本字段，待实测后填值。

E的辅助项为`C_E(t)=clip(lambda*A*L/m,-c,c)*I_e(t)`，加到原终局组优势A上；不访问D，m=0/选择不可用时为0。

主终点80。M−B是主要比较；M−O、M−E次之。独立seed和task配对统计放附录，主表不写CI，但不删实验范围/局限。给出真实成本字段，待实测后填值。

SHA-256 f345575600d0c77bdd22b735a141a30a97a0513e3d65dbe0061cc37063e8cbe1

paper-rl/paper/sections/6_limitations.tex

Final-artifact removal conditions on the other completed files. Later edits on other paths may depend on the removed block, even when no later turn overwrites its paths. The resulting contrast is not an unbiased action value, a Shapley allocation, or the outcome of a naturally continued trajectory. Last-touch and test-path eligibility also restrict which edits can receive this feedback. Snapshot attribution is temporal: background writes captured after a turn can be assigned to that turn, and restoring a last-touch block can reintroduce damage from an earlier block.

Equal signed and absolute coefficient mass does not isolate semantic placement from every optimization effect. Localization changes concentration; adding the terminal advantage can change signs and DPPO masking. O/B/M/E evaluate a specific practical allocation rule. Claims about attribution accuracy require a suitably matched relocation control, and claims that removal information is necessary require controls using simpler factual test feedback. Those conclusions are outside the present comparison.

The planned study uses one base model and a limited RL task pool. It does not establish model-family-wide effectiveness, and base-model pretraining overlap is not measured here. An earlier adapter admitted zeros without grading reports; the affected run was excluded from the comparison. Successor training requires complete grading evidence, but this can also withhold negative feedback for model-broken installation; retry and masking rates must be reported by arm (\cref{app:grading_recovery}). The official evaluation's expected-test lists can omit behavior relevant to an issue; retaining its labels and denominator preserves the benchmark definition rather than repairing that limitation. Matching reports establish reproducibility of the recorded score. Missing test observations can be classified as failures, and reused descriptions can overwrite earlier test statuses. Final evaluation therefore retains raw observations and separates observed failures from outcomes requiring attribution.

SHA-256 c0d214dcd0f06f9efd49d21e934252f7c53c7e32cdfd1c949ca2e73967d72adc

paper-rl/results/O42_retry1_external_termination_20260928.json

"observed_at": "2026-09-28T15:35:17.342206+08:00",

"quality_observed_at": "2026-09-28T15:34:35.821420+08:00",

"host_observed_at": "2026-09-28T15:34:54.658129+08:00",

"training_call_completed": 34,

"termination_class": "external_or_host_termination_unresolved",

"automatic_restart": false,

"oom_killed": false,

"next_action": "Do not resume until a separately reviewed recovery plan binds checkpoint30 and termination evidence"

SHA-256 5b9ce35674fd8114d171fb00d8b0336ffdd278b5f6e6bc9161e12d06244577f9

paper-rl/results/common_profile_runtime_coverage_final_20260930.json

"scope": "All500 common-profile tasks have been exercised across retained waves; unresolved and mismatches remain explicit. No model-effect claim.",

"controls": 1000,

"controls": 1000,

"raw_pass": 995,

"strict_pass": 862,

"report_binary": 887,

"unresolved": 113,

"partial_zeros": 20,

"control_expectation_mismatches": 5,

"control_expectation_mismatches": 5,

"report_binary": "Binary report diagnosis includes unexpected control scores; not a count of passed controls or formal model slots"

SHA-256 45ef82e6cc1980428d679e28ad6f29ab4ec72fa843521448f25a5172a2385a48

paper-rl/PAPER_ACCEPTANCE_CONTRACT.md

status: accepted

assurance: draft

9. 数据表明确实际训练149、2384别名、dev39；官方500及holdout100分母不变；原150映射与未用于后继训练的351不混用。

10. 80步、三种子计划、E42、n=2/n=16及当前执行状态明确区分；成功率按每条rollout求均值，绝不取best-of-n，计划的holdout100 n=16明确包含B；不声称未运行的重复实验已经完成。

11. 所有主结果/成本空缺使用Pending或破折号并解释未测量，不使用模拟数字、臆测曲线或旧无效O/B/M成绩。全文任何位置都不能给出尚未测得的本方法胜负/泛化提升或更低总成本结论。

这份约定针对当前草稿，不把“方法有效”作为可凭空满足的写作要求。最终提交阶段必须在真实终评后重新建立结果版本的约定；不能把本草稿验收等同投稿就绪。

SHA-256 3d9ae3c2667e2c4a0bc56dbf24e83b279b8267c122956181e230d585242208cd

paper-rl/STATUS_20260928.md

**2026-10-01汇总更正：** `common_profile_runtime_coverage_final_20260930.json` 已按23个保留cohort重算：500题/1000控制，raw 995，strict 862，report-binary 887，未决113，部分观测20，参考/环境差异5。原文件中的999/779/798/101/18为聚合错误，已由提交 `fdd83033` 修正；这些是环境验证覆盖数字，不是模型效果。

- **O42-retry1在第34次调用后随父容器外部终止。** 质量观察器最后看到34/80，训练父 `skyrl-rl` 曾 exit137、Docker `OOMKilled=false`；driver控制读取进入 review。step20与step30 HF导出/参数审计已完成；step30为427张量、329个变化张量、823171568个变化元素、无非有限值。`skyrl-rl` 后来被外部重新启动，但没有形成O42恢复证据，不把容器重启当自动恢复，也不把存储变化写成效果。收据 `results/O42_retry1_external_termination_20260928.json`、`O42_retry1_step30_parameter_audit_20260928.json`。

旧O40仍作废于主比较，所有旧分数/存点保留。新runtime、数据、launch和release都保持冻结。gpu08 GPU6/7及GPU0原小服务不动；监督不自动重启。

- 后继入口F0–F3/两轮复核已完成，真实O80、终点参数核验及CF适用性release仍欠；B/M/E未启动。共同候选仍87题/174控制、16未决、413题未新跑，不能用新诊断改变分母。论文仍15页、效果未测，当前没有等待自动重试的CPU控制任务。

下一步：核验首个原生存点的参数；实现共同评测的可信参考/完整成员映射/准入规则；沿已完成后继入口准备真实O80与CF证据绑定的release。旧CPU控制/原始未决和论文数字均未修改。

- 后继入口的CPU核验明确：当前恢复make_plan只收O42，前驱校验只认旧O40停训且拒绝retry1名称。要另建支持固定矩阵和正常80调用前驱的新入口；9份配置计划不是release。见 `results/successor_launch_boundary_20260928.json`。

剩余9个预定训练条件的比较计划已经生成：B/M各3种子、O另2种子、E42。全部通过实际SkyRL配置解析，与活跃O仅臂、种子和输出路径不同。数据、原base、runtime、采样、优化器及14卡布局保持相同；没有启动后继或申请新GPU。入口 `staged/credit-after-recovery-20260928/`，计划 `rl/runs/rl-paper-20260927/credit-after-recovery-plans-v2/`。v1清单相对路径在runtime cwd下无法定位，原失败保留，v2改绝对路径。计划不是release；后继仍要核对新O80实际结束、清理、源、CF与资源证据。

SHA-256 824026dba6207f23184e0c9dd09e37f9e8ea21b534a21ce4b6a878a03db40b2f

paper-rl/results/grading_recovery_paper_update_20260928.json

"scope": "Limitations and appendix recovery/resource-control changes only, no efficacy claim or full160-claim reaudit.",

"review_independence": "same-family",

"acceptance_status": "provisional",

"pdf_pages": 15,

"main_body_end_page": 9,

"submission_ready": false,

"policy_effectiveness_measured": false,

SHA-256 7756cd4c431ab0056467e333e00fcdd6feafe4e39f6a71e6b0fbe4c31fc05b8e