ParetoGTICal 工作台ParetoGTICal workbench

组会 · 2026-10-21（最新一天，完整）Group meeting · 2026-10-21 (latest day, in full)

四个问题，一屏作答Four questions, answered on one screen

现在在哪？

Where does it stand?

P001 已关闭（D006），答案否定：等保留率下率排名在任何主集上都不优于 2.filter。此后项目转向误差预算判据（C001）：多观测时一致性比信噪比重要。2026-10-18 跑完了第一次保留集判决：在 119 个从未参与开发的观测上，83 个冻结候选中 79 个相对归档 2.filter 是 Confirmed —— 但 2.filter 是弱参照，对复现的标准光变曲线滤波器只剩 11/83，且都靠更短的保留时间。2026-10-20 PI 定调：交付物是"每个数据集测试的帕累托族"（D096/D097）—— 没有单一赢家，没有对参照的通过/失败；一个策略只要没被别的策略在所有轴上完全超过就保留。轴包括观测间一致性、单观测 \(\sigma_R\)、保留时间、观测覆盖率。R55 已跑：20 个测试 \(\times\) 103 个策略，2.filter 在 19/20 个测试是家族成员。截至 2026-10-21：观测间轴已改成 between_all（被丢弃的观测按不过滤水平计回，D106），"丢观测换好分"的漏洞已堵；光子拆分证明保留集颜色只是"噪声保留"（D112/D118）；同天区谱形状验证（R57）因在 \(n=8\) 上无法认证 10% 裕度而改为描述性精度研究（D118/D122，待 PI 定）；残余软质子在观测之内的示踪剂斜率 \(\approx0\)，观测间的大斜率是混杂 —— 而混杂已被 R82 查明是指向偏离造成的 Simpson 成分效应（D125），所以"残余软质子驱动过滤后颜色散度"在发展集上已关闭；掩码 + 去偏份额 + 同类后只剩 M31 核心 MOS1 E_k2.5 一个过滤 cell 通过 Holm，且因可能是点源变化或 PSF 泄漏而标为 Provisional（D126/D127）。下一个测试（P007）用变体 C 的 54 个观测，冻结前还差 R63 颜色轴与 R57 协议定稿。

P001 is closed (D006), negative: at equal retention rate-rank does not beat 2.filter on any primary set. The project then turned to the error-budget criterion (C001): over several observations, consistency matters more than S/N. On 2026-10-18 the first held-out verdict ran: on 119 observations that took no part in development, 79 of 83 frozen candidates are Confirmed relative to the archived 2.filter — but 2.filter is a weak reference, so against the reproduced standard light-curve filter only 11 of 83 survive, and all of them by keeping less time. On 2026-10-20 the PI set the headline: the deliverable is "the Pareto family per dataset test" (D096/D097) — no single winner, no pass/fail against a reference; a policy is kept when no other policy beats it on every axis at once. The axes include between-observation consistency, per-observation \(\sigma_R\), kept time and observation coverage. R55 has run: 20 tests \(\times\) 103 policies, 2.filter a family member in 19/20. As of 2026-10-21: the between-observation axis is now between_all (a dropped observation re-enters at its unfiltered level, D106), closing the drop-for-score exploit; the photon split showed held-out colours are "noise-held-out" only (D112/D118); the same-sky spectral-shape validation (R57) is recast as a descriptive precision study because a 10% margin cannot be certified at \(n=8\) (D118/D122, pending the PI); residual soft protons have within-observation tracer slopes \(\approx0\), and the large between-observation slopes are confounding — and R82 has now shown the confounding is a Simpson composition effect from the pointing offset (D125), so "residual soft protons drive the filtered colour scatter" is closed on the development sets; after masking, a debiased share and homogeneous classes only M31 core MOS1 E_k2.5 survives Holm, and it stays Provisional because it may be source variability or PSF leakage (D126/D127). The next test (P007) uses the 54 observations of variant C; before the freeze it still needs the R63 colour axes and the R57 protocol finalised.

还有什么没解决？

What is unsolved?

① 观测间一致性轴已修好（D106 关闭）：轴改为 between_all，被丢弃的观测按不过滤水平与曝光计回，"丢观测换好分"的漏洞堵住；红队 RT-106 确认它仍奖励"丢真异常观测"，但税不超过被丢观测自己的散度。② 同天区谱形状验证（R57）改了性质：\(n=8\) 上无法认证 10% 的 H/S 裕度（S-P30：上界 \(\geq14.38\) log-pp 对 9.53），改为描述性精度研究（R79 协议已出，待 PI 定）；颜色散度在数据集之间差异很大，且尚未与粒子本底的颜色漂移分开。③ 残余软质子：观测之内的示踪剂斜率 \(\approx0\)（R74/R78），观测间的大斜率是混杂；QPB 颜色扣减与逐 CCD B 态剔除基本不降颜色散度（R68/R73）—— R82 去找那个混杂因子。④ 帕累托成员数随阈值 z 的增长是保留时间产物（R76/D122）：去掉 T 只剩 z = 1 与 z = 2 入族；文献建议在固定保留时间下比污染度并预先声明优值（R80）。⑤ Al-K 线门几乎从未触发（每格 0–3 个观测）；谱验证（C2）还没做；P002 F1 复制失败的原因未测。

① The between-observation axis is fixed (closed by D106): the axis is now between_all, a dropped observation re-enters at its unfiltered level and exposure, and the drop-for-score exploit is closed; the red team RT-106 upheld it with the reading that it still rewards dropping a genuine outlier, but the tax never exceeds the dropped observation's own scatter. ② The same-sky spectral-shape validation (R57) changed character: a 10% H/S margin cannot be certified at \(n=8\) (S-P30: the upper bound is \(\geq14.38\) log-pp against 9.53), so it is recast as a descriptive precision study (the R79 protocol is drafted, pending the PI); colour scatter varies strongly between data sets and is not yet separated from particle-background colour drift. ③ Residual soft protons: within-observation tracer slopes are \(\approx0\) (R74/R78) and the large between-observation slopes are confounding; QPB colour subtraction and per-CCD B-state removal barely reduce colour scatter (R68/R73) — R82 hunts the confounder. ④ Pareto membership growth with the threshold z is a kept-time artefact (R76/D122): without T only z = 1 and z = 2 remain members; the literature suggests comparing contamination at fixed kept time with a pre-declared figure of merit (R80). ⑤ The Al-K line gate almost never fires (0-3 observations per cell); spectral validation (C2) is still to come; the causes of the P002 F1 replication failure are untested.

接下来做什么？

What do I do next?

按记录：(a) P007 冻结与运行（主线）—— 数据集已定变体 C（54 观测，D103）；冻结前把 R63 颜色轴并入家族、定稿 R57 协议（默认：描述性精度研究，待 PI 定），判决形式 = 每个数据集测试的家族 + 自举成员频率。(b) R63（颜色轴进家族）、R80（固定保留时间的阈值比较）、R82（找观测间混杂因子）进行中。(c) 谱验证 C2 与 P002 F2 等 PI 派发。步骤格式见"如何接手"页 §5。

Per the records: (a) freeze and run P007 (the main line) — the data set is variant C (54 observations, D103); before the freeze, fold the R63 colour axes into the family and finalise the R57 protocol (default: a descriptive precision study, pending the PI); the verdict takes the form "the family per dataset test + bootstrap membership frequency". (b) R63 (the colour axes into the family), R80 (the threshold comparison at fixed kept time) and R82 (hunting the between-observation confounder) are running. (c) Spectral validation C2 and P002 F2 await the PI's dispatch. Step formats on the "By hand" page §5.

目标表（PI 工作单，D002,2026-10-13）Objectives (the PI's work order, D002, 2026-10-13)

开放问题表Open problems

未来 1–4 周计划（按记录 D061–D123；每步给目标、输入、方法、成败检验、时间估计与先读文件）Plan for the next 1–4 weeks (per the records D061–D123; each step with goal, inputs, method, success/failure test, time estimate and the file to read first)

1. 已完成：P001 关闭（D006,否定）；C001 判据与 C1/C1b 定义并冻结（D007–D012）；P006 预注册、83 个候选冻结（D061）、保留集判决跑完（E024,D068）并三次收窄（D070–D074）；P007 规则草稿与判决代码 v5（D075–D092）；PI 改判：交付物 = 每个数据集测试的帕累托族（D091/D096/D097）；R51 定尺寸、R53 首次精确家族、R55 多轴家族、R59 QFG z 网格、R54 P007 程序草稿（69 钉版）、R57 草稿、R60 负可行性、R61 光子拆分缓存（146/146）；P007 数据集定为变体 C（D103）；R58 防丢弃轴已完成并接受（D106，between_all；红队 RT-106 复核 D111）；R62 文献候选、R63 六轴颜色家族、R64 E_k3、R65 LR Neyman 区间、R66 可能/确定层、R67 折叠审计、R68 QPB 颜色、R70 emanom 普查、R71 IN/OUT 双比值、R72 共享耀斑零假设、R73 逐 CCD、R74/R78 观测内示踪剂、R75 掩码版 IN/OUT、R76 QFG z 复算、R77 修复零假设、R79 R57 协议草稿（D107–D123）；R80 固定保留时间比较（无通用赢家）、R81 B 态时期控制、R82 指向偏离 Simpson 成分效应、R83 掩码 + 去偏份额 + 同类（只剩 M31 核心 MOS1 E_k2.5 通过 Holm，Provisional）、R84/R88 无非劣势通用默认、R85 CCD1 软过量筛查（10/146，与 R81 的 7 个只重叠 2 个，Jaccard 0.13，CCD1 筛查未确立）、R86 NGC 3079 颜色间隙诊断（D125–D131）。Done: P001 closed (D006, negative); the C001 criterion and C1/C1b defined and frozen (D007–D012); the P006 pre-registration, 83 candidates frozen (D061), the held-out verdict run (E024, D068) and narrowed three times (D070–D074); the P007 rule draft and verdict code v5 (D075–D092); the PI's re-frame: the deliverable is the Pareto family per dataset test (D091/D096/D097); R51 sizing, R53 the first exact family, R55 the multi-axis family, R59 the QFG z-grid, R54 the P007 procedure draft (69 pins), R57 the draft, R60 negative feasibility, R61 the photon-split caches (146/146); the P007 data set fixed as variant C (D103); R58 the drop-proof axis done and accepted (D106, between_all; red team RT-106 review D111); R62 literature candidates, R63 the six-axis colour family, R64 E_k3, R65 the LR Neyman interval, R66 the certain/possible layers, R67 the fold audit, R68 QPB colour, R70 the emanom census, R71 the IN/OUT double ratio, R72 the shared-flare null, R73 per-CCD, R74/R78 within-observation tracers, R75 masked IN/OUT, R76 the QFG z recomputation, R77 the repaired null, R79 the R57 protocol draft (D107–D123); R80 the fixed-kept-time comparison (no universal winner), R81 B-state epoch control, R82 the pointing-offset Simpson composition effect, R83 masking plus a debiased share plus homogeneous classes (only M31 core MOS1 E_k2.5 passes Holm, Provisional), R84/R88 no non-inferior universal default, R85 the CCD1 soft-excess screen (10/146, overlapping R81's 7 in only 2, Jaccard 0.13, CCD1 screening not established), R86 the NGC 3079 colour-gap diagnosis (D125–D131).
2. (a) P007 冻结与运行（主线）。目标：在全新的天空上回答"哪些策略在每个数据集测试的帕累托族里"（D096/D097），\(\sigma_R\) 只是轴之一。输入：规则草稿 prereg/P007_rule_draft.md、R54 程序草稿与冻结清单（69 钉版）、家族代码 r58.py（between_all 轴已并入，D106）、变体 C 观测清单（54 个，D103）、判决代码 v5、R79 协议草稿。方法：家族代码（r58.py）\(\to\) R63 颜色轴并入 \(\to\) R57 协议定稿（默认：描述性精度研究，待 PI 定）\(\to\) PI 过目 \(\to\) 冻结规则文本与清单（哈希）\(\to\) 下载/还原变体 C 的 54 个观测（免费执行者）\(\to\) 建缓存 \(\to\) 一次打分 \(\to\) 家族 + 成员频率判决 \(\to\) 红队。成败检验：每个数据集测试给出其家族与每个策略的成员频率；\(\varepsilon\) 容差与参照对比全部降为描述列（D096）。时间估计：前置步骤 1–2 天，冻结半天，还原约 1–2 天（免费执行者），打分与判决数小时。先读：prereg/P007_rule_draft.md §1–§10 与 results/r54_procedure/luna/P007_procedure_draft.md。(a) Freeze and run P007 (the main line). Goal: on fresh sky, answer "which policies are in the Pareto family of each dataset test" (D096/D097), with \(\sigma_R\) only one axis among several. Inputs: the rule draft prereg/P007_rule_draft.md, the R54 procedure draft and freeze manifest (69 pins), the family code r58.py (the between_all axis folded in, D106), the variant C census list (54 observations, D103), the verdict code v5, the R79 protocol draft. Method: the family code (r58.py) \(\to\) fold in the R63 colour axes \(\to\) finalise the R57 protocol (default: a descriptive precision study, pending the PI) \(\to\) the PI reviews \(\to\) freeze the rule text and manifest (hashed) \(\to\) download/reduce the 54 observations of variant C (free executors) \(\to\) build caches \(\to\) one scoring run \(\to\) the family + membership-frequency verdict \(\to\) red team. Success/failure: each dataset test yields its family and every policy's membership frequency; the \(\varepsilon\) tolerances and the reference comparisons all become descriptive columns (D096). Time: the pre-steps 1-2 days, the freeze half a day, reduction 1-2 days (free executors), scoring and verdict hours. Read first: prereg/P007_rule_draft.md §1–§10 and results/r54_procedure/luna/P007_procedure_draft.md.
3. (b) R41（已完成，D080/D086）：给全部 83 个候选排序，定死主列表。结果：按家族取前 5 定为 XJ.MS、KFP_z3_a1、JFR_z3、LG.MS、KFR2_t600_a1（D086；D096 后降为 \(\sigma_R\) 轴描述）。原目标：把 P007 主列表从"暂定 5 个"变成"按发展数据 \(\bar D\) 排序的全部 83 个中的前 5 个"。输入：R38 功效表、83 个候选的发展集分数。方法：补齐 27 个缺发展集分数的候选（可从 gen3 per_obs 打分），排序，取前 5。成败检验：主列表的期望假 Confirmed \(\leq 0.1\)（全局零假设下）。时间估计：数小时。先读：04_DECISION_LOG.md D079。(b) R41 (done, D080/D086): rank all 83 candidates and fix the primary list. Result: the one-per-family top 5 became XJ.MS, KFP_z3_a1, JFR_z3, LG.MS, KFR2_t600_a1 (D086; after D096 a descriptive \(\sigma_R\)-axis view). Original goal: turn the P007 primary list from "5 provisional" into "the top 5 of all 83 ranked by development \(\bar D\)". Inputs: the R38 power table, the 83 candidates' development scores. Method: score the 27 candidates missing development scores (from gen3 per_obs), rank, take the top 5. Success/failure: expected false Confirmed of the primary list \(\leq 0.1\) under a global null. Time: hours. Read first: 04_DECISION_LOG.md D079.
4. (c) R42（已完成，D082/D083）：候选的增益有多少来自"删掉灾难观测"而非"删掉更脏的时间"？结果：XJ.MS、JFR_z3、GATE0_f0.15 的池化增益 57–102% 来自整段丢弃参照保留的观测；D084 加了同观测守卫（后为 D089 曝光匹配形式）。原目标：回答 P007 检验的到底是什么。输入：P006 保留集分数、每观测的 \(\tau^2\) 表。方法：对每个候选分解"增益份额 = 来自被整段删掉的观测 vs 来自同观测内更干净的时间"。成败检验：每个候选打印该份额；若主候选的份额以删观测为主，P007 的解读要写明这一点。时间估计：数小时。先读：04_DECISION_LOG.md D079（5）。(c) R42 (done, D082/D083): how much of a candidate's gain comes from "dropping catastrophic observations" rather than "cleaner time"? Result: 57-102% of the pooled gain of XJ.MS, JFR_z3, GATE0_f0.15 comes from wholly dropping observations the reference keeps; D084 added the same-observation guard (the D089 exposure-matched form). Original goal: say what P007 actually tests. Inputs: the P006 held-out scores, the per-observation \(\tau^2\) tables. Method: for each candidate, decompose the gain share into "from wholly dropped observations" vs "from cleaner time within the kept ones". Success/failure: the share is printed per candidate; if the main candidates' share is drop-dominated, the P007 reading states it. Time: hours. Read first: 04_DECISION_LOG.md D079 (5).
5. (d) 谱验证 C2 与 P002 F2（等 PI 派发）。目标：把快速判据最终落到光谱一致性上。输入：C1 里最好的 2–3 个政策与 2.filter。方法：用到 MosaicSelfCal CDFS 金/银 cell 上，按 LHB 与 halo 归一化的一致性 \(\tau\) 打分，再上 M31CGM；P002 F2 用一次新的 Lockman Hole 还原同时检验。成功标准：快速判据的排名与谱 \(\tau\) 的排名一致。时间估计：由 PI 决定派发时间。先读：prereg/C001_error_budget_criterion.md §4。
6. (e) R58：防丢弃的观测间轴（已完成，D106/D111）。结果：轴改为 between_all —— 被丢弃的观测按它"不过滤"时的水平与曝光计回。Claude 在 20 个测试上暴力复算 FULL_all 与 (between_all, T, coverage) 前沿，0 处不一致；红队 RT-106 复核确认它仍奖励"丢真异常观测"（但税不超过被丢观测自己的散度），且任何共享支撑的变体都堵不住"薄条"漏洞，所以防丢弃必须用插补。P007 家族代码改用 r58.py。(e) R58: the drop-proof between-observation axis (done, D106/D111). Result: the axis is now between_all — a dropped observation is imputed at its unfiltered level and exposure. Claude brute-force re-checked the FULL_all and (between_all, T, coverage) fronts on all 20 tests with 0 mismatches; the red team RT-106 upheld it with the readings that it still rewards dropping a genuine outlier (but the tax never exceeds the dropped observation's own scatter) and that every shared-support variant leaves the sliver exploit open, so drop protection requires imputation. The P007 family code becomes r58.py.
7. (f) R63：颜色/形状轴进家族（进行中）。目标：把能谱形状一致性（H/S、U/S 颜色）作为家族的一轴，且不重新编码 GTI 选择（RT-103 教训）。输入：R61 光子拆分缓存（146/146 精确还原）、R67/RT-111 的折叠审计、S-P06 的逐观测 QPB 色偏 \(d_i\)、S-P08 的 SWCX 旗标、S-P29 的 emanom 旗标。方法：在拆分后的独立半区上算颜色散布（注意：两半共享真实耀斑，所以它是"噪声保留"而非独立数据，D112），非零 \(\tau\) 区间建好后并入 R57。成败检验：颜色轴只加成员不删成员（R56 的先例），且与保留率的相关显著低于 R56 的 \(r=0.82\)。时间估计：1–2 天。先读：04_DECISION_LOG.md D104、D105、D112。(f) R63: the colour/shape axes into the family (running). Goal: add spectral-shape consistency (the H/S and U/S colours) as a family axis without re-encoding the GTI selection (the RT-103 lesson). Inputs: the R61 photon-split caches (146/146 exact), the R67/RT-111 fold audit, the S-P06 per-observation QPB colour offsets \(d_i\), the S-P08 SWCX flag, the S-P29 emanom flags. Method: compute colour scatter on the independent halves after the split (note: the halves share the real flares, so it is "noise-held-out", not independent data, D112); join R57 once the non-zero-\(\tau\) intervals exist. Success/failure: the colour axis adds members only (the R56 precedent) and correlates with the retained rate well below R56's \(r=0.82\). Time: 1-2 days. Read first: 04_DECISION_LOG.md D104, D105, D112.
8. (g) R57：同天区谱形状验证 —— 改为描述性精度研究（默认待 PI 定）。目标：如实报告同天区颜色散度被估计得多准，不给"一致"或"已认证"的标签。输入：R79 协议草稿（区间设计 (b)，\(\tau=0\) 条件覆盖 0.991；\(k=8\) 上附功效曲线与 MDE80 边界约 2.7 log-pp）、R61 缓存、R69 的覆盖表。方法：按协议逐"政策 \(\times\) 相机 \(\times\) 颜色"格报告 \(\hat\tau\)、LR 区间端点与宽度、保留观测数与越界旗标；散度随曝光变化的情形（无条件覆盖 0.77–0.93）作为模型敏感性写明。成败检验：每个格都报告，没有格被静默丢弃；协议里没有一个"consistent/certified"字样。时间估计：半天定稿，加 PI 决定。先读：.agent/jobs/2026-10-20-r79-r57-precision-protocol/R57_descriptive_protocol.md 与 04_DECISION_LOG.md D118、D122。(g) R57: the same-sky spectral-shape validation — recast as a descriptive precision study (default pending the PI). Goal: report honestly how precisely same-sky colour scatter is estimated, with no "consistent" or "certified" label. Inputs: the R79 protocol draft (interval design (b), conditional coverage 0.991 at \(\tau=0\); at \(k=8\) the power curve and an MDE80 boundary of about 2.7 log-pp are reported alongside), the R61 caches, R69's coverage tables. Method: per policy \(\times\) camera \(\times\) colour cell, report \(\hat\tau\), the LR interval endpoints and width, the retained-observation count and boundary flags; the exposure-dependent case (unconditional coverage 0.77–0.93) is stated as a model-sensitivity limit. Success/failure: every cell is reported and none is silently dropped; the protocol contains no "consistent/certified" wording. Time: half a day to finalise, plus the PI's decision. Read first: .agent/jobs/2026-10-20-r79-r57-precision-protocol/R57_descriptive_protocol.md and 04_DECISION_LOG.md D118, D122.
9. (h) R82：找观测之间大颜色斜率的混杂因子（已完成，D125）。目标：解释为什么观测之内的软质子示踪剂斜率 \(\approx0\) 而观测之间的斜率很大。输入：R78 的观测内斜率、R75 的观测间斜率、S-P41/S-P42 的星系发射线索、S-P23 的时期线索、S-P38 的 SWCX 线索。方法：在观测之间分离"星系发射 / 观测时期 / SWCX"三个候选，各自给出可检验的预言。结果：M31 上按指向偏离调整后过滤斜率降 88–92%（QFG M1 +1.2088 → +0.1481，M2 +1.0702 → -0.0842，WLS 精确复算）—— 17 个核心与 10 个盘/晕指向之间的 Simpson 悖论；NGC 3079 是时期混合（去掉 2001/2003 两个早期观测后 M1 \(+0.05\pm0.11\)）。成败检验通过：结论写成"混杂来自指向偏离/时期"，不是"软质子驱动颜色"；"残余软质子驱动过滤后颜色散度"在发展集上关闭。时间：约 1 天。先读：04_DECISION_LOG.md D119、D123、D125。(h) R82: hunt the confounder behind the large between-observation colour slopes (done, D125). Goal: explain why the within-observation soft-proton tracer slope is \(\approx0\) while the between-observation slope is large. Inputs: R78's within-observation slopes, R75's between-observation slopes, the S-P41/S-P42 galaxy-emission clues, the S-P23 epoch clue, the S-P38 SWCX clue. Method: separate the three candidates — galaxy emission / observation epoch / SWCX — between observations, each with its own testable prediction. Result: on M31, adjusting for the pointing offset removes 88-92% of the filtered slope (QFG M1 +1.2088 → +0.1481, M2 +1.0702 → -0.0842, replayed exactly by WLS) — a Simpson's paradox between the 17 core and 10 disc/halo pointings; on NGC 3079 it is an epoch mix (without the two early 2001/2003 observations, M1 \(+0.05\pm0.11\)). Success/failure passed: the conclusion is written "the confounding comes from the pointing offset / the epoch", not "soft protons drive colour"; "residual soft protons drive the filtered colour scatter" is closed on the development sets. Time: about 1 day. Read first: 04_DECISION_LOG.md D119, D123, D125.
10. (i) R80：在固定保留时间下比较阈值的污染度（进行中）。目标：回答"成员数随阈值 z 增长是不是质量"，因为 R76 已证明它全部来自保留时间轴。输入：R76 的 z 网格复算（去掉 T 只剩 z = 1 与 z = 2 入族；\(\sigma_R\) 最优的 z 在 14/20 个测试是 1）、S-P52 的文献结论（没有已发表研究在固定保留时间下比较阈值）。方法：固定保留时间，比较各阈值的污染度，并用预先声明的优值（figure of merit）排序。成败检验：给出的比较不再依赖保留时间轴；优值在跑数据前写定。时间估计：1 天。先读：04_DECISION_LOG.md D122、D123。(i) R80: compare thresholds' contamination at fixed kept time (running). Goal: answer "does membership growth with the threshold z mean quality", since R76 showed it comes entirely through the kept-time axis. Inputs: R76's z-grid recomputation (without T only z = 1 and z = 2 remain members; the \(\sigma_R\)-optimal z is 1 in 14/20 tests), S-P52's literature conclusion (no published study compares thresholds at fixed kept time). Method: hold kept time fixed, compare each threshold's contamination, and rank with a pre-declared figure of merit. Success/failure: the comparison no longer depends on the kept-time axis, and the figure of merit is written down before the data are run. Time: 1 day. Read first: 04_DECISION_LOG.md D122, D123.
11. (j) R81：修 RT-117 的三个小问题（进行中）。目标：让逐 CCD 的 B 态分析没有已知的机械缺陷。输入：RT-117 的发现（76 个零事件 CCD 仍带活时间；面积与 vignetting 缩放 \(\leq0.32\) log-pp；CCD1 无法被 emanom 筛查）、R73 的逐 CCD 计数。方法：修这三处后重跑 CDFS M2 none 的 6.3404 \(\to\) 5.9149 复算，确认结论不变。成败检验：三处修复都有回归测试；R70 的 MOS2 敏感性案例重新标注为"观测子集"而非"异常 CCD 修正"（D121 已定）。时间估计：半天。先读：04_DECISION_LOG.md D123。(j) R81: repair RT-117's three minor items (running). Goal: make the per-CCD B-state analysis free of known mechanical defects. Inputs: RT-117's findings (76 zero-event CCDs still carry live time; area and vignetting scaling \(\leq0.32\) log-pp; CCD1 cannot be screened by emanom), R73's per-CCD counts. Method: repair the three items and re-run the CDFS M2 none 6.3404 \(\to\) 5.9149 replay to confirm the conclusion is unchanged. Success/failure: all three repairs carry a regression test; R70's MOS2 sensitivity case is re-labelled "observation subset", not "anomalous-CCD correction" (settled in D121). Time: half a day. Read first: 04_DECISION_LOG.md D123.
(d) Spectral validation C2 and P002 F2 (await the PI's dispatch). Goal: land the quick criterion on spectral consistency at last. Inputs: the 2–3 best C1 policies and 2.filter. Method: apply them to the MosaicSelfCal CDFS golden/silver cells, score the consistency \(\tau\) of the LHB and halo normalisations, then M31CGM; P002 F2 is tested together with one fresh Lockman Hole reduction. Success: the quick-metric ranking matches the spectral-\(\tau\) ranking. Time: the PI decides the dispatch. Read first: prereg/C001_error_budget_criterion.md §4.

一分钟看懂核心矛盾（动画）The core conflict in one minute (animation)

本次更新What's new

2026-10-21（第八次更新，折入 D125–D130；修复路径泄露）:

2026–10-21 (eighth edition; D125–D130 folded in; path leaks fixed):

• 观测间大颜色斜率的混杂因子已找到（D125，R82）：M31 上按指向偏离调整后过滤斜率降 88–92%（QFG M1 +1.2088 → +0.1481，M2 +1.0702 → -0.0842，WLS 精确复算）—— 17 个核心与 10 个盘/晕指向之间的 Simpson 悖论；NGC 3079 是时期混合。"残余软质子驱动过滤后颜色散度"在发展集上关闭。The confounder behind the large between-observation colour slopes is found (D125, R82): on M31, adjusting for the pointing offset removes 88–92% of the filtered slope (QFG M1 +1.2088 → +0.1481, M2 +1.0702 → -0.0842, replayed exactly by WLS) — a Simpson's paradox between the 17 core and 10 disc/halo pointings; on NGC 3079 it is an epoch mix. "Residual soft protons drive the filtered colour scatter" is closed on the development sets.
• 没有"通用最优 GTI"（D125/D126/D131，R80/R84/R88）：固定保留时间下没有任何政策、任何 z 在全部 20 个测试胜出（L0–L0.7 各有 10/9/11/12 个不同赢家）；也没有任何政策在全部 20 个测试"从不显著更差"（\(\sigma_R\) 单轴最多 15/20，联合 4–5/20；R88 用预先声明的非劣势边界与 Romano–Wolf 也一样）。帕累托族作为报告形式成立。No universal "best GTI" (D125/D126/D131, R80/R84/R88): at fixed kept time no policy and no z wins across all 20 tests (10/9/11/12 distinct winners at L0-L0.7); and no policy is never-significantly-worse in all 20 tests either (at most 15/20 for \(\sigma_R\) alone, 4–5/20 jointly; R88 with a pre-declared non-inferiority margin and Romano-Wolf is the same). The Pareto family as the report form stands.
• 只有一个过滤 cell 通过 Holm，且是 Provisional（D126/D127，R83）：掩码 + 去偏份额 + 同类后，16 个过滤 cell 里只剩 M31 核心 MOS1 E_k2.5（去偏份额 +0.617，p \(0.002\times16\) = 0.032，斜率 +0.640）；不过滤的 none 对照在每一类都被示踪剂解释（+0.46…+0.70）。D127 指出它可能是点源变化或 PSF 泄漏，因此 Provisional。Only one filtered cell passes Holm, and it is Provisional (D126/D127, R83): after masking, a debiased share and homogeneous classes, only M31 core MOS1 E_k2.5 of 16 filtered cells survives (debiased share +0.617, p \(0.002\times16\) = 0.032, slope +0.640); the unfiltered none control is explained by the tracer in every class (+0.46…+0.70). D127 notes it may be source variability or PSF leakage, so it stays Provisional.
• CDFS 的观测内零结果功效不足（D126，RT-120）：95% 区间 [-0.053, +0.040] 包含 +0.025，功效约 18% —— 那里只能说"在此功效下测不到"，不能说"没有"。星系结果（R78）不受影响；RT-119 的衰减校正后斜率（M31 M1 \(+0.165\pm0.064\)）仍在 \(4.9\text{–}8.5\sigma\) 上排除 R75 的观测间斜率。The CDFS within-observation null is underpowered (D126, RT-120): the 95% interval [-0.053, +0.040] contains +0.025 and the power is about 18% — there one can only say "not detectable at this power", not "absent". The galaxy result (R78) is unaffected; RT-119's attenuation-corrected slopes (M31 M1 \(+0.165\pm0.064\)) still exclude R75's between-observation slopes at \(4.9\text{-}8.5\sigma\).
• 30″ 掩码的硬化偏差（D127/D128，S-P62）：30″ 掩码漏掉一个在轴 MOS 点源的 15.5–18.5%（60″ 只漏 8–10%），轴外 5 keV 泄漏是 1.2 keV 的 1.9–2.4 倍 —— 所有 30″ 掩码颜色结果（R75/R78/R83）带一个未量化的硬化偏差；R86/R87 已要求随流量变化的掩码（S-P63 给出规则）。The 30″-mask hardening bias (D127/D128, S-P62): a 30″ mask leaves 15.5–18.5% of an on-axis MOS point source unmasked (60″ leaves 8–10%), and off-axis the 5 keV leakage is \(1.9–2.4\times\) the 1.2 keV leakage — every 30″-mask colour result (R75/R78/R83) carries an unquantified hardening bias; R86/R87 now ask for flux-dependent masks (S-P63 gives the rule).
• NGC 3079 的颜色间隙由单个 2001 观测承载（R86，D131）：晚 − 早间隙 +0.435 ln(H/S)，但 2001 观测 0110930201 是 0.02–0.15，2003 观测 0147760101 已经 0.64–0.70 —— 不是"2002/2003 对 2017"的整期分裂。滤光片/模式/SAS/FWC 全部相同；点源掩码与指向各只解释约 0.10 ln。此前"两个早期观测都在 0.02–0.15"的说法是简报错误，已更正。NGC 3079's colour gap is carried by the single 2001 observation (R86, D131): the late − early gap is +0.435 ln(H/S), but the 2001 observation 0110930201 sits at 0.02–0.15 while the 2003 observation 0147760101 is already 0.64–0.70 — not a whole-epoch "2002/2003 vs 2017" split. Filter, mode, SAS and FWC are all identical; the point-source mask and pointing each explain only about 0.10 ln. The earlier claim that "both early observations sit at 0.02–0.15" was a brief error and is corrected.
• B 态效应与 CCD1 筛查都未确立（D125/D131，R81/R85）：只有 M31 MOS2 的两次整观测 B 态剔除在时期控制下存活（QFG -4.93 log-pp，p 0.014；E_k2.5 -4.61，p 0.026），27 个检验的 Holm 校正后无一存活；CCD1 软过量筛查 10/146 与 R81 的 7 个只重叠 2 个（Jaccard 0.13），两条规则互相矛盾。Neither the B-state effect nor CCD1 screening is established (D125/D131, R81/R85): only two whole-observation B-state removals on M31 MOS2 survive epoch control (QFG -4.93 log-pp, p 0.014; E_k2.5 -4.61, p 0.026), and none survives a Holm correction over 27 tests; the CCD1 soft-excess screen flags 10/146 and overlaps R81's 7 in only 2 (Jaccard 0.13), so the two rules disagree.
• 站点访问将受限制（D129，PI 决定）：本手册在下次部署时加访问门；密码从不写进任何记录。本次刷新同时修复了 3 个页面上的绝对路径/用户名泄露（byhand、policy_ms、physics，共 52 处）。The site will be access-restricted (D129, PI decision): this handbook gets an access gate at the next deploy; the password is never written into any record. This refresh also fixed 52 absolute-path / user-name leaks on 3 pages (byhand, policy_ms, physics).

2026–10-21（第七次更新，折入 D106–D123）:

2026-10-21 (seventh edition; D106–D123 folded in):

• 观测间轴修好并关闭（D106/D111）：轴改为 between_all，被丢弃的观测按不过滤水平与曝光计回；红队 RT-106 复核确认无代码缺陷，并给出三条解读（它把丢观测策略拉向全保留值；2.filter 在 dev/ALL 上 13→12 只因漏滤一个正常观测；防丢弃必须用插补）。The between-observation axis is fixed and closed (D106/D111): the axis is now between_all, a dropped observation re-entering at its unfiltered level and exposure; the red team RT-106 upheld it with no code defect and three readings (it pulls drop-policies toward the keep-all value; 2.filter's 13→12 on dev/ALL comes only from one unfiltered normal observation; drop protection requires imputation).
• 光子拆分的诚实定性（D112/D118）：两半共享真实耀斑过程（中位相关 0.88–0.98），所以保留集颜色只是"噪声保留"；对 fold A 的选择把 fold B 的 H/S 移动 \(-0.124\) ln，R72 的共享耀斑零假设解释了它。NGC 3079 r30"空间结构"读数已撤回（RT-116），R77 修复零假设后 0/16 个 cell 排除零（D122/D123）。The photon split's honest characterisation (D112/D118): the two halves share the real flare process (median correlation 0.88-0.98), so held-out colours are "noise-held-out" only; selecting on fold A shifts fold B's H/S by \(-0.124\) ln, which R72's shared-flare null explains. The NGC 3079 r30 "spatial structure" reading is withdrawn (RT-116); after R77's repaired null, 0 of 16 cells exclude zero (D122/D123).
• R57 改为描述性精度研究（D118/D122，默认待 PI 定）：\(n=8\) 上无法认证 10% 的 H/S 裕度（S-P30 上界 \(\geq14.38\) log-pp 对 9.53；RT-112 证明 CP\(_X\) 更严；R69 设计 (b) 条件覆盖 0.991）。R79 已出协议草稿：逐格报 \(\hat\tau\)、区间、保留数与旗标，不给"一致/已认证"标签。R57 is recast as a descriptive precision study (D118/D122, default pending the PI): a 10% H/S margin cannot be certified at \(n=8\) (S-P30 upper bound \(\geq14.38\) log-pp against 9.53; RT-112 showed CP\(_X\) is stricter; R69 design (b) conditional coverage 0.991). R79 drafted the protocol: per cell report \(\hat\tau\), the interval, retained counts and flags, with no "consistent/certified" label.
• 残余软质子：观测之内 \(\approx0\)，观测间是混杂（R74/R78）：CDFS 的观测内 H/S 斜率 \(-0.0061\pm0.0237\)（M1），M31 核心 \(+0.094\pm0.036\)，NGC 3079 \(+0.136\pm0.086\) —— 比观测间斜率小一个量级；QPB 颜色扣减与逐 CCD B 态剔除基本不降颜色散度（R68/R73）。R82 去找混杂因子。Residual soft protons: \(\approx0\) within observations, confounding between them (R74/R78): the within-observation H/S slope is \(-0.0061\pm0.0237\) on CDFS (M1), \(+0.094\pm0.036\) on the M31 core and \(+0.136\pm0.086\) on NGC 3079 — an order of magnitude below the between-observation slopes; QPB colour subtraction and per-CCD B-state removal barely reduce colour scatter (R68/R73). R82 hunts the confounder.
• 帕累托：QFG 成员数随 z 的增长是保留时间产物（R76/D122）：去掉 T 只剩 z = 1 与 z = 2 入族，\(\sigma_R\) 最优的 z 在 14/20 个测试是 1；文献建议在固定保留时间下比污染度并预先声明优值（S-P52）→ R80。R70 的 MOS2 B 态敏感性案例重新标注为"观测子集"（D121）。Pareto: QFG membership growth with z is a kept-time artefact (R76/D122): without T only z = 1 and z = 2 remain members, and the \(\sigma_R\)-optimal z is 1 in 14/20 tests; the literature suggests comparing contamination at fixed kept time with a pre-declared figure of merit (S-P52) → R80. R70's MOS2 B-state sensitivity case is re-labelled "observation subset" (D121).
• 命名更正（S-P37，再经 D124 修正）：dev-virgo.01 是 M31（17 个核心指向 + 10 个盘/晕指向），dev-virgo.02 是 NGC 3079（"virgo" 是还原服务器的名字）；所有"Virgo 团星系天区"的读法作废。此前"dev-virgo.01 = M31 核心"的标签是错的（一个文件的 OBJECT 不等于整个集合）；R75 的 M31 过滤后正斜率因此是合并读数，只有 none 对照的符号稳健。Naming correction (S-P37, amended by D124): dev-virgo.01 is M31 (17 core + 10 disc/halo pointings) and dev-virgo.02 is NGC 3079 ("virgo" is the reduction server's name); every "Virgo cluster sky" reading is void. The earlier label "dev-virgo.01 = the M31 core" was wrong (one file's OBJECT is not the set); R75's M31 filtered positive slope is therefore a pooled figure, and only the none control's sign is robust.
• 诚实更正（红队，全部复算）：RT-108 撤回 R65 的全局截断（约 2.1 倍过保守，是 2160 个含噪 cell 分位数的最大值）；RT-115 抓住 R71 的 IN 计数从未掩过点源，D119 的份额撤回待 R75；RT-114 纠正 R59 的 z 读数（R76 复算确认 6/4/8/9/11/13）。Honest corrections (red teams, all replayed): RT-108 withdrew R65's global cutoff (about 2.1\(\times\) over-conservative, the maximum over 2160 noisy cell quantiles); RT-115 caught R71's IN counts never source-masked, so D119's shares are withdrawn pending R75; RT-114 corrected R59's z reading (R76 confirmed 6/4/8/9/11/13).
• 页面更新：组会页新增 2026-10-21 块与 10-20 块的追加注；开放问题表关闭"防丢弃轴"、改写 R57 行并新增"观测间混杂"行；计划改为 (a)–(j) 十步；目标表第 3/6 行更新；论文草稿与报告改写受影响章节；物理页新增"为什么同一片天两次观测颜色会不同"的解释与新图。Pages updated: the meetings page gains the 2026-10-21 block plus an appendix note in the 10-20 block; the open-problems table closes the drop-proof axis row, rewrites the R57 row and adds a "between-observation confounder" row; the plan becomes ten steps (a)-(j); objectives rows 3 and 6 updated; the paper draft and seminar rewrite the affected sections; the physics page gains a "why two visits of the same sky differ in colour" explanation with a new figure.

2026-10-20（第六次更新，折入 D080–D105）:

2026-10-20 (sixth edition; D080–D105 folded in):

• 头条：PI 把结果改判为"帕累托族"（D096/D097） —— 在每个明确的数据集测试里，只要没有别的策略在所有轴上至少持平且在一轴上严格更好，这个策略就保留；没有单一赢家，没有对参照的通过/失败；旧的 Confirmed 规则降为 \(\sigma_R\) 轴的描述性读数。保留时间回到判据里当一个轴（D091），不再是可以一票否决的门槛。Headline: the PI re-framed the result as a "Pareto family" (D096/D097) — within each explicit dataset test, a policy is kept when no other is at least as good on every axis and strictly better on one; no single winner, no pass/fail against a reference; the old Confirmed rule becomes a descriptive \(\sigma_R\)-axis view. Kept time returns to the criterion as an axis (D091), no longer a gate.
• R55 家族首跑（D101）：20 个数据集测试 \(\times\) 103 个策略，精确支配下 2.filter 在 19/20 个测试是成员、E_k2.5 20/20、none 20/20；D086 五领导 8-17/20，靠保留时间/覆盖率轴。R59 把 QFG 的 z 网格全部当候选（不再"选" z）。红队 RT-101/RT-103 抓出观测间轴可被"丢观测"操纵、硬度轴重新编码选择 —— 成员随覆盖数打印，两轴保持描述性。The R55 family first run (D101): 20 dataset tests \(\times\) 103 policies; under exact dominance 2.filter is a member in 19/20, E_k2.5 20/20, none 20/20; the D086 five leaders 8-17/20, through the kept-time/coverage axes. R59 puts QFG's whole z-grid in as candidates (no z is "chosen"). The red teams RT-101/RT-103 caught the between axis being gameable by dropping observations and the hardness axis re-encoding the selection — members are printed with coverage counts and both axes stay descriptive.
• P007 定形：数据集定为变体 C（54 观测，D103）；NGC 5907 接受为同天扩展（D100）；同天区谱形状验证（R57）获批并成稿，但未冻结 —— R60 干跑证明旧缓存不可用（负可行性），R61 用伯努利 50/50 光子拆分重建缓存（146/146 精确还原，\(\tau=0\) 区间覆盖 92.5-100%）。P007 takes shape: the data set is variant C (54 observations, D103); NGC 5907 accepted as same-sky extension (D100); the same-sky spectral-shape validation (R57) approved and drafted, but not frozen — the R60 dry run proved the old caches unusable (negative feasibility), and R61 rebuilt them with a Bernoulli 50/50 photon split (146/146 exact, \(\tau=0\) interval coverage 92.5-100%).
• 诚实更正（评审与红队，全部复算）：D085 撤回"功效区间"措辞（发展替身与已用数据是两个情景，不是上下界）；RT-99 撤回 QFG 的"领先者"称呼（优势是 \(\varepsilon_T\) 负载的）；D093 的"未完全定种子"注记被 RT-98 撤回（逐位一致）；R51 Hermes 臂与 codexnim 臂一致。Honest corrections (reviews and red teams, all replayed): D085 withdrew the "power bracket" wording (development stand-ins and spent data are two scenarios, not bounds); RT-99 withdrew QFG's "leader" label (the edge is \(\varepsilon_T\)-contingent); D093's "sizing not fully seeded" note was withdrawn by RT-98 (bit-exact); the R51 Hermes arm agrees with the codexnim arm.
• 文献与专家：没有已发表的"保留时间 \(\times\) 重复观测一致性"联合帕累托处理（S-P01，16 源）；专家建议保留 none 与完整精确族、加 \(\varepsilon=(0,0)\)、A3 可合法地无结论（D098）。Literature and experts: no published joint Pareto treatment of retained time vs repeated-observation consistency (S-P01, 16 sources); the expert advised keeping none and the full exact family, adding \(\varepsilon=(0,0)\), and that A3 may legitimately end unresolved (D098).
• 页面更新：组会页新增 2026-10-20 块与 10-19 块的更正注；目标表加第 6 行（按数据集测试评价家族）；开放问题表加两行（防丢弃轴、R57 冻结）；计划改为 (a)–(g) 七步；论文草稿与报告已改写为帕累托族框架。Pages updated: the meetings page gains the 2026-10-20 block plus a correction note in the 10-19 block; the objectives table gains row 6 (evaluate the family per dataset test); the open-problems table gains two rows (the drop-proof axis, the R57 freeze); the plan becomes seven steps (a)-(g); the paper draft and seminar are rewritten to the Pareto-family framing.

2026-10-19（第五次更新，组会页与论文/报告两页新增；折入 D058–D079）:

2026-10-19 (fifth edition; new group-meeting, paper-draft and seminar pages; D058–D079 folded in):

• 新页面：组会档案（meetings.html）—— 每个 UTC 日一块，四部分固定为成果 / 工作 / 问题 / 待 PI 决定，只用有日期的记录写；首页顶部是最新一天的完整块，更早的是一行链接。New page: the group-meeting archive (meetings.html) — one block per UTC day with the four fixed parts Results / Work / Problems / For the PI, written only from dated records; the home page carries the latest day in full and earlier days as one-line links.
• 新页面：论文草稿（paper.html）与报告幻灯片加完整讲稿（seminar.html）—— 面向不认识本项目的graduate student与教授，按普通期刊与组会实践写；只收已确立的结果，每个数字带来源标记。New pages: the paper draft (paper.html) and the seminar slides with the full talk script (seminar.html) — written for graduate students and professors who do not know the project, following ordinary journal and seminar practice; only established results, every number with a source marker.
• 保留集判决跑完（P006,2026-10-18）：在 119 个从未参与开发的观测上，83 个冻结候选中 79 个相对归档 2.filter 是 Confirmed（23 个家族），中位 \(\delta\) 为 HO-M31 MOS1 \(-10.1\) pp、MOS2 \(-3.7\) pp；HO-blank \(-3.2\) / \(-2.9\)；HO-mixed \(-2.9\) / \(-3.5\) pp。判决由冻结代码链产生，Claude 逐字节重放，红队执行 QA 0 条实质问题。The held-out verdict ran (P006, 2026-10-18): on 119 observations that took no part in development, 79 of 83 frozen candidates are Confirmed against the archived 2.filter (23 families), median \(\delta\) HO-M31 MOS1 \(-10.1\) pp, MOS2 \(-3.7\) pp; HO-blank \(-3.2\) / \(-2.9\); HO-mixed \(-2.9\) / \(-3.5\) pp. The verdict came from the frozen code chain, was replayed byte-identically by Claude, and the red-team execution QA found 0 material.
• 同日三次收窄（诚实限定，全部复算）：① 2.filter 是弱参照 —— 204 个非候选政策里也有 50 个通过同一规则；② 对复现的标准光变曲线滤波器（espfilt 加硬段裁剪，EH_B1000_n2）只剩 11/83，且全部保留更少时间（0.32–0.76 对 0.81–0.93）；③ 对"等保留时间的最好复现滤波器"配对自举 0–1/83（点估计 37/83，对种子脆弱）—— 未确立。Three narrowings the same days (honest limits, all replayed): ① 2.filter is a weak reference — 50 of 204 non-candidate policies also pass the same rule; ② against the reproduced standard light-curve filter (espfilt plus a hard-band clip, EH_B1000_n2) only 11 of 83 survive, all keeping less time (0.32–0.76 vs 0.81–0.93); ③ against the best reproduced-filter setting at matched kept time a paired bootstrap gives 0–1 of 83 (point estimate 37/83, seed-fragile) — not established.
• HO-M31 MOS1 的 \(-10.1\) pp 拆分（R29）：约 4.6 pp 来自一个损坏的归档 2.filter GTI（0202230301），约 5.5 pp 是候选在更难环境下的自身性质（去一观测法 \(-3.2\) 到 \(-6.1\) pp）。Al-K 线门每格只在 0–3 个观测上触发，所以判决没有检验它。The HO-M31 MOS1 \(-10.1\) pp is split (R29): about 4.6 pp from one broken archived 2.filter GTI (0202230301) and about 5.5 pp a candidate property in a harder regime (leave-one-out \(-3.2\) to \(-6.1\) pp). The Al-K line gate fires on 0–3 observations per cell, so the verdict does not test it.
• P007 规则草稿 1 与判决代码就绪（D075–D079）：主估计量 = 六格等权平均 \(\bar D\) 对 EH_B1000_n2（\(\delta_0=0.5\) pp，伤害守卫，六格全有效）；状态 Confirmed / No gain \(\geq 0.5\) shown / Ambiguous / Invalid；规模 \(\leq 3\%\)（集 A）/ 1.4%（集 C）；发展集功效只报区间（0.013–0.164）。R36 的 0.42 功效数字被撤回（样本内值）。The P007 rule draft 1 and the verdict code are ready (D075–D079): primary estimand = the equal-weight six-cell mean \(\bar D\) vs EH_B1000_n2 (\(\delta_0=0.5\) pp, harm guard, all cells valid); states Confirmed / No gain \(\geq 0.5\) shown / Ambiguous / Invalid; size \(\leq 3\%\) (set A) / 1.4% (set C); development power quoted only as a range (0.013–0.164). R36's 0.42 power figure was withdrawn (in-sample).
• 新参照 QFG（只用光变曲线、无 2.filter、无角落）通过发展筛选 8/8；在已用过的保留集数据上对 2.filter Confirmed 6/6，但对 EH_B1000_n2 Not confirmed（\(\bar D=-0.70\) pp，Ambiguous）。The new reference QFG (light curves only, no 2.filter, no corner) passes the development screen 8/8; on the spent held-out data it is Confirmed 6/6 vs 2.filter but Not confirmed vs EH_B1000_n2 (\(\bar D=-0.70\) pp, Ambiguous).
• P002 F1 的"CDFS MOS2 TRACKS"读数已于 2026-10-15 撤回（红队检查 RT-02，E004 更正）：那是观测时期造成的假象，加每时期偏移后 \(b=0.690\pm0.020\)，四集都是"中间"。The P002 F1 "CDFS MOS2 TRACKS" reading was withdrawn on 2026-10-15 (red-team check RT-02, E004 correction): it was an epoch artefact; with per-epoch offsets \(b=0.690\pm0.020\) and all four sets are INTERMEDIATE.

2026-10-16（第四次更新，新增前沿策略对比页；当日二次折入 D2 记录）:

2026-10-16 (fourth edition, a new frontier-policies page, with the D2 records folded in the same day):

• 新页面"前沿策略对比"（frontier.html）：十个 GTI 政策（2.filter、6 个搜索政策、MS_z3.5_core、EH_B1000_n2、Q_d0.05_n2、none）的光变曲线、单观测与全观测叠加的帕累托前沿、31 观测配对一致性、M31 竖直剖面，以及"某政策在前沿哪里"的下拉选择（选一个政策，两张前沿图换成它的高亮版）。每张图背后的"未检出"与"待 PI 决定"都写明。New page "Frontier policies" (frontier.html): light curves, the single-observation and stacked Pareto fronts, the 31-observation paired consistency, and the M31 vertical profile for ten GTI policies (2.filter, 6 search policies, MS_z3.5_core, EH_B1000_n2, Q_d0.05_n2, none), plus a "where does policy X sit" dropdown that switches both front figures to the chosen policy's highlighted version. The "not detected" and the "pending with the PI" are stated alongside every figure.
• 头条数字：在 25 个没人设计过的测试观测上，gen2:hermes.FLR_z3 把叠加 \(\sigma_R\) 从 5.15% 降到 3.81%（MOS1，配对差 −1.34 pp，自举展宽 [−1.98, −0.65]）、从 5.84% 降到 3.93%（MOS2，−1.92 [−2.58, −1.18]），保留 0.85–0.87 的活时间；FLR_z3、MS_z3.5_core、RT3_h2.5_l1 这 3 个政策在 4 个集 \(\times\) 2 台相机共 8 种组合上都优于 2.filter（R9）。Headline: on the 25 test observations no designer saw, gen2:hermes.FLR_z3 takes the stacked \(\sigma_R\) from 5.15% to 3.81% (MOS1, paired difference −1.34 pp, bootstrap spread [−1.98, −0.65]) and from 5.84% to 3.93% (MOS2, −1.92 [−2.58, −1.18]) while keeping 0.85–0.87 of live time; these 3 policies (FLR_z3, MS_z3.5_core, RT3_h2.5_l1) beat 2.filter in all 8 \(\text{set}\times\text{camera}\) cases (4 sets \(\times\) 2 cameras; R9).
• 同日折入的诚实限定：① RT-16 F-01 —— 优势主要是修好 2.filter 失效的观测（M31 MOS1 去掉 2 个最差观测后 FLR_z3 只剩 −0.79 pp，MS_z3.5_core 变成 +0.33 pp 持平）；② RT-18 —— \(\sigma_{RC}\) 降级，只用 \(\sigma_R\) 排名并另报每个政策丢掉的观测数；③ E022 —— FLR_z3 的优势来自绝对硬度门 \(r_0=1.8\)，同一个门在 Virgo-server 集上丢掉 51/142（MOS1，34% 时间）与 54/142（MOS2，38%）个观测，主候选正在被重新询问；④ RT-15 —— 竖直剖面读法取决于曝光模型（平坦曝光下 2.filter 在 2 台相机都排最后）；⑤ E021 —— R2 跨观测静息参考跑完，没有产生新领先者，但场参照门能拒绝 LH 0101；⑥ RT-17/R10 —— "复现 2.filter" 改为文件级。Honest limits folded in the same day: ① RT-16 F-01 — the advantage is mostly the repair of 2.filter's failed observations (on M31 MOS1, removing the 2 worst observations leaves FLR_z3 at only −0.79 pp and turns MS_z3.5_core into +0.33 pp, no significant difference); ② RT-18 — \(\sigma_{RC}\) is demoted, so ranking uses \(\sigma_R\) only with the dropped-observation count reported separately; ③ E022 — FLR_z3's edge comes from its absolute hardness gate \(r_0=1.8\), and the same gate drops 51/142 observations (MOS1, 34% of the time) and 54/142 (MOS2, 38%) on the Virgo-server set, so the main candidate is being re-asked; ④ RT-15 — the vertical-profile reading depends on the exposure model (under a flat exposure 2.filter is last on 2 cameras); ⑤ E021 — R2's cross-observation quiescent reference finished with no new leader, but a field-referenced gate does reject LH 0101; ⑥ RT-17/R10 — "reproducing 2.filter" is now file-level.
• 同时如实记录：竖直剖面的中位 \(\tau_b\) 改善"未检出"（检验功效 \(1\text{–}2\sigma\)，RT-12 F-01）；池化的每观测水平偏移上 FLR_z3 在 MOS2 上 −2.0 pp（95% [−3.3, −0.7]）、Q_d0.05_n2 在 2 台相机都降，但该统计量是事后选择的，待 R7 预注册；Lockman Hole 的观测间项 14.6% / 23.3% 几乎全由 0022740101 一个观测贡献（RT-06，espfilt 锁到耀斑上）。Recorded honestly alongside: the median-\(\tau_b\) profile improvement is "not detected" (the test has \(1\text{–}2\sigma\) power, RT-12 F-01); on the pooled per-observation offset statistic FLR_z3 lowers MOS2 by −2.0 pp (95% [−3.3, −0.7]) and Q_d0.05_n2 lowers 2 cameras, but that statistic was chosen after the fact and awaits R7's pre-registration; the Lockman Hole between term, 14.6% / 23.3%, is carried almost entirely by the single observation 0022740101 (RT-06, espfilt locking onto a flare).

2026-10-14（第三次更新，D007–D012 刷新）:

2026-10-14 (third edition, the D007-D012 refresh):

• PI 把下游成功判据定义为误差预算(C001,D007/D008):单观测看 \(\sigma_{\mathrm{tot}}\)，多观测先看一致性；0.5–2 keV 拼接光变曲线的 \(s_{\mathrm{sys}}=17\text{–}25\%\)（2.filter），对 \(s_{\mathrm{stat}}\) 0.04–0.2% —— 系统误差约大 \(100\) 倍。The PI defined the downstream success criterion as the error budget (C001, D007/D008): \(\sigma_{\mathrm{tot}}\) for a single observation, consistency first over a set; the 0.5–2 keV concatenated-light-curve \(s_{\mathrm{sys}}\) is 17–25% (2.filter) against \(s_{\mathrm{stat}}\) 0.04–0.2% — the systematic is about \(100\times\) larger.
• 历史的 M31 2.1.filter 在多个观测上灾难性失败（3 个 MOS2、5 个 MOS1 完全不切;另一些只留 0.1–2.9 ks），且其"1.6%"被撤回：它在 0.2–10 keV 上选择，包含 0.5–2 keV 打分带 —— 选择与打分同一批事件给出 0.0%。The historical M31 2.1.filter fails catastrophically on several observations (no cut at all on 3 MOS2 and 5 MOS1; 0.1–2.9 ks left on others), and its "1.6%" is withdrawn: it selects on 0.2–10 keV, which contains the 0.5–2 keV score band — selecting and scoring on the same events gives 0.0%.
• C1（D010）：每个观测一个角落项把观测间部分从 ~12% 降到 3–4%（CDFS）;剩下的主导项是 2.filter 留下的 ~12% 观测内变化。C1 (D010): a per-observation corner term takes the between part from ~12% to 3–4% (CDFS); the dominant remainder is the ~12% within-observation variability left by 2.filter.
• C1b（D011）：2.filter 内的不相交能段（FOV 2.5–8 keV）裁剪（B = 300 s,n = 1）把观测内项从 11.6–12.7% 减到 4.7–6.0%（100 s），代价 20–26% 活时间；\(\ge1000\) s 时 0–2.6%，达到 ~2% 目标。CDFS 角落修正后 \(s_{\mathrm{sys}}\):12.5% → 6.1%（MOS2）、12.0% → 5.2%（MOS1）。C1b (D011): a disjoint-band (FOV 2.5–8 keV) clip inside 2.filter (B = 300 s, n = 1) cuts the within term from 11.6–12.7% to 4.7–6.0% (100 s) for 20–26% of live time; at \(\ge1000\) s it is 0–2.6%, meeting the ~2% target. CDFS corner-corrected \(s_{\mathrm{sys}}\): 12.5% → 6.1% (MOS2), 12.0% → 5.2% (MOS1).
• P002 批准、冻结（1eb687b）并跑完 F1（D009）：CDFS MOS2 \(b=1.015\pm0.030\) → TRACKS;其余三集 \(b\approx0.6\text{–}0.7\) → INTERMEDIATE。"Al 线是粒子监测器"未在一般意义上确立。F3 指向线测量里的 K/B 系统误差（T005）。P002 was approved, frozen (1eb687b) and F1 run (D009): CDFS MOS2 \(b=1.015\pm0.030\) → TRACKS; the other three sets give \(b\approx0.6\text{–}0.7\) → INTERMEDIATE. "The Al line is a particle monitor" is not established in general. F3 points to a K/B systematic in the line measurement (T005).
• 100 s cell 上约 5% 的快速"白"残差是新的未解问题：裁剪后仍在、不跟随角落（观测内角落回归只移走 \(\le0.7\) 点）、会从谱中平均掉；其来源未测（事件聚簇 Fano \(\approx1.17\)，或阈值下的快速软质子闪烁）。The ~5% fast "white" residual at 100 s cells is a new open question: it survives the clip, does not follow the corner (a within-observation corner regression removes \(\le0.7\) points), and averages out of spectra; its origin is untested (event clustering, Fano \(\approx1.17\), or fast soft-proton flicker below the clip).
• 下一步（PI,D011/D012）:冻结已完成（H,B = 300 s,n = 1;cell 尺寸 1000 s;commit abc5e602），下一次是一次 Lockman Hole 还原，同时检验 P002 F2 与冻结滤波器；C2 在 P002 之后。Lockman Hole"后面再派"，不预设日期。Next (PI, D011/D012): the freeze is done (H, B = 300 s, n = 1; cell size 1000 s; commit abc5e602); the next step is one Lockman Hole reduction testing both P002 F2 and the frozen filter; C2 after P002. The Lockman Hole is dispatched later; no date invented.
• 新页面："误差预算与 Al 线"(budget.html),从零讲 C001 与 P002,含撤回的 1.6% 一节；两张新图:P002 F1 四集散点、C1b cell 尺寸依赖。New page: "Error budget and the Al line" (budget.html), teaching C001 and P002 from zero, including the withdrawn 1.6% claim; two new figures: the P002 F1 four-set scatter and the C1b cell-size dependence.

2026-10-13（第二次更新，D006 刷新）:P001 关闭，主问题否定;CDFS 盲测完成（E003）;角落按数据集回答;探索性角落–Al 张力（后由 P002 接手）。

2026-10-13 (second edition, the D006 refresh): P001 closed negative; the CDFS blind run complete (E003); the corner answered per set; the exploratory corner-vs-Al tension (later taken over by P002).

2026-10-13(第一版）:本手册随 ParetoGTICal 迁移入 DXC 后建立，覆盖 D001–D005。

2026-10-13 (first edition): this handbook was built after ParetoGTICal migrated into DXC, covering D001–D005.

数据截至Data as of

2026-10-21T04:32Z（所读最新来源：cognition-os/04_DECISION_LOG.md D125–D130（至 2026-10-20 23:32 EDT）与 D131（2026-10-21 00:26 EDT）；.agent/jobs/2026-10-20-r86-ngc3079-epoch-gap/REPORT.codexnim.md（NGC 3079 颜色间隙诊断）；cognition-os/00_AI_BRIEFING.md；cognition-os/08_RUNNING.md；cognition-os/RESEARCH_PLAN.md（R49–R88 行）；cognition-os/REDTEAM_QUEUE.md（RT-96–RT-122）；cognition-os/SEARCH_QUEUE.md（S-P14–S-P70）；cognition-os/06_AGENT_LOG.md 至 2026-10-21 00:26 EDT；.agent/jobs/2026-10-20-r79-r57-precision-protocol/R57_descriptive_protocol.md）。

2026-10-21T04:32Z (newest sources read: cognition-os/04_DECISION_LOG.md D125-D130 (to 2026-10-20 23:32 EDT) and D131 (2026-10-21 00:26 EDT); .agent/jobs/2026-10-20-r86-ngc3079-epoch-gap/REPORT.codexnim.md (the NGC 3079 colour-gap diagnosis); cognition-os/00_AI_BRIEFING.md; cognition-os/08_RUNNING.md; cognition-os/RESEARCH_PLAN.md (rows R49-R88); cognition-os/REDTEAM_QUEUE.md (RT-96-RT-122); cognition-os/SEARCH_QUEUE.md (S-P14-S-P70); cognition-os/06_AGENT_LOG.md to 2026-10-21 00:26 EDT; .agent/jobs/2026-10-20-r79-r57-precision-protocol/R57_descriptive_protocol.md).