组会：2026-10-20（UTC）
Group meeting: 2026-10-20 (UTC)
每次刷新把"组会卡片"放在首页最上面：像代理每天向 PI 汇报一次。本页显示最新一天（UTC 日期）的完整内容；更早的日子见组会存档（2026-10-19）。
Every refresh opens the home page with the group-meeting card: as if the agent reported to the PI once a day. This page shows the latest day (UTC date) in full; earlier days are in the meeting archive (2026-10-19).
成果 / Results
Results
- M31 边缘步长有了校正表（R34，已验证）：0.4–1.25 keV、内部 8–10′ 对外部 12–14′、MOS2 同相机 / 曝光匹配——不扣软质子 \(+9.37\pm3.20\) / \(+9.06\pm3.22\) %，名义 SP \(+6.28\pm4.51\) / \(+4.74\pm3.99\) %，测得 M31 模板 \(+5.82\pm4.34\) / \(+4.38\pm3.94\) %。旧数字（R29 的 \(-1.66\) / +1.62 %，RT-M88 的 \(+8.93\) / +9.32 %）因权重与误差处理错误已废弃。已确立A corrected M31 edge-step table exists (R34, verified): 0.4–1.25 keV, inner 8–10′ vs outer 12–14′, MOS2 same-camera / exposure-matched — without soft-proton subtraction \(+9.37\pm3.20\) / \(+9.06\pm3.22\) per cent, nominal SP \(+6.28\pm4.51\) / \(+4.74\pm3.99\) per cent, measured M31 template \(+5.82\pm4.34\) / \(+4.38\pm3.94\) per cent. The old numbers (R29's \(-1.66\) / +1.62 per cent, RT-M88's \(+8.93\) / +9.32 per cent) are deprecated for wrong weights and error treatment. established
- SP 水平无法从外部定住（R33，已验证）：没有任何一条路能给出独立于被评分图像的 4–13 % 每观测安静 SP 归一化；更多数据也不行——系统误差底不随 \(\sqrt{N}\) 下降（R36），把步长压到 1 个百分点按 i.i.d. 外推需要约 140–170 个科学观测。已确立The SP level cannot be pinned externally (R33, verified): no route gives a certifiable 4–13 per cent per-observation quiet-SP normalisation independent of the scored image; more data will not help — the systematic floor does not fall as \(\sqrt{N}\) (R36), and pressing the step to 1 percentage point would need about 140–170 science observations by i.i.d. extrapolation. established
- 重复指向彼此不一致（R37，已验证）：MOS2 重复对 0109270301/0402561501 在 SP 扣除后从 \(+7.9\pm3.8\) % 变到 \(+23.8\pm5.8\) %（约 \(4\sigma\)）；近重复对给出的绝对 SP 精度 \(\ge\) 200 %。"重复对只是上限"的旧说法被推翻。临时Repeat pointings disagree with each other (R37, verified): the MOS2 repeat pair 0109270301/0402561501 moves from \(+7.9\pm3.8\) to \(+23.8\pm5.8\) per cent after SP subtraction (about \(4\sigma\)); near-repeats give an absolute SP precision of \(\ge\) 200 per cent. The old "repeat pairs give only an upper limit" claim is overturned. provisional
- 叠加消不掉共模边缘（R38，已验证）：随机滚动平均不掉共模径向误差——对齐叠加在 12–14′ 处保留 \(\pm\)4 % 天空，任意 N；8′ 偏移 + N=200 只把边缘-only 偏差降到 +0.7 %，而线性项给 \(-1.4\pm1.3\) %。已确立Stacking cannot average out the common-mode edge (R38, verified): random roll does not average out a common-mode radial error — an aligned stack keeps \(\pm\)4 per cent sky at 12–14′ for any N; an 8′ offset with N=200 reduces the edge-only bias to +0.7 per cent while a linear term gives \(-1.4\pm1.3\) per cent. established
- QPB 外模板假设被检验，不充分（R39/R42，已验证）：10–14′ 处 QPB 模板低估 5.8–9.1 % 会产生 +4–6 个百分点的边缘（那里 QPB 约占计数的 35–36 %）；但全 FWC/径向替换给出非正净项，SP 扣除后的 MOS1/同相机/曝光匹配未定义；只剩年代-FWC 漂移转移（同相机 \(+5.08\pm4.35\) %）。QPB 模板未被证明充分。临时The QPB outer-template lead was tested and is not sufficient (R39/R42, verified): a 5.8–9.1 per cent QPB template underestimate at 10–14′ would produce a +4–6 percentage point edge (QPB is about 35–36 per cent of counts there); but the full FWC/radial replacement gives non-positive net terms, leaving SP-subtracted MOS1/samecam/exposure-matched undefined; only the epoch-FWC drift transfer remains (same-camera \(+5.08\pm4.35\) per cent). The QPB template is not shown sufficient. provisional
- 官方 ESAS 耀发图 CCF 自带负像素（R43，已验证）：负值来自 XMM_FLARE_0001 本身——每条 MOS FLAREMAP 行 608,400 像素里有 111k–289k 个负像素；在 M31 足迹内共 186 个计数；截断到 0 只把步长改 \(+0.17\) / \(+0.23\) 个百分点。已确立The official ESAS flare-map CCF carries negative pixels (R43, verified): the negatives come from XMM_FLARE_0001 itself — every MOS FLAREMAP row has 111k–289k negative pixels out of 608,400; 186 counts fall within M31 footprints; clipping at 0 changes the step by only \(+0.17\) / \(+0.23\) percentage points. established
- 目录默认 DR14（D029，PI 决定）与 fast_regions 提速（R21-6，已验证）：有官方目录覆盖时用 4XMM-DR14 屏蔽源，省下每 ObsID 中位数约 60 分钟的 edetect_stack；fast_regions 热缓存 100 个 MOS2 区域 Virgo 16.0 s（0.16 s/区域）。已确立Catalogue default DR14 (D029, PI decision) and fast_regions speed-up (R21-6, verified): mask sources with 4XMM-DR14 whenever the observation is covered, saving edetect_stack at a median of about 60 minutes per ObsID; fast_regions extracts 100 MOS2 regions in 16.0 s on Virgo with a warm cache (0.16 s per region). established
来源：Source: cognition-os/08_RUNNING.md revision 22–28（2026-10-20 02:35–08:43 EDT），revisions 22–28 (2026-10-20 02:35–08:43 EDT), MosaicSelfCal/RESEARCH_PLAN.md R24/R33–R38/R39/R42/R43、R21-6、T-cat 行，the R24/R33–R38/R39/R42/R43, R21-6, T-cat rows, cognition-os/04_DECISION_LOG.md D029D029
工作 / Work
Work
- 答完并验证：R24（测得软 SP 模板 \(\approx\) 耀发图）、R26（CCD2 条带掩膜）、R27（冻结锚）、R29/R30（测得模板不解决边缘）、R31（验收图）、R32（安静模板不可迁移）、R33（外部 SP 归一化不可达）、R34（校正表）、R35（MOS1 残差一半是泊松）、R36（功效/档案）、R37（重复对）、R38（叠加代价）、R39（QPB 铅）、R40（文献策略）、R41（拼接交付物）、R42（QPB 不充分）、R43（CCF 负像素）、R21-6（提速）、目录工具任务（sol 为主，codexnim 复核）。红队 RT-M90–RT-M102 全部三分派并归档；文献检索 S15–S24 完成。Answered and verified: R24 (measured soft SP template \(\approx\) flare map), R26 (CCD2 strip mask), R27 (frozen anchor), R29/R30 (the measured template does not solve the edge), R31 (acceptance map), R32 (quiet-template transfer not certifiable), R33 (external SP normalisation unreachable), R34 (corrected table), R35 (half the MOS1 residual is Poisson), R36 (power/archive), R37 (repeat pairs), R38 (stacking cost), R39 (QPB lead), R40 (literature strategies), R41 (mosaic deliverable), R42 (QPB not sufficient), R43 (CCF negative pixels), R21-6 (speed-up), the catalogue tool task (sol main, codexnim re-check). Red teams RT-M90–RT-M102 all triaged and archived; literature searches S15–S24 done.
- 派出：R42、R43、R44（FWC 库存）、R45（边缘随能段）、R21-7（持久 prepare 缓存，PI 追加）、手册刷新 2（Hermes，本页）；专家顾问 astra（DXC 首次，02:35 EDT 启动）。Dispatched: R42, R43, R44 (FWC inventory), R45 (edge vs energy), R21-7 (persistent prepare cache, PI addendum), handbook refresh 2 (Hermes, this page); the expert advisor astra (first for DXC, launched 02:35 EDT).
- 工具：esas dev/catalogue-default（3703be6，34 测试）与 dev/fast-regions（042bb39）待合并评审；本日无部署。Tools: esas dev/catalogue-default (3703be6, 34 tests) and dev/fast-regions (042bb39) await merge review; no deploy this day.
来源：Source: cognition-os/06_AGENT_LOG.md 2026-10-20 各行，the 2026-10-20 rows, cognition-os/08_RUNNING.md §3§3
问题 / Problems
Problems
- 专家顾问判定：三步目标未达成（astra，rev 23 三分）：验收图（R31）没有一个"步骤 \(\times\) 轴"单元完全达标——第 1 步缺 GTI \(\times\) 掩膜 S/N 预算，第 2 步只有条件性证据，第 3 步的 SP 扣除残差随分解/模板/合并方式变号。行动：R31–R38 按其建议展开。The expert advisor's verdict: the three-step goal is not reached (astra, triaged in rev 23): in the acceptance map (R31) no step \(\times\) axis cell is fully qualified — step 1 lacks a GTI \(\times\) mask S/N budget, step 2 is conditional only, and step 3's SP-subtracted residual flips sign with the decomposition/template/pooling. Action: R31–R38 follow its recommendations.
- R29 的"边缘已解决"被撤回（RT-M98）：测得模板的 \(-1.66\)/+1.62 % 用错历史权重；校正后为 \(+5.82\pm4.34\) / \(+4.38\pm3.94\) %（与 R30 独立一致）。行动：全站把旧数字标注废弃，改用 R34 校正表。R29's "edge solved" is withdrawn (RT-M98): the measured template's \(-1.66\)/+1.62 per cent used the wrong historical weights; corrected they are \(+5.82\pm4.34\) / \(+4.38\pm3.94\) per cent (independently matching R30). Action: the old numbers are marked deprecated across the site and replaced by the R34 corrected table.
- 重复指向推翻"上限"旧说（R37）：重复对在 SP 扣除后不一致（约 \(4\sigma\)），近重复对的绝对 SP 精度 \(\ge\) 200 %。行动：工作台"未解决问题"表相应改写。Repeat pointings overturn the old "upper limit" reading (R37): repeat pairs disagree after SP subtraction (about \(4\sigma\)), and near-repeats give an absolute SP precision of \(\ge\) 200 per cent. Action: the workbench open-problems row is rewritten accordingly.
- QPB 假设落空（R42）：预先声明的充分性规则不满足；SP 扣除后的 MOS1/同相机/曝光匹配未定义。行动：R44（FWC 库存）继续，能段签名（R45）另开一线。The QPB hypothesis collapses (R42): the pre-stated sufficiency rule is not met; SP-subtracted MOS1/samecam/exposure-matched is undefined. Action: R44 (FWC inventory) continues, and the energy signature (R45) opens a separate line.
- 代理自己的错误：R23 简报把 XRT3 写成 MOS2 的镜（XRT3=pn，MOS2=XRT2）；已在记录中更正并在 R26 里复核 XRT1 vs XRT2（晒图相等）。行动：全站不再引用"XRT1/XRT3 相等"这一错误前提。The agent's own error: the R23 brief named XRT3 as MOS2's mirror (XRT3=pn, MOS2=XRT2); corrected in the record, and R26 re-checked XRT1 vs XRT2 (vignetting equal). Action: the site no longer repeats the wrong "XRT1/XRT3 equal" premise.
来源：Source: cognition-os/08_RUNNING.md revision 23–28，revisions 23–28, MosaicSelfCal/RESEARCH_PLAN.md R23/R29/R31/R37/R42 行，the R23/R29/R31/R37/R42 rows, cognition-os/REDTEAM_QUEUE.md RT-M98 条目the RT-M98 entry
待 PI 决定 / For the PI
For the PI
- D028 边缘步长归因规则（草稿，待采纳）：外部 SP 约束到位前不得确认任何归因；99.73 % 置信集；采纳前须封存输入、估计器代码与未见的确认输出编号。D028 edge-step attribution rule (draft, pending adoption): no attribution may be confirmed before an external SP constraint exists; 99.73 per cent confidence sets; before adoption the inputs, estimator code and the identifiers of the unseen confirmation outputs must be sealed.
- R21-7 PI 追加要求（一次性持久 prepare 缓存）已在跑；pn 通道与两个 held-out 限额案例仍开。R21-7 PI addendum (one-off persistent prepare cache) is running; the pn channel and two held-out limit cases remain open.
来源：Source: cognition-os/04_DECISION_LOG.md D028、MosaicSelfCal/RESEARCH_PLAN.md R21-7 行the R21-7 row
DiffuseXrayCalibration：把 XMM 的弥漫辐射测量校准到可引用
DiffuseXrayCalibration: making XMM diffuse-emission measurements quotable
XMM-Newton 的 EPIC 相机看弥漫辐射时，隔着两种仪器自身的背景：准静态粒子背景（QPB）与软质子（SP）。本项目校准 ESAS 减数据 + Sherpa 拟合这条流水线，使同一天空区域在任何观测、任何离轴角、任何 SP/QPB 状态下给出同一个答案；做不到，就给出这个差异的测量值——系统误差预算。最显眼的症状是每个指向在拼接图边缘变亮（edge brightening）。为什么要做，见为什么存在一页。
XMM-Newton's EPIC cameras see diffuse emission through two backgrounds of the instrument's own making: the quiescent particle background (QPB) and soft protons (SP). This project calibrates the pipeline that removes them — ESAS reduction plus Sherpa fitting — so that one piece of sky gives the same answer in every observation, at every off-axis angle, and under every SP/QPB state; where it cannot, it delivers the measured difference itself: a systematic budget. The visible symptom is that every pointing brightens toward the edge of a mosaic (edge brightening). Why this matters is on the why the project exists page.
这一页回答四个问题：这是什么（本段）；现在到哪里了（目标表）；什么还没解决（未解决问题表）；下一步做什么（1–4 周计划与如何用手继续）。
This page answers four questions: what this is (this paragraph); where it stands (the objectives table); what is unsolved (the open-problems table); what to do next (the 1–4 week plan and how to continue by hand).
来源：Source: cognition-os/00_AI_BRIEFING.md, cognition-os/02_CURRENT_QUESTION.md
最新骨架：一条流水线，三个考题
Latest skeleton: one pipeline, three exam steps
来源：Source: cognition-os/02_CURRENT_QUESTION.md, MosaicSelfCal/BENCHMARK.md, cognition-os/tool_iteration.md
目标表：PI 的三步目标与基准测试目标
Objectives: the PI's three-step goal and the benchmark goals
| 目标Objective | 状态Status | 当前最佳数字Best current number | 趋势Trend |
|---|---|---|---|
| 第 1 步：抽谱（region + spectra）——高信噪比、低 SP、低未分辨点源/展源残差Step 1: regions and spectra — high S/N, low SP, low unresolved point/extended residual | 进行中in progress | CDFS P0：\(\tau/\bar{I}=0.246\)，\(\sigma/\bar{I}=0.245\)CDFS P0: \(\tau/\bar{I}=0.246\), \(\sigma/\bar{I}=0.245\) | → 掩膜规则已定（D012/D021），分数尚未收敛→ mask rule decided (D012/D021); score not yet converged |
| 第 2 步：拟合（fit sky 和 Src）——同一点指向内跨观测稳定；跨离轴角稳定Step 2: fit sky and Src — stable across observations of the same pointing, and across off-axis angle | 进行中in progress | M31 配对斜率跨 SP 分解横跨 \(-0.73\) 到 \(+0.91\) %/arcmin（仅 QPB \(+1.51\pm0.17\)；自由谱形 \(-0.73\pm0.29\)；时间切片颜色 \(+0.91\)；颜色先验 \(+0.01\pm0.26\)）；联合多环带（耀发模板）\(-0.161\pm0.262\)，与零一致M31 pair slope spans \(-0.73\) to \(+0.91\) per cent per arcmin across SP decompositions (QPB only \(+1.51\pm0.17\); free shape \(-0.73\pm0.29\); time-slice colour \(+0.91\); colour prior \(+0.01\pm0.26\)); joint multi-annulus with the flare template \(-0.161\pm0.262\), consistent with zero | ↔ 斜率由 SP 分解决定；重复指向在 SP 扣除后彼此不一致（R37，约 \(4\sigma\)）↔ the slope is decided by the SP decomposition; repeat pointings disagree with each other after SP subtraction (R37, about \(4\sigma\)) |
| 第 3 步：成像（image，去掉 QPB 与 SP）——同指向叠加一致；不同指向拼接看不到 vignettingStep 3: images with QPB and SP removed — same-pointing stacks agree; mosaics show no visible vignetting | 未解决open | M31 边缘步长（MOS2，校正估计器 R34）：不扣 SP \(+9.37\pm3.20\) % 同相机、\(+9.06\pm3.22\) % 曝光匹配；名义 SP \(+6.28\pm4.51\) / \(+4.74\pm3.99\)；测得 M31 模板 \(+5.82\pm4.34\) / \(+4.38\pm3.94\)M31 edge step, MOS2, corrected estimator (R34): no-SP subtraction \(+9.37\pm3.20\) per cent same-camera, \(+9.06\pm3.22\) per cent exposure-matched; nominal SP \(+6.28\pm4.51\) / \(+4.74\pm3.99\); measured M31 template \(+5.82\pm4.34\) / \(+4.38\pm3.94\) | ↔ 不扣 SP 的步长稳健；扣 SP 后的残差不稳健，且已测的 QPB 模板修正都消不掉它（R42）↔ the unsubtracted step is robust; the SP-subtracted residual is not, and no tested QPB-template fix removes it (R42) |
| 基准测试基线：同一固定天域跨 ObsID 的 \(\tau/\bar{I}\) 与离轴斜率（CDFS 与 M31）Benchmark baseline: between-ObsID \(\tau/\bar{I}\) and off-axis slope for one fixed sky cell (CDFS and M31) | 已建立established | CDFS P0：\(\tau/\bar{I}=0.246\)，斜率 \(+0.152/10'\)；M31 P0：\(\tau/\bar{I}=0.14\)，斜率 \(+0.25/10'\)CDFS P0: \(\tau/\bar{I}=0.246\), slope \(+0.152/10'\); M31 P0: \(\tau/\bar{I}=0.14\), slope \(+0.25/10'\) | → 这是待改进的基线，不是成果→ this is the baseline to improve, not a result |
| 考试集（exam v4 r240）：7 个 problem、18 个 ObsID、45 个 case 被减出来并审计Exam set (v4 r240): 7 problems, 18 ObsIDs, 45 cases reduced and audited | 完成done | 40 PASS，5 EXCLUDED，0 FAIL40 PASS, 5 EXCLUDED, 0 FAIL | → 冻结集（D020），后续分数都在它上面算→ frozen set (D020); every later score is computed on it |
| 快速抽谱（R21，PI 2026-10-19 11:23 EDT 的新目标）：对任意区域快速抽出谱与响应（RMF、ARF、QPB）Fast extraction (R21, the PI's new goal of 2026-10-19 11:23 EDT): spectra and responses (RMF, ARF, QPB) for arbitrary regions, fast | 进行中in progress | 热缓存 0.16 s/区域（Virgo 上 100 个 MOS2 区域 16.0 s）；冷/热输出一致；原生 prepare 的 ARF+RMF 每相机约 930–990 swarm cache 0.16 s per region on Virgo (100 MOS2 regions in 16.0 s); outputs identical cold/warm; native prepare ARF+RMF ~930–990 s per camera | → R21-2 对齐已过（BACKSCAL 6/6 在 0.1 % 内，ARF 6/6 在 1 % 内）；R21-6 提速完成，R21-7 持久缓存在跑→ R21-2 parity passed (BACKSCAL 6/6 within 0.1 %, ARF 6/6 within 1 %); R21-6 speed-up done, R21-7 persistent cache running |
来源：Source: MosaicSelfCal/BENCHMARK.md, cognition-os/02_CURRENT_QUESTION.md, MosaicSelfCal/results/cdfs_pipeline_comparison_v1.csv, MosaicSelfCal/results/exam_v4_r240/SUMMARY.md, cognition-os/tool_iteration.md 趋势箭头：→ 未收敛；↔ 两个估计矛盾；"持平"= 配对区间含 0。Trend arrows: → not converged; ↔ two estimates disagree; "No significant difference" = the paired interval contains 0.
未解决的问题
Open problems
| 问题Problem | 为何重要Why it matters | 当前最佳假设Best hypothesis | 下一个检验Next test | 负责人Owner | 状态Status |
|---|---|---|---|---|---|
| 边缘增亮（edge brightening）到底是 SP、QPB，还是仪器响应？Is edge brightening soft protons, the QPB, or the instrument response? | 这是 mosaics 里最显眼的失败，也是 PI 三步目标的第 3 步It is the most visible failure in mosaics and step 3 of the PI goal | 步长稳健、归因未定，且数据不足以定归因：校正估计器下不扣 SP 时 \(+9.37\pm3.20\) %（同相机）跨相机一致；扣掉后落在 \(+4.4\) 到 \(+6.3\pm4\) %（R34）。SP 归一化无法从外部定住（R33：最好也只 13 %，需 4–13 %），重复指向彼此不一致（R37），QPB 模板修正不充分（R42）——归因是数据受限的，不是分析疏漏The step is robust, its attribution is not, and the data cannot pin it: under the corrected estimator the unsubtracted step is \(+9.37\pm3.20\) per cent (same-camera), consistent across cameras; after SP subtraction it sits at \(+4.4\) to \(+6.3\pm4\) per cent (R34). The SP normalisation cannot be pinned externally (R33: best route gives 13 %, 4–13 % needed), repeat pointings disagree (R37), and the QPB-template fix is not sufficient (R42) — the attribution is data-limited, not an analysis slip | R45：校正 M31 边缘步长的能段签名（0.4–0.7 一直到 5–7 keV）对照 SP/QPB/相机项的预测；R44：FWC 数据按相机/年代清点，决定 QPB 假设还能不能往前走；D028 归因规则（草稿）待采纳后按它打分R45: the energy signature of the corrected M31 edge step (0.4–0.7 up to 5–7 keV) against the SP/QPB/camera-term predictions; R44: an FWC inventory per camera and epoch, deciding whether the QPB hypothesis can move further; the D028 attribution rule (draft) once adopted, to score against | R45 / R44 | 运行中running |
| SP 在探测器上的空间分布是什么形状？What is the spatial distribution of SP across the detector? | 拟合用的模型 \( \text{rsp\_flat}(A_{\text{sky}}\cdot\text{SP}) \) 默认 SP 在探测器上是均匀的；如果不是，所有 SP 扣除都带系统误差The fit model \( \text{rsp\_flat}(A_{\text{sky}}\cdot\text{SP}) \) assumes SP is uniform across the detector; if it is not, every SP subtraction carries a systematic | 已测（R24）：从耀发时段直接测得的 M31 软带 SP 空间模板与 ESAS 耀发图一致；R20y 联合多环带斜率 \(+0.011\pm0.261\) %/arcmin（名义 \(-0.161\)），与零一致。闭合证据：R24 的 31/31 安静 PHA 重放精确一致，MOS2−MOS1 超出名义预测 \(+0.023\pm0.014\)（1.7\(\sigma\)）Measured (R24): the M31 soft-band SP spatial template taken directly from flare intervals matches the ESAS flare map; the R20y joint multi-annulus slope is \(+0.011\pm0.261\) per cent per arcmin (nominal \(-0.161\)), consistent with zero. Closing evidence: R24's 31/31 quiet-PHA replay is exact, and MOS2−MOS1 exceeds the nominal prediction by \(+0.023\pm0.014\) (1.7\(\sigma\)) | 已闭合：软带模板已由 R24 从耀发时段直接测得。剩下的问题不再是形状，而是每观测的归一化（见上一行）Closed: the soft-band template was measured directly from flare intervals by R24. What remains is not the shape but the per-observation normalisation (previous row) | R24 | 已闭合closed |
| MOS1/MOS2 的 SP 归一比在 2012 年发生台阶（\(0.68\to1.47\)），原因是什么？The MOS1/MOS2 SP-normalisation ratio steps in 2012 (\(0.68\to1.47\)); what causes it? | 任何跨 2012-12 合并 SP 先验或比较都是错的；这也决定了 SP 扣除能否跨年代使用Any SP prior or comparison pooled across 2012-12 is wrong, and it decides whether an SP subtraction can be used across epochs | 没有单一原因站得住：面积效应被排除；掩膜足迹、角区模板陈旧、硬带 QPB 标度都被逐一排除；自由 SP 谱形下看不到 2012 台阶（R20s：1.61\(\sigma\)）No single cause stands: the area effect is refuted; the mask footprint, stale corner templates and the hard-band QPB scale are excluded one by one; with free SP shapes no 2012 step is seen (R20s: 1.61\(\sigma\)) | R20w 的软带 SP 谱形与径向模板直接测量——它同时是 R20v 边缘数字的输入R20w's direct measurement of the soft-band SP shape and radial template — which is also the input to R20v's edge numbers | R20w | 运行中running |
| 重复指向能否测出 SP 的绝对水平？Can repeat pointings measure the absolute SP level? | SP 水平是第 3 步成像的关键输入；测不出来就无法给 SP 扣除定标The SP level is the key input to step 3; without it the SP subtraction cannot be normalised | 不能，而且重复对彼此不一致（R37）：0109270301/0402561501 的 MOS2 在 SP 扣除后从 \(+7.9\pm3.8\) % 变到 \(+23.8\pm5.8\) %（约 \(4\sigma\)）；近重复对给出的绝对 SP 精度 \(\ge\) 200 %——比需要的 4–13 % 差一个量级以上No, and repeat pairs disagree with each other (R37): the MOS2 pair 0109270301/0402561501 moves from \(+7.9\pm3.8\) to \(+23.8\pm5.8\) per cent after SP subtraction (about \(4\sigma\)); near-repeats give an absolute SP precision of \(\ge\) 200 per cent — an order of magnitude worse than the 4–13 per cent needed | R36 的结论已把这条路标为数据受限：1 个百分点的步长需要约 140–170 个科学观测（i.i.d. 外推），且系统底不随 \(\sqrt{N}\) 下降；改走外部定标（D028 的外部 SP 约束门）R36 has marked this route data-limited: a 1-percentage-point step needs about 140–170 science observations (i.i.d. extrapolation), and the systematic floor does not fall as \(\sqrt{N}\); the route moves to external calibration (D028's external-SP gate) | D028 （草稿）(draft) | 未解决open |
| 叠加（stacking）能否把边缘误差平均掉？Can stacking average out the edge error? | CGM 叠加是本项目下游的主要用途；若共模误差消不掉，所有叠加晕的径向轮廓都带 \(\pm\)4 % 的系统Stacking CGM observations is the main downstream use; if a common-mode error survives, every stacked halo profile carries a \(\pm\)4 per cent systematic | 不能（R38）：随机滚动平均不掉共模径向误差——对齐叠加在 12–14′ 处保留 \(\pm\)4 % 天空，任意 N；8′ 偏移 + N=200 只把边缘-only 偏差降到 +0.7 %，但线性项给 \(-1.4\pm1.3\) %No (R38): random roll does not average out a common-mode radial error — an aligned stack keeps \(\pm\)4 per cent sky at 12–14′ for any N; an 8′ offset with N=200 reduces the edge-only bias to +0.7 per cent but a linear term gives \(-1.4\pm1.3\) per cent | 已到结论：没有方案能把全体天空降到 1 % 以内；叠加前必须先修边缘，或把 \(\pm\)4 % 写进误差预算Concluded: no scheme brings the whole sky within 1 per cent; fix the edge before stacking, or carry the \(\pm\)4 per cent in the budget | R38 | 已回答answered |
| QPB 空间模板在 10–14′ 是否被低估？Is the QPB spatial template underestimated at 10–14′? | 这是 R39 提出的非 SP 机制候选：那里 QPB 约占计数的 35–36 %，模板低估 5.8–9.1 % 就会产生 +4–6 个百分点的边缘The non-SP candidate raised by R39: QPB is about 35–36 per cent of counts there, and a 5.8–9.1 per cent template underestimate would produce a +4–6 percentage point edge | 不充分（R42）：预先声明的规则（区间含 0 且修正误差 \(<\)2 个百分点）不满足——全 FWC/径向替换给出非正净项（0690600401 MOS1），SP 扣除后的 MOS1/同相机/曝光匹配未定义；只剩年代-FWC 漂移转移（同相机 \(+5.08\pm4.35\) %）Not sufficient (R42): the pre-stated rule (interval contains 0 and correction error \(<\)2 pp) is not met — the full FWC/radial replacement gives non-positive net terms (0690600401 MOS1), leaving SP-subtracted MOS1/samecam/exposure-matched undefined; only the epoch-FWC drift transfer remains (same-camera \(+5.08\pm4.35\) per cent) | R44（FWC 库存，运行中）决定数据是否够；R45 的能段签名区分 QPB 与 SP/相机项R44 (FWC inventory, running) decides whether the data exist; R45's energy signature separates QPB from the SP/camera terms | R44 / R45 | 检验中testing |
| pn 相机的 QPB 为什么被高估？Why is the pn camera's QPB over-normalised? | pn 目前只能做交叉检查（T01）；修不好就不能升级为主要相机pn is a cross-check only (T01); until it is fixed it cannot be promoted | 不是角区光子泄漏（OOT 只占 \(2.5\ \%\)，边缘 \(0.5\ \%\)）；更像 FWC 模板迁移，最佳修正是 FoV 12-14 keV 归一化Not corner photon leakage (OOT \(2.5\ \%\), edge \(0.5\ \%\)); likelier an FWC template transfer, best corrected by the FoV 12-14 keV normalisation | R5b：pn 减数据 + pn/MOS 联合拟合，held-out 检验 pn QPB 标度R5b: pn reduction plus joint pn/MOS fits, with a held-out test of the pn QPB scale | R5b | 检验中testing |
| 暗天区里 SP 与天空的简并如何打破？How is the SP-sky degeneracy broken in faint fields? | WCDFS 是最暗的检验场；在那里谱拟合给出的 SP 归一不可靠，成像又过度扣除WCDFS is the faintest testbed; there the spectral SP norm is unreliable and the image over-subtracts | 分组方式本身就是一个系统误差（G1 把 2-12 keV 打进一个 bin，天空差 \(18\ \%\)）；用 G2 式边界The grouping choice is itself a systematic (G1 packs 2-12 keV into one bin, moving the sky by \(18\ \%\)); use G2-style boundaries | R3c：联合似然比检验（天空共享 vs 每观测自由，SP 每观测自由）R3c: joint likelihood-ratio test (sky shared vs free per observation, SP free per observation) | R3c | 部分完成partly done |
| SAS 22.1 后端能否用于生产？Can the SAS 22.1 backend be used for production? | SAS 20 与 22.1 的 ARF 差 \(+7\) 到 \(+22\ \%\)；用错后端会让所有天空水平偏移The SAS 20 and 22.1 ARFs differ by \(+7\) to \(+22\ \%\); the wrong backend shifts every sky level | 差异可归因：ARF 来自 arfgen 开关，QPB 来自 FWC 来源 + MOS1 自建角区；但 pn 端到端与可复现性声明还没过The differences are attributable (ARF from arfgen switches, QPB from the FWC source plus MOS1's own corner), but pn end-to-end and the reproducibility statement have not passed | R15 / esas v1.8.0：pn 端到端、ANOMFL、可复现性声明，然后 PI 签字R15 / esas v1.8.0: pn end to end, ANOMFL, a reproducibility statement, then PI sign-off | R15 | 检验中testing |
来源：Source: cognition-os/REDTEAM_QUEUE.md, MosaicSelfCal/RESEARCH_PLAN.md, cognition-os/08_RUNNING.md
未来 1–4 周
The next 1–4 weeks
计划顺序照 MosaicSelfCal/RESEARCH_PLAN.md 的 SP 路线图（PI 2026-10-17）与 cognition-os/08_RUNNING.md §7 的建议方向。每条后面是它要回答的问题。
The order follows the SP roadmap in MosaicSelfCal/RESEARCH_PLAN.md (PI 2026-10-17) and the proposed directions in cognition-os/08_RUNNING.md §7. Each item names the question it answers.
| 时间When | 做什么What | 回答什么问题Question answered | 状态State |
|---|---|---|---|
| 现在进行in flight | R45 边缘随能段变化（sol）；R44 FWC 库存（codexnim）；R21-7 持久 prepare 缓存（sol，Virgo）；RT-M101 复核 R39（hermes）R45 edge versus energy (sol); R44 FWC inventory (codexnim); R21-7 persistent prepare cache (sol on Virgo); RT-M101 re-checking R39 (hermes) | 边缘步长随能段怎么变（区分 SP、QPB 与相机项）；QPB 假设的数据是否够；快速抽谱能否把准备时间压到一次性How the edge step varies with energy (separating SP, QPB and camera terms); whether the data for the QPB hypothesis exist; whether fast extraction can make the preparation a one-off | 运行中running |
| 接下来next | 归因规则落地：D028（草稿）采纳后按它给 M31 MOS2 边缘步长打分——外部 SP 约束、99.73 % 置信集、预登记的输出编号Land the attribution rule: once D028 (draft) is adopted, score the M31 MOS2 edge step by it — external SP constraint, 99.73 per cent confidence sets, pre-registered output identifiers | 在没有外部 SP 约束的情况下，归因只能判为"未识别"；规则先行的价值就在此Without an external SP constraint the attribution can only be judged "unidentified"; that is the value of the rule-first approach | 待定pending |
| 并行in parallel | R21 快速抽谱家族（R21-2/R21-4/R21-6 已完成，R21-7 在跑，pn 未完成）；R5b pn 减数据与联合拟合；R15 / esas v1.8.0 READY 门the R21 fast-extraction family (R21-2/R21-4/R21-6 done, R21-7 running, pn not done); R5b pn reduction and joint fits; the R15 / esas v1.8.0 READY gate | 任意区域快速抽谱（PI 当前目标）；pn 能否成为第二台主相机；SAS 22.1 能否用于生产fast spectra for arbitrary regions (the PI's current goal); can pn become a second primary camera; can SAS 22.1 be used for production | 运行中running |
| 2–4 周2–4 weeks | 工具发布：edgestep 包、xsherpa SP 响应、esas SP 成像模板；然后 DXC 用固定版本重测Tool releases: the edgestep package, the xsherpa SP response, the ESAS SP image template; then a DXC re-test pinned to the version | 修正能否落到工具里并被独立复现Whether the fixes land in the tools and are independently reproduced | 待定pending |
| 验收标准acceptance | 高 SP / 低 SP 时间桶给同一面亮度；M31/LMC/WCDFS 配对斜率归零；mosaic 看不到 vignettingHigh- and low-SP time bins of one pointing give the same surface brightness; M31/LMC/WCDFS pair slopes zero; mosaics show no visible vignetting | PI 三步目标是否达成Whether the PI's three-step goal is met | 未达成not met |
来源：Source: MosaicSelfCal/RESEARCH_PLAN.md, cognition-os/08_RUNNING.md, cognition-os/tool_iteration.md
本次更新 / 数据截至
What's new / Data as of
本次更新What's new
2026-10-20 09:25 EDT：刷新 2（H-DXC-2）。把内容推进到 cognition-os/08_RUNNING.md revision 28（2026-10-20 08:43 EDT），并新增 2026-10-20 的组会段。要点：
2026-10-20 09:25 EDT: refresh 2 (H-DXC-2).The content is brought up to cognition-os/08_RUNNING.md revision 28 (2026-10-20 08:43 EDT), and a 2026-10-20 group-meeting block is added. The essentials:
- M31 边缘步长有了校正表（R34，已验证）：0.4–1.25 keV、内部 8–10′ 对外部 12–14′、MOS2——不扣软质子 \(+9.37\pm3.20\) %（同相机）、\(+9.06\pm3.22\) %（曝光匹配）；名义 SP 扣除后 \(+6.28\pm4.51\) / \(+4.74\pm3.99\)；测得模板后 \(+5.82\pm4.34\) / \(+4.38\pm3.94\)。旧的 \(+8.9\pm2.4\) / \(+9.3\pm2.4\)（RT-M88）与 \(-1.66\)/\(+1.62\)（R29）废弃
- SP 归一化无法从外部定住（R32/R33/R36）：没有任何路线能给到每观测 4–13 % 的可认证精度；按 i.i.d. 外推，1 个百分点的步长需要约 140–170 个科学观测，且系统底不随 \(\sqrt{N}\) 下降
- 重复指向彼此不一致（R37）：0109270301/0402561501 的 MOS2 在 SP 扣除后从 \(+7.9\pm3.8\) % 变到 \(+23.8\pm5.8\) %（约 \(4\sigma\)）；近重复对给出的绝对 SP 精度 \(\ge\) 200 %
- 叠加保留共模边缘（R38）：随机滚动平均不掉共模径向误差——对齐叠加在 12–14′ 处保留 \(\pm\)4 % 天空，任意 N；没有方案能把全体天空降到 1 % 以内
- QPB 外模板假设被检验，不充分（R39/R42）：预先声明的规则（区间含 0 且修正误差 \(<\)2 个百分点）不满足；只剩年代-FWC 漂移转移（同相机 \(+5.08\pm4.35\) %）
- 官方 ESAS 耀发图 CCF 里有负像素（R43）：每行 MOS FLAREMAP 有 111k–289k 个负像素（共 608,400）；截断到 0 只改变步长 \(+0.17\) / \(+0.23\) 个百分点
- 目录默认 DR14（D029）：4XMM-DR14 按观测逐个使用，DR14s 仅在点名理由时；edetect 每观测省 39–86 分钟（中位 60）；34 项测试通过
- 快速抽谱提速（R21-6）：热缓存 0.16 s/区域（Virgo 上 100 个 MOS2 区域 16.0 s），冷/热输出一致；R21-7 持久缓存在跑；专家顾问判定三步目标未达成
- The M31 edge step has a corrected table (R34, verified): 0.4–1.25 keV, inner 8–10′ vs outer 12–14′, MOS2 — unsubtracted \(+9.37\pm3.20\) per cent (same-camera), \(+9.06\pm3.22\) per cent (exposure-matched); after nominal SP \(+6.28\pm4.51\) / \(+4.74\pm3.99\); after the measured template \(+5.82\pm4.34\) / \(+4.38\pm3.94\). The old \(+8.9\pm2.4\) / \(+9.3\pm2.4\) (RT-M88) and \(-1.66\)/\(+1.62\) (R29) are deprecated
- The SP normalisation cannot be pinned externally (R32/R33/R36): no route gives a certifiable 4–13 per cent per-observation precision; by i.i.d. extrapolation a 1-percentage-point step needs about 140–170 science observations, and the systematic floor does not fall as \(\sqrt{N}\)
- Repeat pointings disagree with each other (R37): the MOS2 pair 0109270301/0402561501 moves from \(+7.9\pm3.8\) to \(+23.8\pm5.8\) per cent after SP subtraction (about \(4\sigma\)); near-repeats give an absolute SP precision of \(\ge\) 200 per cent
- Stacks keep the common-mode edge (R38): random roll does not average out a common-mode radial error — an aligned stack keeps \(\pm\)4 per cent sky at 12–14′ for any N; no scheme brings the whole sky within 1 per cent
- The QPB outer-template lead was tested and is not sufficient (R39/R42): the pre-stated rule (interval contains 0 and correction error \(<\)2 percentage points) is not met; only the epoch-FWC drift transfer remains (same-camera \(+5.08\pm4.35\) per cent)
- The official ESAS flare-map CCF has negative pixels (R43): every MOS FLAREMAP row has 111k–289k negative pixels of 608,400; clipping at 0 changes the step by only \(+0.17\) / \(+0.23\) percentage points
- Catalogue default DR14 (D029): 4XMM-DR14 per observation by default, DR14s only with a named reason; edetect saves 39–86 minutes (median 60) per ObsID; 34 tests pass
- Fast extraction speed-up (R21-6): warm cache 0.16 s per region (100 MOS2 regions in 16.0 s on Virgo), outputs identical cold/warm; R21-7 persistent cache running; the expert advisor judged the three-step goal not reached
本页所有数字取自下方列出的记录文件；没有一项是本手册自己算出来的。
Every number on this page is read from the record files listed below; none is computed by this handbook.
数据截至Data as of
项目状态：2026-10-20 08:43 EDT（cognition-os/08_RUNNING.md revision 28）。
Project state: 2026-10-20 08:43 EDT (cognition-os/08_RUNNING.md revision 28).
本手册读到的最新科学记录：2026-10-20 08:43 EDT（rev 28：校正估计器下不扣软质子时 M31 边缘步长 \(+9.37\pm3.20\) %（同相机）、\(+9.06\pm3.22\) %（曝光匹配）；扣掉 SP 后落在 \(+4.4\) 到 \(+6.3\pm4\) %；SP 归一化无法从外部定住，重复指向彼此不一致，QPB 模板修正不充分——归因是数据受限的）。
The newest science record this handbook read: 2026-10-20 08:43 EDT (rev 28: under the corrected estimator the unsubtracted M31 edge step is \(+9.37\pm3.20\) per cent same-camera and \(+9.06\pm3.22\) per cent exposure-matched; after SP subtraction it sits at \(+4.4\) to \(+6.3\pm4\) per cent; the SP normalisation cannot be pinned externally, repeat pointings disagree, and the QPB-template fix is not sufficient — the attribution is data-limited).
更晚的记录（评审、决定、新结果）以它们为准；本页在下次刷新时更新。
Later records (reviews, decisions, new results) take precedence; this page is updated at the next refresh.