M31 CGM 工作台M31 CGM / New-ANGELS III
目录Contents

组会（最新一天：2026-10-16）Group meeting (latest day: 2026-10-16)

成果：A1d 已验收（14 场 × 117 点 = 1638/1638，0 个拟合失败），A2 基线在它上面重算——集中核（r_c ≤ 8 kpc）20 kpc 上限 0.12–0.16（β 0.4）／0.075–0.10（β 0.5）／0.049–0.071（β 0.6），排除 Zhang 名义中值 ≳4 倍；红队 RT-13 复核：算术与边界阈值用法成立，但上限主要靠一个视场（去掉它放宽约一半），措辞改为「全样本 ≳4 倍；去掉约束最强的视场仍 ≥3 倍」；报出的值偏保守（插值与流量换算再紧 2–23%）。红队 RT-08：「上限已校准良好」不成立——toy-MC 只在简化模型下有效，Δstat = 2.71 本身形式正确但尚未用能谱模拟校准（E028 运行中）。外场流量随观测年代下降（2009 年及以前 1.05 ± 0.08 vs 2010 年以后 0.55 ± 0.09，约 2σ）——原因未定，「怀疑探测器老化」已撤回（公布的 MOS 校准只有几个百分点，实测 ARF 软端只变 +1.6%/−3.7%），1.3 keV 特征是一条自由的 1.2–1.4 keV 高斯残差而非仪器线；外/内比与基于外场的极限改标「受观测年代混杂」。红队 RT-11（阻断级）：E026 的 runner 会让全部 28 个场失败，修正单已备好，通过前不批处理。E027：预注册修订 v2（场间 τ 只作描述；主通道改为软质子归一化与软波段面亮度），覆盖收益改为「约 4 个有用 + 5 个边缘指向」，203 个 GTI 文件已验收。工作：A1d 完成；E026 修正中；E028 stage 0、RT-14-T1、RT-15、RT-16 运行中。问题：Virgo ssh 中断 06:16Z–~08:50Z（已自行恢复）；我用 200 个玩具覆盖了 RT-08 的 1000 玩具 CSV（已确定性恢复）；实时条目 L006 的「探测器老化」断言已撤回改写；RT-13 简报把距离写成 776 kpc（项目用 785 kpc）。待 PI 决定：无。

Results: A1d is accepted (14 fields × 117 points = 1638/1638, 0 failed fits) and the A2 baseline is recomputed on it — the compact-core (r_c ≤ 8 kpc) limit at 20 kpc is 0.12–0.16 (β 0.4) / 0.075–0.10 (β 0.5) / 0.049–0.071 (β 0.6), excluding Zhang's nominal median by ≳4×; red team RT-13 re-checked it: the arithmetic and the boundary-threshold usage hold, but the limit leans on one field (dropping it loosens by about half), so the wording becomes "≳4× for the full sample; still ≥3× without the most constraining field"; the quoted values are conservative (the interpolant and flux conversion would tighten them by 2–23 %). Red team RT-08: "the limits are well calibrated" does not survive — the toy-MC holds only under its simplified model, and while the Δstat = 2.71 rule is right in form it is not yet calibrated by spectral simulation (E028 running). The outer-field flux falls with observing year (≤2009 1.05 ± 0.08 vs ≥2010 0.55 ± 0.09, about 2σ) — the cause is unresolved, "detector ageing suspected" is withdrawn (the published MOS calibration is a few percent and the fitted ARFs' soft end changes only +1.6 %/−3.7 %), and the 1.3 keV feature is a free 1.2–1.4 keV Gaussian residual, not an instrumental line; the outer/inner ratio and the outer-based limits are marked "epoch-confounded". Red team RT-11 (blocking): E026's runner would fail all 28 fields, the correction list is ready, and no batch before it passes. E027: pre-registration amendment v2 (the field-to-field τ is descriptive only; the primary channel becomes the soft-proton norm and a soft-band surface brightness), the coverage gain reworded to "about 4 useful and 5 marginal pointings", and 203 GTI files accepted. Work: A1d done; E026's corrections being applied; E028 stage 0, RT-14-T1, RT-15 and RT-16 running. Problems: the Virgo ssh outage 06:16Z–~08:50Z (it came back by itself); my 200-toy rerun overwrote RT-08's 1000-toy CSVs (restored deterministically); live entry L006's "detector ageing" claim was withdrawn and rewritten; my RT-13 brief wrote the distance as 776 kpc (the project uses 785 kpc). For the PI: none.

来源：cognition-os/experiments/E013_joint_spatial_profile_likelihood.md（「A1d accepted and A2 baseline on A1d」「Red-team RT-08 triage」「Red-team RT-13 triage」节）、E025_outer_classB_fits.md（RT-09/RT-09-T1/RT-12/RT-09-T2/RT-14 与 Hermes 臂修订节）、E027_gti_policy_screen.md（RT-10 修订 v2、stage 1a 节）；cognition-os/REDTEAM_QUEUE.md 第 15–30 行；.agent/jobs/2026-10-15-e026-virgo-reextract-stage0/CORRECTION_01_RT11.md；cognition-os/insights/I033 末两段；cognition-os/06_AGENT_LOG.md 181–196 行；handbook/live_updates.jsonl L001–L008。前两天：2026-10-15（RT-01 分诊、D50、E025、A2 v2 final）· 2026-10-14（D48 冻结、D49、E014/A2 v1/E016/E019/E020/E021）。

Source: cognition-os/experiments/E013_joint_spatial_profile_likelihood.md (the "A1d accepted and A2 baseline on A1d", "Red-team RT-08 triage" and "Red-team RT-13 triage" sections), E025_outer_classB_fits.md (the RT-09/RT-09-T1/RT-12/RT-09-T2/RT-14 sections and the Hermes-arm amendment), E027_gti_policy_screen.md (the RT-10 amendment v2 and stage-1a sections); rows 15–30 of cognition-os/REDTEAM_QUEUE.md; .agent/jobs/2026-10-15-e026-virgo-reextract-stage0/CORRECTION_01_RT11.md; the last two paragraphs of cognition-os/insights/I033; rows 181–196 of cognition-os/06_AGENT_LOG.md; handbook/live_updates.jsonl L001–L008. Earlier days: 2026-10-15 (RT-01 triaged, D50, E025, A2 v2 final) · 2026-10-14 (D48 frozen, D49, E014/A2 v1/E016/E019/E020/E021).

最新状态（实时，下次刷新写入正文）Live status (folded into the pages at the next refresh)

以下条目在本次网页刷新之后出现，由 Claude 在分诊时从记录中摘写，未经 Hermes 改写。These items arrived after this site refresh; Claude wrote them from the records at triage time, and Hermes has not yet rewritten them.

全部 4 条最新状态 →All 4 live entries →

工作台：M31 的星系周介质（New-ANGELS III）Workbench: the circumgalactic medium of M31 (New-ANGELS III)

这是什么项目？What is this project?

M31（仙女座星系）是离我们最近的大型旋涡星系。宇宙学认为星系的大部分普通物质（重子）不在星系内部，而在星系周围几十万光年的星系周介质（CGM）——百万度的稀薄气体。本项目用 XMM-Newton 的 22 个指向（proposal 080073，PI J. Li）在 M31 内晕（投影半径 10.5–30.8 kpc）逐视线测量弥漫软 X 射线发射，把「银河系前景 + M31 内晕」的总和（记号 CGMsum）测准，再用空间形状限制其中有多少属于 M31。测量已冻结（D48，2026-10-14）：14 个主样本场的 CGMsum 吸收流量为 0.965 ± 0.044（单位 10⁻¹⁵ erg cm⁻² s⁻¹ arcmin⁻²），共温 kT = 0.17531 ± 0.00436 keV。

M31 (Andromeda) is the nearest large spiral galaxy. Cosmology says most of a galaxy's ordinary matter (baryons) is not inside the galaxy but in the circumgalactic medium (CGM), million-degree thin gas around it. Using 22 XMM-Newton pointings (proposal 080073, PI J. Li) toward the M31 inner halo (projected radii 10.5–30.8 kpc), this project measures the diffuse soft X-ray emission sightline by sightline, pins down the sum "Milky Way foreground + M31 inner halo" (notation CGMsum), and then limits how much belongs to M31 from its spatial shape. The measurement is frozen (D48, 2026-10-14): the 14 primary fields give CGMsum absorbed flux 0.965 ± 0.044 (units 10⁻¹⁵ erg cm⁻² s⁻¹ arcmin⁻²), common kT = 0.17531 ± 0.00436 keV.

来源：cognition-os/00_AI_BRIEFING.md（2026-10-14 D48 块）；cognition-os/04_DECISION_LOG.md D48；cognition-os/insights/I033（0.965 ± 0.044，χ²=11.99/13）。

Source: cognition-os/00_AI_BRIEFING.md (2026-10-14 D48 block); cognition-os/04_DECISION_LOG.md D48; cognition-os/insights/I033 (0.965 ± 0.044, χ²=11.99/13).

现在进展如何？Where does it stand?

测量冻结；v19 基线（14 场）可信，但红队 RT-04（已分诊）确认它是局部单起始基线——同模型多元起始把 14 场均值移到 0.891（−7.7%），最终总量将来自 A1d 剖面曲面，模型 B/C 作拓扑系统误差；±0.044 只是统计误差。轮廓在 10.5–29 kpc 内平（χ²=11.99/13）。A2 基线已在 A1d 曲面上重算（临时，阈值未校准）：无 M31 分量被偏好，集中式 M31 β 晕（r_c ≤ 8 kpc，MW 常数）在 R=20 kpc 的单侧 95% 上限（插值版）为 0.049–0.16（β 0.6…0.4），排除 Zhang 名义中值 ≳4 倍；红队 RT-13（已分诊）确认算术与阈值用法，但上限主要靠一个视场——措辞「全样本 ≳4 倍；去掉约束最强的视场仍 ≥3 倍」；红队 RT-08（已分诊）判定 toy-MC 校准只在简化模型下有效，Δstat = 2.71 尚未用能谱模拟校准（E028 运行中）；插值仍未收敛（线性/样条 0.77–0.82）。历史值仅作记录：A2 v1 0.047–0.164（对插值敏感 2.0–2.3 倍，RT-02）；A2 v2（模型 C 曲面，敏感性变体）β = 0.5 时 r_c = 4/8/15/30 kpc 0.111/0.138/0.215/0.484。外样本从 8 个方向扩到 26 个已拟合方向（23 个有双侧误差）：干净外样本（13 场，无未掩展源）吸收流量 IVW 0.846 ± 0.059 vs 内场 0.965 ± 0.044，固定效应外/内 = 0.877 ± 0.072；含未掩展源的全部 23 场 0.831 ± 0.045。红队 RT-03（已分诊）：随机效应给外/内 = 0.94 ± 0.11，逐一剔除 0.83–0.99——稳妥说法是外场与内场在约 10–15% 内一致，「外场更暗」不稳健。E023 完成：内+外联合拟合把大核 M31 晕在 20 kpc 的单侧 95% 上限压到 0.46（r_c=50）／0.59（r_c=100）／0.90（r_c=200）（β = 0.5）——Zhang 名义中值剖面 0.83 被排除到 r_c ≲ 125/175 kpc（β = 0.4/0.5）；红队 RT-02 后「β = 0.6 时 r_c ≤ 200 全部排除」已撤回（15% 外场斑块底噪给 UL 0.90–1.06，该端边际），E023 将在流量空间于谱曲面上重做；EAGLE fiducial 0.51 到 r_c ≲ 45/65/85 kpc（β = 0.4/0.5/0.6）。红队 RT-01 已分诊：E020 模型 C 的「−15%」不是 SP 谱形系统误差。红队 RT-06（已分诊）：D49 原文成立——RASS 只排除宽的 R45 亮超出，OMNI 只排除 OMNI 追踪的高通量 SWCX，NW 四场 SP/QPB 比值约为其余十场 3×（事后）——NW 超出仍是「未解释」，但倾向软质子。红队 RT-09/RT-09-T1/RT-12/RT-14（均已分诊）：外场流量随观测年代下降（≤2009 1.05 ± 0.08 vs ≥2010 0.55 ± 0.09，约 2σ）——「怀疑探测器老化」已撤回，1.3 keV 特征是自由的 1.2–1.4 keV 高斯残差，外/内比与基于外场的极限改标「受观测年代混杂」。红队 RT-11（阻断级）：E026 的 runner 会让全部 28 个场失败，修正通过前不批处理。运行中（待定）：E028、RT-14-T1、RT-15、RT-16、E026 修正、E027 1b/1c、RT-09-T3。

The measurement is frozen and the v19 baseline (14 fields) is trusted, but red team RT-04 (triaged) confirms it is a local single-start baseline — same-model multi-start moves the 14-field mean to 0.891 (−7.7%), the final total will come from the A1d profile surfaces with Models B/C as a topology systematic, and ±0.044 is statistical only. The profile is flat over 10.5–29 kpc (χ²=11.99/13). The A2 baseline has been recomputed on the A1d surfaces (provisional, threshold uncalibrated): no M31 component is preferred, and the one-sided 95% upper limit (interpolated) on a concentrated M31 beta halo (r_c ≤ 8 kpc, MW constant) at R = 20 kpc is 0.049–0.16 (β 0.6…0.4), excluding Zhang's nominal median by ≳4×; red team RT-13 (triaged) confirms the arithmetic and the threshold's usage, but the limit leans on one field — the wording is "≳4× for the full sample; still ≥3× without the most constraining field"; red team RT-08 (triaged) rules that the toy-MC calibration holds only under its simplified model, so Δstat = 2.71 is not yet calibrated by spectral simulation (E028 running); interpolation is still not converged (linear/spline 0.77–0.82). Historical values, for the record only: A2 v1 0.047–0.164 (interpolation-sensitive by 2.0–2.3×, RT-02); A2 v2 (on the Model-C surfaces, a sensitivity variant) gives, for β = 0.5 at r_c = 4/8/15/30 kpc, limits of 0.111/0.138/0.215/0.484. The outer sample has grown from 8 to 26 fitted directions (23 with two-sided errors): the clean outer sample (13 fields without an unmasked extended target) has absorbed-flux IVW 0.846 ± 0.059 vs the inner 0.965 ± 0.044, fixed-effect outer/inner = 0.877 ± 0.072; all 23 fields including unmasked targets give 0.831 ± 0.045. Red team RT-03 (triaged): random effects give outer/inner = 0.94 ± 0.11 and leave-one-out 0.83–0.99 — the safe statement is that outer and inner agree within ~10–15%, and "the outer fields are fainter" is not robust. E023 is done: a joint inner+outer fit limits a large-core M31 halo at 20 kpc to a one-sided 95% UL of 0.46 (r_c=50) / 0.59 (r_c=100) / 0.90 (r_c=200) for β = 0.5 — the Zhang nominal median profile 0.83 is excluded up to r_c ≲ 125/175 kpc (β = 0.4/0.5); after red team RT-02, "for all r_c ≤ 200 at β = 0.6" is withdrawn (a 15% outer patchiness floor gives UL 0.90–1.06, so that end is marginal), and E023 will be redone in flux space on spectral surfaces; the EAGLE fiducial 0.51 up to r_c ≲ 45/65/85 kpc (β = 0.4/0.5/0.6). Red team RT-01 is triaged: E020's Model-C "−15%" is not an SP-shape systematic. Red team RT-06 (triaged): D49 stands as written — RASS excludes only a broad R45-bright excess, OMNI excludes only OMNI-traced high-flux SWCX, and the NW four's SP/QPB ratio is about 3× the other ten (a posteriori) — the NW excess is still "unexplained" but leans towards soft protons. Red teams RT-09/RT-09-T1/RT-12/RT-14 (all triaged): the outer-field flux falls with observing year (≤2009 1.05 ± 0.08 vs ≥2010 0.55 ± 0.09, about 2σ) — "detector ageing suspected" is withdrawn, the 1.3 keV feature is a free 1.2–1.4 keV Gaussian residual, and the outer/inner ratio and the outer-based limits are marked "epoch-confounded". Red team RT-11 (blocking): E026's runner would fail all 28 fields, and no batch starts before the correction passes. Running (Pending): E028, RT-14-T1, RT-15, RT-16, E026's corrections, E027 1b/1c, RT-09-T3.

还有什么没解决？What is unsolved?

① 前景与 M31 谱形几乎相同，无法按距离分解——集中核（r_c ≤ 8 kpc）已被 A2 基线（A1d 曲面）压到 0.049–0.16（见状态页；红队 RT-13 已分诊：算术与 Δstat = 2.71 的边界用法成立，但上限主要靠一个视场，措辞「全样本 ≳4 倍；去掉约束最强的视场仍 ≥3 倍」；红队 RT-08 已分诊：阈值尚未用能谱模拟校准，E028 运行中；插值仍未收敛，线性/样条 0.77–0.82），大核（r_c ≥ 50 kpc）被 E023 首次限制（0.46–0.90 @20 kpc，β = 0.5；β = 0.6 大核端边际，「r_c ≤ 200 全部排除」已撤回）；② 外 NW 四场偏亮：E014 排除宽的 R45 亮天空结构与 OMNI 追踪的高通量太阳风；红队 RT-01 分诊后 NW 超出仍是「未解释」——固定 SP 谱形的说法不成立（等权重下模型 C 内 NW/其余仍 1.34），D49 原文生效（0.25 未解释场级系统误差），D49 更新（−15% SP 谱形系统误差）已撤回；红队 RT-06 后「未解释」倾向软质子（NW 四场 SP/QPB 约为其余十场 3×，事后 p ≈ 0.005–0.008）；③ v19 拟合 QA：模型 C（SP 谱形 + 线能量自由）只是敏感性变体——−15% 的 IVW 移动一半来自 v19 重收敛（A 步 −7.65%，A→B −5.29%，B→C −2.87%），且模型 C 的 SP 谱形未经物理验证（13/14 场 SP 升、MWhalo 降）；④ PSF 翼泄漏实测 0.16（原值）／0.33（零点校正）vs E016 模型 0.12——落在 E016 的 0.12–0.33 括号内（红队 RT-05：高括号 0.33，不写「0.28 保守」）；⑤ Locatelli 中心预言 1.29 > 0.965，但其参数范围 0.93–1.57 含 0.965 且 M31 在拟合所用 eRASS1 半球之外——写作「有条件紧张」，不写「高估」（红队 RT-04）；⑥ 300 kpc 图：外样本已扩到 26 个已拟合方向（23 个有双侧误差）（固定效应干净外/内 0.877 ± 0.072，随机效应 0.94 ± 0.11——稳妥说法「约 10–15% 内一致」），但全部在负银纬；12 个 P1 场与 10 个 P2 a+b 场的 v19 掩膜盖不住引文的展源圆（≤ 0.066），需 Virgo 重提取（E026，28 场——红队 RT-11（阻断级）发现 runner 会让全部 28 个场在审计这一步失败，修正单正在试点场上应用，通过前不批处理）；⑦ 外场流量随观测年代下降（红队 RT-09/RT-09-T1/RT-12/RT-14，均已分诊）：≤2009 的 10 个场 1.05 ± 0.08 vs ≥2010 的 3 个场 0.55 ± 0.09（内场全是 2017–2018 年），年代分割约 2σ；「怀疑探测器老化」已撤回（公布的 MOS 校准只有几个百分点，实测 ARF 软端只变 +1.6%/−3.7%），1.3 keV 特征是自由的 1.2–1.4 keV 高斯残差而非仪器线，观测年份与 CCD 状态高度共线（方差膨胀因子 20–39）——原因未定，外/内比与基于外场的极限（E022/E023 大核上限）标为「受观测年代混杂」，等 RT-14-T1 与 RT-09-T3。为什么这叫「混杂」：内场全部采于 2017–2018 年，外场横跨 2001–2022 年，所以「外场比内场暗」这句话把两种可能混在一起——要么外场方向上的天空本来就更暗，要么晚期的观测整体上低估了流量。打个比方：用 2001 年的相机拍的一批照片和用 2022 年的相机拍的另一批照片比亮度，如果两批的曝光标定差了几个百分点，而你要找的信号只有 10%，那么「新照片更暗」既可能是物体真的暗了，也可能只是新相机的标定更保守。要分开它们，必须在同一套校准与同一套背景处理下重采一批早期和一批晚期外场；在那之前，任何依赖外场的内外比较都同时带着方向差异和年代差异，不能单独归因于任一个。

(i) The foreground and M31 have nearly identical spectra, so distance separation fails — the compact core (r_c ≤ 8 kpc) is already pushed to 0.049–0.16 by the A2 baseline on the A1d surfaces (see the state page; red team RT-13 is triaged: the arithmetic and the Δstat = 2.71 boundary usage hold, but the limit leans on one field, with the wording "≳4× for the full sample; still ≥3× without the most constraining field"; red team RT-08 is triaged: the threshold is not yet calibrated by spectral simulation, E028 running; interpolation is still not converged, linear/spline 0.77–0.82), and the large core (r_c ≥ 50 kpc) is limited for the first time by E023 (0.46–0.90 at 20 kpc, β = 0.5; the β = 0.6 large-core end is marginal, "for all r_c ≤ 200" withdrawn); (ii) the bright outer-NW four: E014 excluded a broad R45-bright sky structure and OMNI-traced high-flux solar wind; after red team RT-01's triage the NW excess is still "unexplained" — the fixed-SP-shape explanation does not hold (with equal weights NW/other stays 1.34 even under Model C), D49 stands as written (the 0.25 unexplained field-level systematic), and the D49 update (a −15% SP-shape systematic) is withdrawn; after red team RT-06 "unexplained" leans towards soft protons (the NW four's SP/QPB is about 3× the other ten, a posteriori p ≈ 0.005–0.008); (iii) v19 fit QA: Model C (free SP shape and line energies) is only a sensitivity variant — half of the −15% IVW shift comes from re-converging the v19 fits (step A −7.65%, A→B −5.29%, B→C −2.87%), and Model C's SP shapes are not physically validated (SP up, MWhalo down in 13/14 fields); (iv) the PSF-wing leak is 0.16 (raw) / 0.33 (null-corrected) empirically vs E016's model 0.12 — inside E016's 0.12–0.33 bracket (red team RT-05: the high bracket is 0.33, not "0.28 conservative"); (v) Locatelli's central prediction 1.29 > 0.965, but its parameter range 0.93–1.57 contains 0.965 and M31 lies outside the eRASS1 half used for the fit — write "conditional tension", not "over-predicts" (red team RT-04); (vi) the 300-kpc map: the outer sample has grown to 26 fitted directions (23 with two-sided errors) (fixed-effect clean outer/inner 0.877 ± 0.072, random effects 0.94 ± 0.11 — the safe statement "agree within ~10–15%") but all at negative Galactic latitude; the v19 masks of the 12 P1 fields and 10 P2 a+b fields do not cover the cited extended-target circles (≤ 0.066), so they need Virgo re-extraction (E026, 28 fields — red team RT-11 (blocking) found the runner would fail all 28 fields at the audit step; the correction list is being applied on the pilot field and no batch starts before it passes); (vii) the outer-field flux falls with observing year (red teams RT-09/RT-09-T1/RT-12/RT-14, all triaged): 10 fields ≤2009 give 1.05 ± 0.08 vs 3 fields ≥2010 giving 0.55 ± 0.09 (the inner fields are all 2017–2018), an epoch split of about 2σ; "detector ageing suspected" is withdrawn (the published MOS calibration is a few percent and the fitted ARFs' soft end changes only +1.6 %/−3.7 %), the 1.3 keV feature is a free 1.2–1.4 keV Gaussian residual not an instrumental line, and observation year is strongly collinear with CCD state (variance inflation 20–39) — the cause is unresolved, and the outer/inner ratio and the outer-based limits (E022/E023's large-core ULs) are marked "epoch-confounded" pending RT-14-T1 and RT-09-T3. Why this is called a "confound": the inner fields were all taken in 2017–2018 while the outer fields span 2001–2022, so the sentence "the outer fields are fainter than the inner ones" mixes two possibilities together — either the sky in the outer directions really is fainter, or the later observations systematically underestimate the flux. An analogy: comparing the brightness of one batch of photographs taken with a 2001 camera against another batch taken with a 2022 camera, if the two batches' exposure calibrations differ by a few percent while the signal you are hunting is only 10%, then "the new photographs are dimmer" could mean either that the object really got dimmer or that the new camera's calibration is simply more conservative. Separating them requires re-taking a batch of early and a batch of late outer fields under one calibration and one background treatment; until then every outer-based inner/outer comparison carries both a direction difference and an epoch difference and cannot be attributed to either alone.

下一步做什么？What do I do next?

按 编号计划：A1d 已验收，A2 基线已在它上面重算（临时，阈值未校准）→ 等 E028 用能谱模拟校准 Δstat = 2.71（stage 0 运行中，RT-15 审其设计）；在谱曲面上以流量空间重做 E023（加 MW 纬度项与外场系统项）；等 E026 的 28 场 Virgo 重提取（红队 RT-11 发现 runner 会让全部 28 个场失败，修正单正在试点场上应用，通过前不批处理）→ 扩大后的内外样本联合拟合；红队 RT-14-T1（只用响应文件的 CCF 审计）与 RT-09-T3（早/晚配对重提取）决定外场年代混杂的原因。已完成：E013 A2 v1 + toy MC、A2 v2（敏感性变体）、E019、E020、E021、E022、E023、E017a、V01、E024、E025、A1c、A1d + A2 基线、RT-01…RT-14 分诊（RT-15/RT-16 运行中）。命令见 byhand。

Follow the numbered plan: A1d is accepted and the A2 baseline has been recomputed on it (provisional, threshold uncalibrated) → wait for E028 to calibrate Δstat = 2.71 with spectral simulation (stage 0 running, RT-15 reviews its design); redo E023 in flux space on spectral surfaces (with an MW latitude term and an outer systematic); wait for E026's 28-field Virgo re-extraction (red team RT-11 found the runner would fail all 28 fields; the correction list is being applied on the pilot field and no batch starts before it passes) → the joint fit on the enlarged sample; red teams RT-14-T1 (a response-only CCF audit) and RT-09-T3 (an early/late paired re-extraction) decide the cause of the outer epoch confound. Done: E013 A2 v1 + toy MC, A2 v2 (a sensitivity variant), E019, E020, E021, E022, E023, E017a, V01, E024, E025, A1c, A1d + the A2 baseline, and the RT-01…RT-14 triages (RT-15/RT-16 running). Commands at byhand.

最新骨架：流水线与各部分如何衔接Latest skeleton: the pipeline and how the parts connect

pipeline diagram
图：从 XMM 数据到 New-ANGELS III 论文的流水线。左：数据与前景输入；中：v19 逐 ObsID Sherpa 拟合；右：冻结的 CGMsum 测量与四条解释线（联合似然、NW 交叉检验、未解源、SP 谱形重拟合）。 Figure: the pipeline from XMM data to the New-ANGELS III paper. Left: data and foreground inputs; middle: the per-ObsID v19 Sherpa fit; right: the frozen CGMsum measurement and the four interpretation lines (joint likelihood, NW cross-check, unresolved sources, SP-shape refit).

来源：流程依据 cognition-os/00_AI_BRIEFING.md 与 background_spectrum_fitting_2T_basedon_obsid_v19.py 的模型结构；数值见 cognition-os/insights/I033。

Source: the flow follows cognition-os/00_AI_BRIEFING.md and the model structure of background_spectrum_fitting_2T_basedon_obsid_v19.py; numbers from cognition-os/insights/I033.

目标表（O1–O5）Objectives (O1–O5)

目标Objective 状态Status 当前最好数值（估计器已命名）Best current number (estimator named) 趋势Trend
O1 M31 内晕方向冷视线总和（CGMsum）的测量Total cool line-of-sight emission (CGMsum) toward the M31 inner halo done 14 场逆方差平均：吸收 0.5–2 keV 流量 0.965 ± 0.044；共温 kT = 0.17531 ± 0.00436 keV（v19，局部协方差误差）。红队 RT-04（已分诊）：该值是局部单起始基线——同模型多元起始（E020 步 A）把均值移到 0.891（−7.7%）；最终总量将来自 A1d 剖面曲面，模型 B/C 作拓扑系统误差；±0.044 只是统计误差14-field inverse-variance mean: absorbed 0.5–2 keV flux 0.965 ± 0.044; common kT = 0.17531 ± 0.00436 keV (v19, local-covariance errors). Red team RT-04 (triaged): that value is a local single-start baseline — same-model multi-start (E020 step A) moves the mean to 0.891 (−7.7%); the final total will come from the A1d profile surfaces with Models B/C as a topology systematic, and ±0.044 is statistical only →冻结（总量口径待 A1d）frozen (the total's definition awaits A1d)
O2a 由轮廓形状约束集中式 M31 热晕（小核）的大小/幅度Constrain the size/amplitude of a concentrated M31 hot halo (compact core) from the profile shape provisional A2 基线在 A1d 上（2026-10-16 13:2xZ，临时，阈值未校准）：更密的似然网格（14 场 × 117 点 = 1638/1638，0 个拟合失败，65.7 worker-小时）已验收。同一估计器（MW 常数、kT 在最佳共温 ±1 行内 profile、Δstat = 2.71、插值）给出 20 kpc 处单侧 95% 上限：核半径 ≤ 8 kpc 时 0.12–0.16（β 0.4）／0.075–0.10（β 0.5）／0.049–0.071（β 0.6）；r_c 30 为 0.30–0.56，r_c 50 为 0.68–1.26（单位 1e-15 erg/cm²/s/arcmin²）。插值收敛：线性/样条上限比 0.77–0.82（A1 上是 0.43–0.50），仍差于 10% 目标——取较松的样条值，线性值为插值系统误差（−18 至 −23%）。Zhang 名义中值 0.8285 / 上限（r_c ≤ 8）= 5.2–6.7（β 0.4）、8.1–11.1（β 0.5）、11.7–17.1（β 0.6），对外措辞 「全样本 ≳4 倍；去掉约束最强的视场仍 ≥3 倍」。红队 RT-13（已分诊）：算术与输入成立，A ≥ 0 的最小值在全部 39 个配置都落在 A = 0，所以 Δstat = 2.71 是「参数不能为负」边界情形下正确的 50:50 分位数；上限主要靠一个视场（0800730601）——去掉它放宽 +46…+62%（r_c 8、β 0.4：0.161 → 0.253，Zhang/上限 5.2 → 3.3）；报出的自然样条值是四种插值里最松的（去重/PCHIP/线性紧 2–23%），kT 一致的流量换算再紧约 9.7%。红队 RT-08（已分诊）：「上限已校准良好」不成立——toy-MC 的位置族模型只在简化下有效（高斯抽样下覆盖率 0.78–0.88），Δstat = 2.71 本身在理想高斯情形下仍正确，但阈值尚未用能谱模拟校准（E028 stage 0 运行中）。RT-02 的插值警告仍在：A1 上敏感 2.0–2.3 倍，A1d 上降到 0.77–0.82。RT-07：只排除 Zhang 的名义中值剖面，不写「叠加预期」（Zhang 只发布 16/84% 边际、无协方差；v19 APEC 桥给 0.76–0.90）。历史值仅作记录：A2 v1（0.047–0.164，toy-MC 覆盖率 0.975–0.984）与敏感性变体 A2 v2（模型 C 曲面，β = 0.5 时 r_c = 4/8/15/30 = 0.111/0.138/0.215/0.484，覆盖率 0.918–0.933）The A2 baseline on A1d (2026-10-16 13:2xZ; provisional, threshold NOT calibrated): the denser likelihood grid (14 fields × 117 points = 1638/1638, 0 failed fits, 65.7 worker-hours) is accepted. The same estimator (MW constant, kT profiled over the best shared kT ±1 row, Δstat = 2.71, interpolated) gives one-sided 95% ULs at 20 kpc: for core radius ≤ 8 kpc, 0.12–0.16 (β 0.4) / 0.075–0.10 (β 0.5) / 0.049–0.071 (β 0.6); r_c 30 gives 0.30–0.56 and r_c 50 gives 0.68–1.26 (units 1e-15 erg/cm²/s/arcmin²). Interpolation convergence: the linear/spline UL ratio is 0.77–0.82 (0.43–0.50 on A1), still outside the 10 % target — the looser spline values are quoted and the linear ones are the interpolation systematic (−18 to −23 %). Zhang's nominal median 0.8285 / UL for r_c ≤ 8 = 5.2–6.7 (β 0.4), 8.1–11.1 (β 0.5), 11.7–17.1 (β 0.6); the public wording is "≳4× for the full sample; still ≥3× without the most constraining field". Red team RT-13 (triaged): the arithmetic and inputs hold, the A ≥ 0 minimum sits at A = 0 in all 39 configurations, so Δstat = 2.71 is the correct 50:50 quantile for a parameter bounded at zero; the limit leans on one field (0800730601) — leaving it out loosens by +46…+62 % (r_c 8, β 0.4: 0.161 → 0.253, Zhang/UL 5.2 → 3.3); the quoted natural-spline UL is the loosest of four interpolants (de-duplicated/PCHIP/linear are 2–23 % tighter) and a kT-consistent flux conversion another ≈9.7 % tighter. Red team RT-08 (triaged): "the limits are well calibrated" does not survive — the toy-MC location-family model holds only under its simplification (coverage 0.78–0.88 with a Gaussian sampler), and while the Δstat = 2.71 rule itself is still right in the ideal Gaussian case the threshold is not yet calibrated by spectral simulation (E028 stage 0 running). RT-02's interpolation warning still stands: 2.0–2.3× sensitive on A1, down to 0.77–0.82 on A1d. RT-07: exclude Zhang's nominal median profile only, never "the stacked expectation" (Zhang publishes only 16/84 % marginals, no covariance; the v19 APEC bridge gives 0.76–0.90). Historical values, for the record only: A2 v1 (0.047–0.164, toy-MC coverage 0.975–0.984) and the sensitivity variant A2 v2 (Model-C surfaces, for β = 0.5 at r_c = 4/8/15/30 = 0.111/0.138/0.215/0.484, coverage 0.918–0.933) ↗
O2b 约束大核 M31 热晕（r_c ≥ 50 kpc，A2 够不到的区间）Constrain a large-core M31 hot halo (r_c ≥ 50 kpc, the regime A2 cannot reach) provisional E022（外场参考）+ E023（联合拟合，22 个 v19 norm，高斯似然）：β = 0.5 时 20 kpc 单侧 95% 上限 0.46（r_c=50）／0.59（r_c=100）／0.90（r_c=200）；内场单独拟合在 r_c ≥ 50 kpc 无约束力（r_c=100 时 3.61，r_c=200 时 13.8）。Zhang 名义中值 0.83 被排除到 r_c ≲ 125/175 kpc（β = 0.4/0.5）；红队 RT-02（已分诊）：「β = 0.6 时 r_c ≤ 200 全部排除」已撤回——15% 外场斑块底噪给 β = 0.6、r_c = 200 的上限 0.90–1.06，该端边际；E023 还因单一流量/norm 转换（外场跨度 1.40–3.22e-10）而将在流量空间于谱曲面上重做（加 MW 纬度项与外场系统项）。EAGLE fiducial 0.51 到 r_c ≲ 45/65/85 kpc（β = 0.4/0.5/0.6）。E022 的内禀 norm 外/内 = 0.80 ± 0.13（两组样本统计均匀），平 M31 项 ≲ 0.4–0.46（2σ，RT-03 后两种写法并列）。E025（D50 的 P1+P2）把外样本扩到 26 个已拟合方向（23 个有双侧误差）：固定效应干净外/内 = 0.877 ± 0.072；红队 RT-03（已分诊）：随机效应 0.94 ± 0.11，逐一剔除 0.83–0.99——稳妥说法「外场与内场在约 10–15% 内一致」，「外场更暗」不稳健E022 (the outer reference) + E023 (joint fit of the 22 v19 norms, Gaussian likelihood): for β = 0.5 the one-sided 95% UL at 20 kpc is 0.46 (r_c=50) / 0.59 (r_c=100) / 0.90 (r_c=200); inner-only fits are uninformative at r_c ≥ 50 kpc (3.61 at r_c=100, 13.8 at r_c=200). The Zhang nominal median 0.83 is excluded up to r_c ≲ 125/175 kpc (β = 0.4/0.5); red team RT-02 (triaged): "for all r_c ≤ 200 at β = 0.6" is withdrawn — a 15% outer patchiness floor gives a β = 0.6, r_c = 200 limit of 0.90–1.06, so that end is marginal; E023 is also to be redone in flux space on spectral surfaces (with an MW latitude term and an outer systematic) because of its single flux/norm conversion (the outer fields span 1.40–3.22e-10). The EAGLE fiducial 0.51 up to r_c ≲ 45/65/85 kpc (β = 0.4/0.5/0.6). E022's intrinsic norm outer/inner = 0.80 ± 0.13 (both samples statistically homogeneous), and a flat M31 term is ≲ 0.4–0.46 (2σ, both wordings after RT-03). E025 (D50's P1+P2) has grown the outer sample to 26 fitted directions (23 with two-sided errors): fixed-effect clean outer/inner = 0.877 ± 0.072; red team RT-03 (triaged): random effects 0.94 ± 0.11, leave-one-out 0.83–0.99 — the safe statement is "outer and inner agree within ~10–15%", and "the outer fields are fainter" is not robust ↗
O3 外 NW 亮场的起源Origin of the bright outer-NW fields E014 + E018a + E020 + RT-01 + RT-06 done 红队 RT-01 分诊（2026-10-15，Claude 复核全部 6 项）：E020 模型 C 的「NW 超出大半是固定 SP 谱形伪迹」不成立——−15% 的 IVW 移动中 v19 重收敛（A 步）贡献 −7.65%、SP 谱形（A→B）−5.29%、线能量（B→C）−2.87%；用共同的模型 C 误差时 NW/其余比值 1.345（v19）→ 1.133（A）→ 1.134（C）——下降发生在重收敛，不在 SP 谱形；等权重下 1.381 → 1.274（A）→ 1.341（C），模型 C 内 NW 超出仍在。I033 第 5 条回到 Questioned；D49 原文生效（0.25 未解释场级系统误差）；D49 更新撤回；模型 C 只是敏感性变体。红队 RT-06 分诊（2026-10-16，Claude 复算 4/4）：D49 原文的措辞收窄——RASS 只排除宽的 R45 亮超出（足迹稀释：28% 足迹内超出稀释到 1.09–1.14，观测 0.91 ± 0.12，≈1.8σ），不排除足迹尺度结构或弱映射进 R45 的谱；OMNI 只排除 OMNI 追踪的高通量 SWCX（日球层 SWCX 未检验）；NW 四场的 SP/QPB 比值约为其余十场的 3×（3.06×，事后 p ≈ 0.005–0.008）——「未解释」倾向软质子。E014（D49）：RASS 比值 0.91 ± 0.12 vs XMM 1.28；OMNI 无窗口超 4e8 cm⁻² s⁻¹；I034（ParetoGTICal 交叉核对）：v19 用 2.1.filter GTI，比 2.filter 更严，NW 四场没有被欠过滤——剩余耀斑时间不是解释；稳态 SP 对 GTI 不可见，E018b 预期变化不大Red team RT-01 triaged (2026-10-15, Claude reproduced all 6 checks): E020's claim that "the NW excess is mostly a fixed-SP-shape artefact" does not hold — of the −15% IVW shift, the v19 re-convergence (step A) gives −7.65%, the SP shape (A→B) −5.29%, the line energies (B→C) −2.87%; with common Model-C errors the NW/other ratio goes 1.345 (v19) → 1.133 (A) → 1.134 (C) — the drop happens at re-convergence, not at the SP shape; with equal weights 1.381 → 1.274 (A) → 1.341 (C), so the NW excess survives under Model C. I033 point 5 is back to Questioned; D49 stands as written (the 0.25 unexplained field-level systematic); the D49 update is withdrawn; Model C is a sensitivity variant only. Red team RT-06 triaged (2026-10-16, Claude reproduced 4/4): D49's wording is narrowed — RASS excludes only a broad R45-bright excess (footprint dilution: a 28% excess inside the footprints dilutes to 1.09–1.14 against the observed 0.91 ± 0.12, ≈1.8σ), not footprint-scale structure or a spectrum that maps weakly into R45; OMNI excludes only OMNI-traced high-flux SWCX (heliospheric SWCX is untested); the NW four's SP/QPB ratio is about 3× the other ten (3.06×, a posteriori p ≈ 0.005–0.008) — "unexplained" leans towards soft protons. E014 (D49): RASS ratio 0.91 ± 0.12 vs XMM 1.28; no OMNI window above 4e8 cm⁻² s⁻¹; I034 (the ParetoGTICal cross-check): v19 uses the 2.1.filter GTIs, stricter than 2.filter, and the NW four are not under-filtered — residual flare time is not the explanation; a steady SP floor is invisible to GTIs, so E018b is expected to change little ↗
O4 银河系前景模型 + Local Hot Bridge 检验（300 kpc 图）MW foreground model + Local Hot Bridge test (300-kpc map) E021/E022/E023/E024/E025 完成；E026 修正中（RT-11 阻断）E021/E022/E023/E024/E025 done; E026's corrections being applied (RT-11 blocking) E021 清点 + E022（8 个 class-A 外场拟合）之后，D50（PI，2026-10-15「同意第二个」）采纳 E024 建议：P1+P2 立即拟合（E025），6 个真正外围场上 Virgo 重提取，跳过 P4，弃 NGC 147（仅剩 22% 视场）。E025 结果：P1 12/12 未通过掩膜审计（现有 v19 谱的掩膜只盖住引文展源圆的 0–9%），全部转入重提取清单；P2 19 场拟了 18 个（1 个 dof=0 剔除），其中 3 个无有限通量误差。干净外样本（13 场，无未掩展源）IVW 0.846 ± 0.059（χ² 27.3/12）vs 内场 0.965 ± 0.044，固定效应外/内 = 0.877 ± 0.072；全部 23 场（含未掩展源）0.831 ± 0.045；与 b（ρ = 0.04/−0.10）、N_H（ρ = −0.13/−0.19）均无相关；未掩展源不抬高流量（b-only 0.865 ± 0.133 vs a+b 0.809 ± 0.069）。红队 RT-03（已分诊）：干净 IVW 有 27% 权重集中在一个场（0884992601）且过度离散（χ²/dof 2.27）——随机效应外/内 = 0.94 ± 0.11（全部 23 场 0.89 ± 0.08），逐一剔除 0.83–0.99；稳妥说法是「外场与内场在约 10–15% 内一致」，「外场更暗」不稳健；误差模型敏感（按剖面半宽重加权把 E022 均值移到 0.897，占干净差距的 46%）。correction 1：10 个 P2 a+b 场也未通过掩膜审计（MOS 最低 0–0.066）→ 重提取清单 = 6（D50）+ 12（P1）+ 10（a+b）= 28 场（E026，运行中，待定）。全部新增方向仍在 b < 0（−33.2 … −9.6），无正纬度覆盖。Locatelli 模型在 M31 足迹投影 1.290 vs 测量 0.965（残差 −0.326，有条件紧张——其参数范围 0.93–1.57 含 0.965，且 M31 在拟合所用 eRASS1 半球之外，红队 RT-04）；Ueda 0.312、HaloSat 0.565。红队 RT-09/RT-09-T1/RT-12（已分诊）：外场流量随观测年代下降——干净 13 场 2009 年及以前（n = 10）1.05 ± 0.08 vs 2010 年以后（n = 3）0.55 ± 0.09，而内场全部是 2017–2018 年；年代分割约 2σ（置换 p = 0.027），2001–2009 的下降可能部分是太阳周期 SWCX，2010 年后的偏低不是。红队 RT-14（luna + Hermes 臂均已分诊）：原因未定——「怀疑探测器老化」已撤回（公布的 MOS 校准只有几个百分点；拟合实际使用的 82 个 ARF 逐一核对哈希无误，其软端有效面积在 2010 年前后只变 +1.6%/−3.7%，离解释 43% 所需差 12–27 倍）；那条 1.3 keV 特征是自由的 1.2–1.4 keV 高斯残差（补偿 QPB 切除后残留的铝荧光，占该能段 17–38%），不是仪器荧光线示踪量；观测年份与探测器 CCD 状态高度共线（方差膨胀因子 20–39）。因此外/内比与基于外场的极限（E022/E023 大核上限）一律标为「受观测年代混杂」，直到 RT-14-T1（运行中）与 RT-09-T3（排队）出结果。红队 RT-11（阻断级）：E026 的 Virgo runner 会让全部 28 个场在审计这一步失败（3 处阻断级缺陷），CCF 也未锁定到 v19 集；修正单 CORRECTION_01_RT11.md 已备好，修正通过前不启动批处理After the E021 inventory and E022 (8 class-A outer fields fitted), D50 (the PI, 2026-10-15, "agree with the second") adopts E024's suggestion: fit P1+P2 now (E025), re-extract 6 genuinely-outer fields on Virgo, skip P4, drop NGC 147 (only 22% of the field of view left). E025 results: all 12 P1 fields fail the mask audit (the existing v19 spectra's mask covers only 0–9% of the cited extended-target circle), all moving to the re-extraction list; 18 of 19 attempted P2 fits are accepted (1 rejected, dof = 0), 3 of them without a finite flux error. The clean outer sample (13 fields without an unmasked extended target) has IVW 0.846 ± 0.059 (χ² 27.3/12) vs the inner 0.965 ± 0.044, fixed-effect outer/inner = 0.877 ± 0.072; all 23 fields (including unmasked targets) give 0.831 ± 0.045; no trend with b (ρ = 0.04/−0.10) or N_H (ρ = −0.13/−0.19); unmasked extended targets do not raise the flux (b-only 0.865 ± 0.133 vs a+b 0.809 ± 0.069). Red team RT-03 (triaged): the clean IVW mean carries 27% of its weight in one field (0884992601) and is overdispersed (χ²/dof 2.27) — random effects give outer/inner = 0.94 ± 0.11 (all 23 fields 0.89 ± 0.08), leave-one-out 0.83–0.99; the safe statement is "outer and inner agree within ~10–15%", and "the outer fields are fainter" is not robust; the error model matters (reweighting by profile half-widths moves E022's mean to 0.897, 46% of the clean gap). Correction 1: the 10 P2 a+b fields also fail the mask audit (MOS minimum 0–0.066) → the re-extraction list = 6 (D50) + 12 (P1) + 10 (a+b) = 28 fields (E026, running, Pending). All added directions remain at b < 0 (−33.2 … −9.6); no positive-latitude coverage. The Locatelli model projected at the M31 footprint is 1.290 vs the measured 0.965 (residual −0.326, conditional tension — its parameter range 0.93–1.57 contains 0.965, and M31 lies outside the eRASS1 half used for the fit, red team RT-04); Ueda 0.312, HaloSat 0.565. Red teams RT-09/RT-09-T1/RT-12 (triaged): the outer-field flux falls with observing year — in the clean 13, 10 fields up to 2009 give 1.05 ± 0.08 vs 3 fields from 2010 on giving 0.55 ± 0.09, while the inner fields are all from 2017–2018; the epoch split is about 2σ (permutation p = 0.027), part of the 2001–2009 decline may be solar-cycle SWCX, and the post-2010 deficit is not. Red team RT-14 (both the luna and Hermes arms triaged): the cause is unresolved — "detector ageing suspected" is withdrawn (the published MOS calibration is only a few percent; all 82 ARFs actually used in the fits are hash-verified and their soft effective area changes by only +1.6 %/−3.7 % across 2010, a factor 12–27 short of what a 43 % drop needs); the 1.3 keV feature is a free 1.2–1.4 keV Gaussian residual (absorbing aluminium fluorescence left after the QPB cut-out, 17–38 % of that band), not an instrumental-fluorescence tracer; observation year is strongly collinear with detector CCD state (variance inflation 20–39). Therefore the outer/inner ratio and the outer-based limits (E022/E023's large-core ULs) are all marked "epoch-confounded" until RT-14-T1 (running) and RT-09-T3 (queued) report. Red team RT-11 (blocking): E026's Virgo runner would fail all 28 fields at the audit step (3 blocking defects) and its CCF is not pinned to the v19 set; the correction list CORRECTION_01_RT11.md is ready and no batch starts before it passes ↗
O5 New-ANGELS III 论文The New-ANGELS III manuscript draft 16 页内部稿完成（AASTeX，5 表 3 图）；提交门槛见 I013（径向+侧向联合似然、边界感知区间、系统误差网格等）A 16-page internal draft exists (AASTeX, 5 tables, 3 figures); submission gates in I013 (shared radial-plus-side likelihood, boundary-aware intervals, the systematic grid, ...) ↗

来源：O1 cognition-os/insights/I012 与 I033；O2a cognition-os/experiments/E013_joint_spatial_profile_likelihood.md（「A1d accepted and A2 baseline on A1d」「Red-team RT-08 triage」「Red-team RT-13 triage」节；插值表 a2_ul_vs_rc_interp_A1d.csv 与 a2_ul_vs_rc_interp_A1d_linear.csv）、cognition-os/insights/I033 末两段、cognition-os/REDTEAM_QUEUE.md 第 15、26 行（RT-02 的 F3 发现仍在）；O2b cognition-os/experiments/E022（correction 1 节）、E023（limits.csv）与 E025（correction 1 节；correction1_ivw_table.csv）；O3 cognition-os/experiments/E020（「Red-team RT-01 triage」节）、E014、cognition-os/insights/I034、cognition-os/04_DECISION_LOG.md D49 及其 update 节（撤回）、cognition-os/REDTEAM_QUEUE.md；O4 cognition-os/experiments/E021（inventory.csv）、E024（policy_table.csv）、E025（RT-09、RT-09-T1、RT-12、RT-09-T2、RT-14 与 Hermes 臂修订节）、.agent/jobs/2026-10-15-e026-virgo-reextract-stage0/CORRECTION_01_RT11.md、cognition-os/04_DECISION_LOG.md D50、newangels3_discussion/E024_hermes/latitude_reach.csv、00_AI_BRIEFING.md（D42–D46 数值）与 02_CURRENT_QUESTION.md；O5 cognition-os/insights/I013 与 paper_apj_v19_cgmsum_draft/README_draft_status_and_submission_gaps.md。

Source: O1 cognition-os/insights/I012 and I033; O2a the "A1d accepted and A2 baseline on A1d", "Red-team RT-08 triage" and "Red-team RT-13 triage" sections of cognition-os/experiments/E013_joint_spatial_profile_likelihood.md (the interpolated tables a2_ul_vs_rc_interp_A1d.csv and a2_ul_vs_rc_interp_A1d_linear.csv), the last two paragraphs of cognition-os/insights/I033, and rows 15 and 26 of cognition-os/REDTEAM_QUEUE.md (RT-02's F3 finding still stands); O2b cognition-os/experiments/E022 (the correction-1 section), E023 (limits.csv) and E025 (the correction-1 section; correction1_ivw_table.csv); O3 cognition-os/experiments/E020 (the "Red-team RT-01 triage" section), E014, cognition-os/insights/I034, cognition-os/04_DECISION_LOG.md D49 with its update section (withdrawn), cognition-os/REDTEAM_QUEUE.md; O4 cognition-os/experiments/E021 (inventory.csv), E024 (policy_table.csv), E025 (the RT-09, RT-09-T1, RT-12, RT-09-T2 and RT-14 sections and the Hermes-arm amendment), .agent/jobs/2026-10-15-e026-virgo-reextract-stage0/CORRECTION_01_RT11.md, cognition-os/04_DECISION_LOG.md D50, newangels3_discussion/E024_hermes/latitude_reach.csv, cognition-os/00_AI_BRIEFING.md (the D42–D46 numbers) and 02_CURRENT_QUESTION.md; O5 cognition-os/insights/I013 and paper_apj_v19_cgmsum_draft/README_draft_status_and_submission_gaps.md.

未解问题表Open problems

问题Problem 为什么重要Why it matters 当前最佳假设Best current hypothesis 下一个检验Next test 负责Owner 状态Status
MW 与 M31 气体同温同形，EPIC 谱无法按距离分解（I006）MW and M31 gas share temperature and spectral shape; EPIC spectra cannot separate them by distance (I006) 决定论文能声明什么：只能测总和 CGMsum，任何分量归属都需要外部假设Determines what the paper can claim: only the sum CGMsum is measured; any component split needs external assumptions 分离只能来自空间形状（半径衰减、南北对称）与外部前景约束Separation can only come from spatial shape (radial decline, N/S symmetry) and external foreground constraints A2 基线已在 A1d 上重算（见状态页插值表）：集中核（r_c ≤ 8 kpc）20 kpc 上限 0.12–0.16（β 0.4）／0.075–0.10（β 0.5）／0.049–0.071（β 0.6）；红队 RT-13 已分诊——算术与输入成立、Δstat = 2.71 是「参数不能为负」边界情形下正确的分位数，但上限主要靠一个视场（去掉它放宽约一半），措辞「全样本 ≳4 倍；去掉约束最强的视场仍 ≥3 倍」；红队 RT-08 已分诊——toy-MC 校准只在简化模型下有效，阈值尚未用能谱模拟校准（E028 stage 0 运行中）；插值仍未收敛（线性/样条 0.77–0.82，RT-02 在 A1 上的 2.0–2.3 倍警告仍在）The A2 baseline has been recomputed on A1d (see the interpolated table on the state page): the compact-core (r_c ≤ 8 kpc) limit at 20 kpc is 0.12–0.16 (β 0.4) / 0.075–0.10 (β 0.5) / 0.049–0.071 (β 0.6); red team RT-13 is triaged — the arithmetic and inputs hold and Δstat = 2.71 is the correct quantile for a parameter bounded at zero, but the limit leans on one field (dropping it loosens by about half), with the wording "≳4× for the full sample; still ≥3× without the most constraining field"; red team RT-08 is triaged — the toy-MC calibration holds only under its simplified model, so the threshold is not yet calibrated by spectral simulation (E028 stage 0 running); interpolation is still not converged (linear/spline 0.77–0.82, RT-02's 2.0–2.3× warning on A1 still stands) E013 A1d/A2 基线已出（临时）；RT-08/RT-13 已分诊；E028 校准运行中A1d/A2 baseline delivered (provisional); RT-08/RT-13 triaged; E028's calibration running
轮廓平（χ²=11.99/13；E020 模型 C 下同样平，斜率 −0.0005 ± 0.006 /kpc）：M31 集中晕（r_c ≤ 8 kpc）在 20 kpc ≤0.049–0.16（A2 基线在 A1d 上，插值版，β 0.6…0.4；阈值未校准），远低于 Zhang 名义中值剖面 0.83The profile is flat (χ²=11.99/13; still flat under E020's Model C, slope −0.0005 ± 0.006 /kpc): a concentrated M31 halo (r_c ≤ 8 kpc) is ≤0.049–0.16 at 20 kpc (the A2 baseline on A1d, interpolated, β 0.6…0.4; threshold uncalibrated), far below the Zhang nominal median profile 0.83 M31 内晕热气要么比 Zhang 名义中值剖面暗几倍，要么在 10–30 kpc 内分布很平（核 ≳30 kpc）M31's inner-halo hot gas is either several times fainter than the Zhang nominal median profile, or flat over 10–30 kpc (core ≳30 kpc) 「平/大核」这条退路现在也被限制：E023 用 8 个 82–253 kpc 外场把大核晕压到 0.46（r_c=50）／0.59（r_c=100）／0.90（r_c=200）@20 kpc（β=0.5），Zhang 名义中值 0.83 被排除到 r_c ≲ 125/175 kpc（β = 0.4/0.5）；「β = 0.6 时 r_c ≤ 200 全部排除」已被红队 RT-02 撤回（15% 外场斑块底噪给 β = 0.6、r_c = 200 的上限 0.90–1.06，该端边际）；E025 再把外样本扩到 26 个已拟合方向（23 个有双侧误差）（干净外/内 0.877 ± 0.072；RT-03 后稳妥说法是「约 10–15% 内一致」）；r_c = 50 时上限 0.318–0.601（A2 v1，内场单独）已被联合拟合取代the "flat/large-core" escape route is now limited too: using the 8 outer fields at 82–253 kpc, E023 pushes the large-core halo to 0.46 (r_c=50) / 0.59 (r_c=100) / 0.90 (r_c=200) at 20 kpc (β=0.5), excluding the Zhang nominal median 0.83 up to r_c ≲ 125/175 kpc (β = 0.4/0.5); "for all r_c ≤ 200 at β = 0.6" was withdrawn by red team RT-02 (a 15% outer patchiness floor gives a β = 0.6, r_c = 200 limit of 0.90–1.06, so that end is marginal); E025 has since grown the outer sample to 26 fitted directions (23 with two-sided errors) (clean outer/inner 0.877 ± 0.072; after RT-03 the safe wording is "agree within ~10–15%"); the A2 v1 inner-only limit at r_c = 50 (0.318–0.601) is superseded by the joint fit A2 基线（A1d 曲面）已复核集中核上限（RT-13 已分诊）；E028 用能谱模拟校准 Δstat = 2.71 阈值（stage 0 运行中，RT-15 审其设计）；E023 在谱曲面上重做大核上限（加 MW 纬度项）；E016 的未解源项空间上平，不削弱形状上限the A2 baseline (on the A1d surfaces) has re-checked the compact-core limit (RT-13 triaged); E028 calibrates the Δstat = 2.71 threshold with spectral simulation (stage 0 running, RT-15 reviews its design); E023 is redone on spectral surfaces for the large-core limit (with an MW latitude term); E016's unresolved-source terms are spatially flat, so they do not weaken the shape limit E013 provisional
外 NW 四场偏亮（v19 外四/其余十 = 1.28）The four outer-NW fields are bright (v19 outer-four/other-ten = 1.28) 若为银河系纬度结构，会挤占 M31 的解释空间；若为 SWCX 则是时间效应；若为仪器/软质子则影响系统误差预算If it is Galactic-latitude structure it eats into the M31 budget; if SWCX it is a time effect; if instrumental/soft-proton it changes the systematic budget 红队 RT-01 分诊后回到「未解释」：−15% 的 IVW 移动一半来自 v19 重收敛（0800730301 通量 0.975 → 0.559，Δstat 仅 −0.04），固定 SP 谱形只贡献 −5.29%；用共同模型 C 误差时 NW/其余 1.345 → 1.133（A）→ 1.134（C）——下降在重收敛，不在 SP 谱形；等权重下模型 C 内仍 1.34。E014 已排除大尺度天空（RASS 0.91 ± 0.12 vs XMM 1.28）与 OMNI 太阳风；I034：NW 四场没有被 GTI 欠过滤（2.1.filter 比 2.filter 严，NW 四场仅损失 0–28% 时间），稳态 SP 对时间筛选不可见。D49 原文生效（0.25 未解释场级系统误差）；D49 更新（−15% SP 谱形系统误差）已撤回；模型 C 只是敏感性变体back to "unexplained" after red team RT-01's triage: half of the −15% IVW shift comes from re-converging the v19 fits (0800730301's flux 0.975 → 0.559 at Δstat only −0.04), the fixed SP shape contributes just −5.29%; with common Model-C errors NW/other goes 1.345 → 1.133 (A) → 1.134 (C) — the drop is at re-convergence, not the SP shape; with equal weights it stays 1.34 under Model C. E014 excluded large-scale sky (RASS 0.91 ± 0.12 vs XMM 1.28) and OMNI solar wind; I034: the NW four are not under-filtered by the GTIs (2.1.filter is stricter than 2.filter; the NW four lose only 0–28% of live time), and a steady SP floor is invisible to time filtering. D49 stands as written (the 0.25 unexplained field-level systematic); the D49 update (a −15% SP-shape systematic) is withdrawn; Model C is a sensitivity variant only A1d 曲面重跑 A2 基线；红队 RT-02/RT-03 分诊；E018b（Virgo 更严耀斑过滤）预期变化不大（I034），保持可选rerun the A2 baseline on the A1d surfaces; triage red teams RT-02/RT-03; E018b (a stricter Virgo flare filter) is expected to change little (I034) and stays optional E020/RT-01/PI 回到未解释；红队进行中back to unexplained; red team running
v19 拟合 QA：SP 谱形过刚、仪器线能量钉边（E018a/E013）——E020 已量化v19 fit QA: the SP shape was too rigid and line energies pinned (E018a/E013) — now quantified by E020 决定系统误差网格里必须有哪一行（I013 提交门槛 3）；不修会以场级伪迹污染科学结论Decides which row the systematic grid must contain (I013 submission gate 3); unfixed, it contaminates science with field-level artefacts E020（20 场，多元起始）：SP 谱形自由在 0401/0501/0901/1101/2101 改善 8–16；线能量自由在 0401 改善 8.5、1001 改善 15。模型 C 下 14 场逆方差均值 0.964 → 0.819 ± 0.041。红队 RT-01 分解：−15.04% 中重收敛（A）−7.65%、SP 谱形（A→B）−5.29%、线能量（B→C）−2.87%——不是一个 SP 谱形系统误差；模型 C 的 SP 谱形未经物理验证（13/14 场 SP 升、MWhalo 降，log-ratio 相关 −0.49；0800730901 的 SP ×1150 且 PhoIndx2 钉在极限）。模型 C 的剖面误差在固定 kT 下算得（低估），4 场半宽仅 0.15–0.27 σ_v19；残差散布 χ² 24.5/13（p = 0.026），单个 −15% 标量不能替代场级项E020 (20 fields, multi-start): freeing the SP shape improves 0401/0501/0901/1101/2101 by 8–16; freeing line energies improves 0401 by 8.5 and 1001 by 15. Under Model C the 14-field inverse-variance mean is 0.964 → 0.819 ± 0.041. Red team RT-01's decomposition: of −15.04%, re-convergence (A) gives −7.65%, the SP shape (A→B) −5.29%, the line energies (B→C) −2.87% — not a single SP-shape systematic; Model C's SP shapes are not physically validated (SP up, MWhalo down in 13/14 fields, log-ratio correlation −0.49; 0800730901's SP ×1150 with PhoIndx2 at its limit). Model C's profile errors were computed at fixed kT (an underestimate); 4 fields have half-widths only 0.15–0.27 σ_v19; the residual scatter is χ² 24.5/13 (p = 0.026), so a single −15% scalar cannot replace the field-to-field term A1d（v19 模型、每点多元起始）曲面成为 A2 基线（运行中）；模型 C 曲面（A1c）降级为敏感性变体；D49 更新撤回——系统误差预算继续携带场级未解释项the A1d (v19-model, multi-start at every point) surfaces become the A2 baseline (running); the Model-C surfaces (A1c) are demoted to a sensitivity variant; the D49 update is withdrawn — the systematic budget keeps carrying the unexplained field-level term E020 E020 完成E020 done
PSF 翼泄漏：掩膜亮源的翼落在提取区里（E016）PSF-wing leak: the wings of masked bright sources fall inside the extraction region (E016) E016 里它是最大的单一未解源项（模型估计 ≈0.12 F [最大 0.32]，主要是幂律）；若实测偏大，会压低「仅气体」的总和并改变系统误差预算In E016 it is the single largest unresolved-source term (model estimate ≈0.12 F [max 0.32], mostly power law); if the empirical value is larger it lowers the gas-only total and changes the systematic budget E019（已完成）：123 个源相机测量里的 81 个孤立类点源叠加——只有最亮 bin（中位 1.2e-14 cgs）可用：f_leak = 10.3 ± 8.6%（原值）／20.4 ± 7.8%（零点校正）；King 模型对同样椭圆给 14.3%。14 主场的逐场泄漏 SB：0.16（原值）／0.33（零点校正） vs E016 模型 0.12（高括号 0.22）；共享标度误差 ≈ ±0.14。泄漏是幂律型、空间上平——不改变 M31 形状上限；对总量是又一个向下项（主要落在幂律/CXB 项，不在 APEC）E019 (done): a stack of 81 isolated point-like sources out of 123 source-camera measurements — only the brightest bin (median 1.2e-14 cgs) is usable: f_leak = 10.3 ± 8.6% raw / 20.4 ± 7.8% null-corrected; the King model gives 14.3% for the same ellipses. Per-field leak SB over the 14 primaries: 0.16 raw / 0.33 null-corrected vs E016's model 0.12 (high bracket 0.22); the shared-scale error is ≈ ±0.14. The leak is power-law-like and spatially flat — it does not change the M31 shape limit; for the total it is one more downward term (mostly in the power-law/CXB term, not the APEC) 暗 bin 低于零点底噪（椭圆外净计数为负），不可用；2801（E016 异常掩膜标记）在 1.06 > F，0101 在 0.50 ≈ Fthe faint bins sit below the background floor (net counts out of the ellipse are negative), unusable; 2801 (an anomalous mask, E016 flag) is at 1.06 > F and 0101 at 0.50 ≈ F E019 E019 完成E019 done
Locatelli MW 模型在 M31 方向预言 1.29 > 测量 0.965The Locatelli MW model predicts 1.29 > the measured 0.965 toward M31 若模型对，则无正的 M31 中心残差；紧张来自模型、转换或吸收之一If the model is right there is no non-negative central M31 residual; the tension lies in the model, the conversion, or absorption 模型高估、线-宽带转换失配、或 v19 总和偏低（三种解释并列）Model over-prediction, line-to-band conversion mismatch, or a low-biased v19 sum (three live explanations) 300 kpc RASS 图建立本地 MW 模型（O4）A 300-kpc RASS map to build a local MW model (O4) 计划planned 计划planned
300 kpc 图：30 kpc 以外的档案样本很薄（E021）The 300-kpc map: the archival sample beyond 30 kpc is thin (E021) O4（本地 MW 模型 + Local Hot Bridge 检验）需要 30 kpc 以外的视线；样本太薄会限制图的能力O4 (a local MW model + the Local Hot Bridge test) needs sightlines beyond 30 kpc; a thin sample limits what the map can do 145 个 ObsID 里 class-A 只有 22 个，30 kpc 以外仅 8 个（30–60 kpc 为零）；全部在 b = −5° 到 −33°，无正纬度场。E022 已完成这 8 个场的拟合（外 IVW 0.842 ± 0.065；内禀 norm 外/内 0.80 ± 0.13）。D50 已定（2026-10-15）：P1+P2 立即拟合、6 场 Virgo 重提取、跳过 P4、弃 NGC 147。E025 已跑：P1 12/12 掩膜审计未过（掩膜仅盖展源圆 0–9%）→ 全部转重提取；P2 18/19 接受；干净外样本（13 场）0.846 ± 0.059，外/内 0.877 ± 0.072；与 b、N_H 无相关。correction 1 后重提取清单 = 6 + 12 + 10 = 28 场（E026 运行中，待定）；4 个引文半径是估值，任何提升结果前需从图像重推Of 145 ObsIDs only 22 are class A, and just 8 lie beyond 30 kpc (zero at 30–60 kpc); all at b = −5° to −33°, no positive-latitude field. E022 has fitted those 8 fields (outer IVW 0.842 ± 0.065; intrinsic norm outer/inner 0.80 ± 0.13). D50 is decided (2026-10-15): fit P1+P2 now, re-extract 6 fields on Virgo, skip P4, drop NGC 147. E025 has run: all 12 P1 fields fail the mask audit (the mask covers only 0–9% of the extended-target circle) → all move to re-extraction; 18/19 P2 fits accepted; the clean outer sample (13 fields) gives 0.846 ± 0.059, outer/inner 0.877 ± 0.072; no trend with b or N_H. After correction 1 the re-extraction list = 6 + 12 + 10 = 28 fields (E026 running, Pending); 4 cited radii are estimates, to be re-derived from images before any promoted result E026（28 场 Virgo 重提取，D50 条件：复现 v19 的 2.1.filter GTI 约定并逐相机报告相对 2.filter 的保留时间）→ 重提取场按 v19 规则拟合 → 在扩大样本上重做 E023 式联合拟合（加 MW 纬度项）；红队 RT-03 审查外/内比较。红队 RT-11（阻断级，已分诊）：runner 会让全部 28 个场在审计这一步失败（3 处阻断级缺陷）且 CCF 未锁定——修正单 CORRECTION_01_RT11.md 已备好，修正通过前不启动批处理；E027 的 1b Virgo 提取也排在其后。外场年代混杂（RT-09/RT-14）使重提取成为决定性检验：E026 清单里有 15 个 2010 年后的外场，配上锁定校准与冻结污染模型的早/晚配对重提取（RT-09-T3）才能分开年代与 CCD 状态E026 (28-field Virgo re-extraction, D50 conditions: reproduce the v19 2.1.filter GTI convention and report retained time vs 2.filter per camera) → the re-extracted fields are fitted under v19 rules → the E023-style joint fit is redone on the enlarged sample (with an MW latitude term); red team RT-03 reviews the outer/inner comparison. Red team RT-11 (blocking, triaged): the runner would fail all 28 fields at the audit step (3 blocking defects) and its CCF is unpinned — the correction list CORRECTION_01_RT11.md is ready and no batch starts before it passes; E027's 1b Virgo extraction is queued behind it too. The outer epoch confound (RT-09/RT-14) makes the re-extraction the decisive test: the E026 list holds 15 post-2010 outer fields, and only a pinned-calibration, frozen-contamination early/late paired re-extraction (RT-09-T3) can separate year from CCD state E026/RT-03/RT-11 E021/E022/E024/E025 完成；E026 修正中（RT-11 阻断，未批处理）E021/E022/E024/E025 done; E026's corrections being applied (RT-11 blocking, no batch)
源掩膜的出处缺口（I013 提交门槛 5）：v19 掩膜规则不可复原（经验上是 2–4.5 keV 硬带掩膜）The source-mask provenance gap (I013 submission gate 5): the v19 mask rule is unrecoverable (empirically a hard-band 2–4.5 keV mask) 审稿人会问「掩膜怎么来的、漏了多少」；E017a 现在给出实测答案a referee will ask "where did the mask come from and what does it miss"; E017a now answers it empirically E017a（已完成）：v19 掩膜按 0.5–2 keV 流量掩住 Paper I 亮源（>1e-14）的 95.5 %（233/244；11 个未掩，6 个主场亮源完备度低于 0.95）；未掩亮源量级 ≈ 11 × ~2e-14 摊在 ≈ 4200 arcmin² ≈ 0.05（1e-15 单位，多为幂律/CXB 型，小的向上偏置）。掩膜椭圆 8784 个里 87 % 没有 10″ 内的 Paper I 源——v19 掩膜来自更深的逐观测探测表（E016：有效 ML ≈ 0.55），切掉更多（多为暗源/伪源），花的是天空面积不是偏差E017a (done): by 0.5–2 keV flux the v19 mask covers 95.5 % of the bright Paper I sources (>1e-14; 233/244; 11 unmasked, 6 primary fields below 0.95 bright completeness); the unmasked-bright order of magnitude is ≈ 11 × ~2e-14 spread over ≈ 4200 arcmin² ≈ 0.05 (1e-15 units, mostly power-law/CXB-like, a small upward bias). Of the 8784 mask ellipses, 87 % have no Paper I source within 10″ — the v19 mask comes from a much deeper per-observation detection list (E016: effective ML ≈ 0.55), so it cuts many more (mostly faint or spurious) regions; it costs sky area, not bias E017b（低优先，除非审稿人要）：只用 Paper I 星表重建掩膜（需 Virgo 重提取），对比 v19 的 skyarea 与通量E017b (low priority unless a referee asks): rebuild the mask from the Paper I catalogue only (needs re-extraction on Virgo), compare skyarea and flux with v19 E017a E017a 完成E017a done
时间过滤（GTI）与场间底噪：v19 的 GTI 是历史 2.1.filter，场间散布的来源是什么？（I034）Time filtering (GTIs) and the between-field floor: v19's GTIs are the historical 2.1.filter — what causes the field-to-field scatter? (I034) 若散布是残余耀斑，更严的 GTI 能压低它（E018b）；若是稳态 SP/未遮光的 QPB，GTI 无效，系统误差要靠场级模型项携带if the scatter were residual flares, a stricter GTI would suppress it (E018b); if it is a steady SP or unvignetted QPB, GTIs are powerless and the systematic must ride as a field-level model term I034（ParetoGTICal 交叉核对）：v19 的诊断产品与 2.1.filter 逐字节一致（sha256 核对三个场）；对全部 14 主场 2.1.filter 保留的时间少于 2.filter（更严）；NW 四场损失 0–28%，而 1101/2301/2801 损失 17–25% 却不亮——残余耀斑不是 NW 超出的解释；稳态 SP 对 GTI 不可见；ParetoGTICal E020 给出同 z 处 9–12% 的场间 rms 底噪（未遮光 QPB/SP 之差），与谱拟合的场间散布同量级I034 (the ParetoGTICal cross-check): v19's diagnostic products are byte-identical to 2.1.filter (sha256-checked on three fields); on all 14 primaries 2.1.filter keeps less time than 2.filter (stricter); the NW four lose 0–28% while 1101/2301/2801 lose 17–25% without being bright — residual flares do not explain the NW excess; a steady SP floor is invisible to GTIs; ParetoGTICal E020 gives a 9–12% rms between-field floor at the same z (unvignetted QPB/SP differences), the same scale as the spectral fits' field-to-field scatter E018b 预期变化不大（保持可选）；E027 的谱级 stage 1（31 个观测 × 泳道）用预注册的 S1–S4 指标直接回答「GTI 分支有多大」——RT-04-T1（2.filter 分支测试）已并入；D50 的 Virgo 重提取必须复现 2.1.filter 约定并报告相对 2.filter 的保留时间；系统误差预算继续携带场级项E018b is expected to change little (stays optional); E027's spectral stage 1 (31 observations × lanes) answers "how big is the GTI branch" directly with the pre-registered S1–S4 metrics — RT-04-T1 (the 2.filter branch test) is merged into it; D50's Virgo re-extractions must reproduce the 2.1.filter convention and report retained time vs 2.filter; the systematic budget keeps the field-level term I034/E027 I034 完成I034 done
GTI 政策（ParetoGTICal 的 2.filter/FLR_z3/MS_z3.5_core/GATE）能为我们保留多少有价值时间、降低多少误差？（E027，PI 2026-10-16）How much valuable time do the GTI policies (ParetoGTICal's 2.filter/FLR_z3/MS_z3.5_core/GATE) keep for us, and how much error do they reduce? (E027, the PI, 2026-10-16) v19 用历史 2.1.filter；若别的策略能在同一样本上给出更小的谱误差，或把更多观测变成可用场，样本与系统误差预算都会变；同时要给 ParetoGTI 的标准设计反馈v19 uses the historical 2.1.filter; if another policy gives smaller spectral errors on the same sample, or turns more observations into usable fields, both the sample and the systematic budget change; and ParetoGTI's criterion design needs the feedback Stage 0/0b（灯光曲线级，已完成）：GATE 不可用（丢 5–6 个主场，含 NW 四场中的 2–3 个，与 ParetoGTICal 不采用它一致）；FLR_z3 与 MS_z3.5_core 清洗效果相同（NW 四场总水平约 −2%、其余约 −0.5%），代价不同（MS 损失 2–3% 曝光，FLR 约 10%）；对 NW 超出的预期效应小（NW−其他差变 −1.4 至 −2.3% 的总水平，最多约 CGMsum 中 ~25% NW 超出的四分之一）；场间 rms 无改善。Stage 0b（31 个观测）：总量掩盖再分配——各策略保留总量相近（none 的 0.61–0.65）但落在不同观测上，主要价值是覆盖（RT-10 复核后：约 4 个有用 + 5 个边缘非 v19 指向）。红队 RT-10 分诊 + 预注册修订 v2（任何谱数据之前）：S2 场间 τ 只作描述（功效太低且误差模型敏感），策略间不排座次；泳道不嵌套 → S4 改为集合差 + 公共核分解；S1 的 CGMsum 与 SP 简并 → 主通道改为软质子归一化与不依赖模型的软波段面亮度。Stage 1a 完成（203 个 GTI 文件，kept live 与 ParetoGTICal 的 K 逐秒相等）；谱级结论待定Stage 0/0b (light-curve level, done): GATE is unusable (it drops 5–6 primaries, including 2–3 of the NW four, agreeing with ParetoGTICal not adopting it); FLR_z3 and MS_z3.5_core clean identically (about −2% on the NW four's gross level, about −0.5% on the others) at different cost (MS costs 2–3% of the exposure, FLR_z3 about 10%); the expected effect on the NW excess is small (the NW−other difference moves by −1.4 to −2.3% of the gross level, at most about a quarter of CGMsum's ~25% NW excess); no improvement in the between-field rms. Stage 0b (31 observations): totals hide the redistribution — the policies keep similar totals (0.61–0.65 of none) but on different observations, so the main value is coverage (after RT-10's review: about 4 useful and 5 marginal non-v19 pointings). Red team RT-10 triaged + pre-registration amendment v2 (before any spectral data): S2's field-to-field τ becomes descriptive only (too little power and error-model dependent), so it does not rank policies; the lanes are not nested → S4 becomes the set difference plus a common-core decomposition; S1 on CGMsum is degenerate with the SP term → the primary channel becomes the soft-proton norm and a model-independent soft-band surface brightness. Stage 1a done (203 GTI files, kept live equal to ParetoGTICal's K to the second); the spectral verdict is Pending E027 stage 1：1a GTI 导出（lang，06:26Z → ~07:10Z）已完成并验收（203 个 GTI 文件）→ 1b Virgo 逐观测逐泳道提取（等 E026 修正后的 runner）→ 1c lang 拟合 → 1d 反馈备忘录 FEEDBACK_ParetoGTI.md。RT-04-T1（2.filter GTI 分支测试）已并入 stage 1E027 stage 1: 1a GTI export (lang, 06:26Z → ~07:10Z) done and accepted (203 GTI files) → 1b per-observation per-lane extraction on Virgo (waits for E026's corrected runner) → 1c fits on lang → 1d the feedback memo FEEDBACK_ParetoGTI.md. RT-04-T1 (the 2.filter GTI-branch test) is merged into stage 1 E027 stage 0/0b + 1a 完成；1b/1c 待定stages 0/0b + 1a done; 1b/1c Pending
南北差异 0.212 ± 0.210（1.01σ）——一致性，不是非对称探测North-minus-South 0.212 ± 0.210 (1.01σ) — consistency, not an asymmetry detection 非对称上限是论文的直接结果之一；两侧半径分布不匹配（中位 19.28 vs 14.97 kpc）The asymmetry limit is one of the paper's direct results; the two sides have unmatched radius distributions (medians 19.28 vs 14.97 kpc) 总和在两侧一致；任何分量级非对称结论都需要 MW 均匀或 M31 对称假设The sum is consistent between sides; any component-level asymmetry claim needs MW-uniformity or M31-symmetry assumptions 共享径向项 + 侧向偏移的联合拟合（I013 提交门槛 1）A joint fit with a shared radial term plus a side offset (I013 submission gate 1) 计划planned 计划planned

来源：cognition-os/02_CURRENT_QUESTION.md（2026-10-14 D48 块）；cognition-os/insights/I033（E022 + E023 段；第 5 条经 RT-01 回到 Questioned）、I006、I012、I017、I034；cognition-os/experiments/E013（A2 v1、toy-MC 与「A2 v2 final」节）、E014、E016、E018a、E019、E020（「Red-team RT-01 triage」节）、E021、E022（correction 1 节）、E023、E024、E025（correction 1 节）、E017a；cognition-os/04_DECISION_LOG.md D49 及其 update 节（撤回）与 D50；cognition-os/REDTEAM_QUEUE.md（RT-02 的 F3 发现）；.agent/LOOP_STATE.md（A1d/RT-02/RT-03/E026 运行与队列）；03_TRUST_TABLE.md 末行。

Source: cognition-os/02_CURRENT_QUESTION.md (2026-10-14 D48 block); cognition-os/insights/I033 (the E022 + E023 paragraph; point 5 back to Questioned after RT-01), I006, I012, I017, I034; cognition-os/experiments/E013 (the A2 v1, toy-MC and "A2 v2 final" sections), E014, E016, E018a, E019, E020 (the "Red-team RT-01 triage" section), E021, E022 (the correction-1 section), E023, E024, E025 (the correction-1 section), E017a; cognition-os/04_DECISION_LOG.md D49 with its update section (withdrawn) and D50; cognition-os/REDTEAM_QUEUE.md (RT-02's F3 finding); .agent/LOOP_STATE.md (the A1d/RT-02/RT-03/E026 runs and the queue); the last rows of 03_TRUST_TABLE.md.

未来 1–4 周计划（依据 D48 与 02_CURRENT_QUESTION）Plan for the next 1–4 weeks (from D48 and 02_CURRENT_QUESTION)

A2 baseline UL vs core radius on A1d
图：A2 基线（A1d 曲面）——集中型 M31 β 模型晕在 R = 20 kpc 的单侧 95% 上限随核半径 r_c 的变化（β = 0.4/0.5/0.6；MW 常数，14 个主场，吸收 0.5–2 keV，单位 1e-15 erg/cm²/s/arcmin²）。水平线为外部预言：Zhang 2024 名义中值 0.83、EAGLE fiducial 0.51、EAGLE AGNdT9 0.26、EAGLE NoAGN 1.08、SIMBA 1.86；灰色粗线为测得的总和 0.965。判定阈值 Δstat = 2.71 尚未用能谱模拟校准（E028 运行中）；红队 RT-13：上限主要靠一个视场，去掉它 Zhang/上限从 5.2 降到 3.3。怎么读这张图：横轴越往右，晕越「铺开」——核半径越大，同样的总亮度在 20 kpc 处的面亮度就越低，所以上限随 r_c 上升；曲线整体落在 Zhang 名义中值线之下，意味着「M31 内晕的热气集中在很小一个核里、总量与模拟的中值星系相当」这种图像被数据排除；而核半径大到约 30–50 kpc 时曲线接近甚至超过预言线，此时数据不再有约束力——一个很平的晕可以藏进总亮度里而不被形状发现。三条曲线（β = 0.4/0.5/0.6）描述的是晕边缘掉得有多快：β 越小边缘越陡，同样总亮度下中心越亮，因而上限越低。 Figure: the A2 baseline (on the A1d surfaces) — the one-sided 95% upper limit on a concentrated M31 beta-model halo at R = 20 kpc versus core radius r_c (β = 0.4/0.5/0.6; MW constant, 14 primary fields, absorbed 0.5–2 keV, units 1e-15 erg/cm²/s/arcmin²). Horizontal lines are external predictions: Zhang 2024 nominal median 0.83, EAGLE fiducial 0.51, EAGLE AGNdT9 0.26, EAGLE NoAGN 1.08, SIMBA 1.86; the thick grey line is the measured total 0.965. The Δstat = 2.71 threshold is not yet calibrated by spectral simulation (E028 running); red team RT-13: the limit leans on one field, and without it Zhang/UL falls from 5.2 to 3.3. How to read it: further right on the horizontal axis the halo is more "spread out" — a larger core radius puts less of the same total brightness at 20 kpc, so the limit rises with r_c; the curves sit below the Zhang nominal-median line, which excludes the picture "M31's inner-halo hot gas is concentrated in a small core with the total brightness of the median simulated galaxy"; while at core radii of about 30–50 kpc the curves approach or cross the prediction lines and the data stop being constraining — a very flat halo can hide inside the total brightness without being found by its shape. The three curves (β = 0.4/0.5/0.6) describe how fast the halo's edge falls: a smaller β means a steeper edge and a brighter centre for the same total, hence a lower limit.

数据：newangels3_discussion/E013_A2_claude/a2_ul_vs_rc_interp_A1d.csv（A2 基线，Claude；RT-13 复核 39 行全部一致）；图由 a2_fig_ul_vs_rc.py 以 A2_SURF/A2_TAG 指向 A1d 曲面生成（a2_ul_vs_rc_A1d.png，已查看）。红队 RT-08/RT-13 见 E013 对应节。

Data: newangels3_discussion/E013_A2_claude/a2_ul_vs_rc_interp_A1d.csv (the A2 baseline, Claude; RT-13 reproduced all 39 rows); the figure made by a2_fig_ul_vs_rc.py with A2_SURF/A2_TAG pointing at the A1d surfaces (a2_ul_vs_rc_A1d.png, viewed). Red teams RT-08/RT-13: the matching sections of E013.

E018a fluxes
图：E018a 的八个场（左：外 NW 四场；右：同期对照四场）的吸收 0.5–2 keV CGMsum 通量。黑点 = 冻结的 v19；蓝方 = 模型 A（v19 重收敛，6 个起始）；红菱 = 模型 B（A + SP 谱形自由）。0301 的流量在近简并处移动（0.975→0.56，Δstat −0.04），0601 在线能量解钉后统计量降 14.7；外四/对照四的逆方差比值从 1.295 降到 1.130（≈1σ）。 Figure: E018a's absorbed 0.5–2 keV CGMsum fluxes for the eight fields (left: the outer-NW four; right: the same-epoch control four). Black = frozen v19; blue squares = Model A (v19 re-converged, 6 starts); red diamonds = Model B (A + free SP shape). 0301's flux moves at a near-degeneracy (0.975→0.56 at Δstat −0.04), and 0601's statistic drops by 14.7 once the line energy unpins; the outer-four/control-four inverse-variance ratio falls from 1.295 to 1.130 (≈1σ).

数据：newangels3_discussion/E018a_luna/fits_per_field.csv（E018a，luna，Claude 已核）；比值 1.295/1.168/1.130 由本站用 v19 权重从该 CSV 复算。图由 handbook/hermes_figscripts/make_fig_e018a.py 生成。

Data: newangels3_discussion/E018a_luna/fits_per_field.csv (E018a, luna, Claude-checked); the 1.295/1.168/1.130 ratios recomputed by this site from that CSV with v19 weights. Figure made by handbook/hermes_figscripts/make_fig_e018a.py.

E021 sky coverage
图：E021 档案清点的天区覆盖（Claude 评审时已查看）。lang 的 data/ 有 145 个 ObsID：class-A（v19 可用）22 个、class-B 114 个、class-C 9 个。class-A 按到 M31 的距离：0–30 kpc 14 个（New-ANGELS 核心样本）、30–60 kpc 0 个、60–100 kpc 3 个、100–150 kpc 1 个、150–200 kpc 3 个、200–300 kpc 1 个——30 kpc 以外只有 8 个，且全部在负银纬。 Figure: the sky coverage of E021's archival inventory (viewed by the reviewer). lang's data/ holds 145 ObsIDs: 22 class A (v19-ready), 114 class B, 9 class C. Class A by distance from M31: 14 at 0–30 kpc (the New-ANGELS core), 0 at 30–60 kpc, 3 at 60–100, 1 at 100–150, 3 at 150–200, 1 at 200–300 — only 8 beyond 30 kpc, all at negative Galactic latitude.

数据：newangels3_discussion/E021_hermes/inventory.csv 与 class_counts.csv（E021，Hermes，已接受）。图来自 newangels3_discussion/E021_hermes/fig_e021_sky_coverage.png（评审已查看），复制入 assets/。

Data: newangels3_discussion/E021_hermes/inventory.csv and class_counts.csv (E021, Hermes, accepted). Figure from newangels3_discussion/E021_hermes/fig_e021_sky_coverage.png (viewed by the reviewer), copied into assets/.

  1. E013 A2 v1 + toy MC（已完成）：谱联合似然在 14 主场的 v19 曲面上插值，对 N_MW、梯度、kT（0.185 keV）做 profile；无 M31 分量被偏好（Δstat ≤ 1.07）。单侧 95% 上限（Δstat = 2.71，插值版）见状态页表；toy-MC 覆盖率 0.975–0.984（略保守），误报率 0.069–0.086，观测上限 = toy 中位数。产出 newangels3_discussion/E013_A2_claude/（a2_ul_vs_rc_interp.csv、a2_toymc_threshold.csv、a2_ul_vs_rc.png）。 E013 A2 v1 + toy MC (done): the spectral joint likelihood interpolated on the 14 primary fields' v19 surfaces, profiling over N_MW, the gradient and kT (0.185 keV); no M31 component is preferred (Δstat ≤ 1.07). One-sided 95% UL (Δstat = 2.71, interpolated) in the state-page table; toy-MC coverage 0.975–0.984 (slightly conservative), false-detection rate 0.069–0.086, observed limits = the toy median. Outputs newangels3_discussion/E013_A2_claude/ (a2_ul_vs_rc_interp.csv, a2_toymc_threshold.csv, a2_ul_vs_rc.png).
  2. E020（已完成；RT-01 已分诊）：样本级重收敛（20 场，模型 A/B/C，多元起始）。模型 C 下 14 场均值 0.964 → 0.819 ± 0.041。红队 RT-01 分解（Claude 复核）：−15.04% 中 v19 重收敛（A）−7.65%、SP 谱形（A→B）−5.29%、线能量（B→C）−2.87%——「SP 谱形单侧 −15% 系统误差」不成立；模型 C 的 SP 谱形未经物理验证；D49 更新撤回。模型 C 曲面（A1c）降级为敏感性变体；A1d（v19 模型多元起始）成为 A2 基线。产出 newangels3_discussion/E020_luna/。 E020 (done; RT-01 triaged): the sample-wide re-convergence (20 fields, Models A/B/C, multi-start). Under Model C the 14-field mean moves 0.964 → 0.819 ± 0.041. Red team RT-01's decomposition (Claude-reproduced): of −15.04%, the v19 re-convergence (A) gives −7.65%, the SP shape (A→B) −5.29%, the line energies (B→C) −2.87% — the "one-sided −15% SP-shape systematic" does not hold; Model C's SP shapes are not physically validated; the D49 update is withdrawn. The Model-C surfaces (A1c) are demoted to a sensitivity variant; A1d (v19-model multi-start) becomes the A2 baseline. Outputs newangels3_discussion/E020_luna/.
  3. E019（已完成）：PSF 翼泄漏的实测——81 个孤立类点源叠加，最亮 bin f_leak 10.3 ± 8.6%（原值）／20.4 ± 7.8%（零点校正）；14 主场泄漏 SB 0.16（原值）／0.33（零点校正）vs E016 模型 0.12。产出 newangels3_discussion/E019_codexnim/。 E019 (done): the empirical PSF-wing leak — a stack of 81 isolated point-like sources, brightest-bin f_leak 10.3 ± 8.6% raw / 20.4 ± 7.8% null-corrected; per-field leak SB over the 14 primaries 0.16 raw / 0.33 null-corrected vs E016's model 0.12. Outputs newangels3_discussion/E019_codexnim/.
  4. E021（已完成）：档案清点与 300 kpc 图设计（无拟合）——145 ObsID，class-A 22 个，30 kpc 以外仅 8 个；产出 newangels3_discussion/E021_hermes/（inventory.csv、DESIGN.md、天区覆盖图）。 E021 (done): the archival inventory and the 300-kpc map design (no fitting) — 145 ObsIDs, 22 class A, only 8 beyond 30 kpc; outputs newangels3_discussion/E021_hermes/ (inventory.csv, DESIGN.md, the sky-coverage figure).
  5. E022（已完成）：8 个外 class-A 档案场（82–253 kpc）的 v19 + 模型 C 拟合（N_H 从 HI4PI 重建，与 v19 差 0.05–0.31 %）。v19 结果被接受为外场参考：外 IVW 0.842 ± 0.065；内禀 norm 外/内 0.80 ± 0.13；模型 C 仅作系统变体（其外场剖面误差 χ² 168/7，不能用于加权平均）。产出 newangels3_discussion/E022_luna/。 E022 (done): v19 + Model-C fits of the 8 outer class-A archival fields (82–253 kpc; N_H rebuilt from HI4PI, agreeing with v19 to 0.05–0.31 %). The v19 results are accepted as the outer reference: outer IVW 0.842 ± 0.065; intrinsic norm outer/inner 0.80 ± 0.13; Model C is a systematic variant only (its outer profile errors give χ² 168/7, unusable for a weighted mean). Outputs newangels3_discussion/E022_luna/.
  6. E023（已完成）：22 个 v19 norm（14 内 + 8 外）的联合 MW 常数 + M31 β 拟合（高斯似然，局部协方差误差）——大核晕在 20 kpc 的单侧 95% 上限 0.46（r_c=50）／0.59（r_c=100）／0.90（r_c=200）（β = 0.5）；Zhang 0.83 被排除到 r_c ≈ 125–175 kpc (β = 0.4/0.5)；无 M31 分量被偏好（Δχ²(A=0) ≤ 1.3）。Claude 已独立复算 r_c = 50/100 的上限，数值一致。产出 newangels3_discussion/E023_luna/limits.csv。下一步：在谱曲面上重做。 E023 (done): a joint MW-constant + M31-beta fit of the 22 v19 norms (14 inner + 8 outer; Gaussian likelihood on local-covariance errors) — the large-core halo's one-sided 95% UL at 20 kpc is 0.46 (r_c=50) / 0.59 (r_c=100) / 0.90 (r_c=200) for β = 0.5; Zhang 0.83 is excluded up to r_c ≈ 125–175 kpc (β = 0.4/0.5); no M31 term is preferred (Δχ²(A=0) ≤ 1.3). Claude independently recomputed the r_c = 50/100 limits and got the same values. Outputs newangels3_discussion/E023_luna/limits.csv. Next: redo it on spectral surfaces.
  7. E017a（已完成）：Paper I 星表与 v19 掩膜的交叉匹配（14 主场，MOS1）——亮源（>1e-14）掩住 95.5 %（233/244，11 个未掩）；8784 个掩膜椭圆里 87 % 无 10″ 内的 Paper I 源（掩膜来自更深的逐观测探测表）。掩膜出处缺口就此闭合。产出 newangels3_discussion/E017a_luna/。 E017a (done): the Paper I catalogue cross-matched against the v19 masks (14 primary fields, MOS1) — 95.5 % of bright sources (>1e-14) masked (233/244, 11 unmasked); of the 8784 mask ellipses 87 % have no Paper I source within 10″ (the masks come from a deeper per-observation detection list). The mask-provenance gap is thereby closed. Outputs newangels3_discussion/E017a_luna/.
  8. V01（已完成）：E020/E022 数值的独立复核（luna，直接从四个 CSV 复算，不用任何先前报告）——20 个标量比较 18 个 PASS；2 个「FAIL」只是 E020 加权径向斜率的舍入差（复算 −0.0005348 vs 声称 −0.0005；0.006051 vs 0.0061；四位小数下一致，超出 0.5 % 严格容差）。产出 newangels3_discussion/V01_luna/verify_table.csv。 V01 (done): an independent verification of the E020/E022 numbers (luna, recomputed directly from the four CSVs, no prior reports as inputs) — 18 of 20 scalar comparisons PASS; the 2 "FAILs" are rounding-only differences in E020's weighted radial slope (recomputed −0.0005348 vs claimed −0.0005; 0.006051 vs 0.0061; identical at four decimals, outside the strict 0.5 % tolerance). Outputs newangels3_discussion/V01_luna/verify_table.csv.
  9. A1c 完成 + A2 v2 final（模型 C 曲面，敏感性变体，2026-10-15 18:26Z）：A1c（codexnim）14/14 场、加密后 117 点/场，门检全过；A2 v2 final 给出插值上限（β = 0.5，r_c = 4/8/15/30 = 0.111/0.138/0.215/0.484；β = 0.4/0.6 在 r_c = 4 为 0.176/0.075），toy MC 1000 次覆盖率 0.918–0.933（低于 0.95）——Δstat = 2.71 在模型 C 曲面上略不保守，校准阈值待做（Claude TODO）；Zhang/上限（r_c ≤ 8）= 3.9–12 倍。产出 newangels3_discussion/E013_A2_claude/（a2_ul_vs_rc_interp_v2C.csv、a2_ul_vs_rc_v2C.png）。RT-01 后定位：敏感性变体，不是基线。 A1c done + A2 v2 final (Model-C surfaces, a sensitivity variant, 2026-10-15 18:26Z): A1c (codexnim) 14/14 fields, 117 points per field after densification, all gates exact; A2 v2 final gives interpolated ULs (β = 0.5, r_c = 4/8/15/30 = 0.111/0.138/0.215/0.484; β = 0.4/0.6 at r_c = 4: 0.176/0.075), with a 1000-toy MC coverage of 0.918–0.933 (below 0.95) — the Δstat = 2.71 rule is slightly anti-conservative on the Model-C surfaces, a calibrated threshold is TODO (Claude); Zhang/UL for r_c ≤ 8 = 3.9–12x. Outputs newangels3_discussion/E013_A2_claude/ (a2_ul_vs_rc_interp_v2C.csv, a2_ul_vs_rc_v2C.png). After RT-01 its status is: a sensitivity variant, not the baseline.
  10. E024 完成（2026-10-15 ~13:2xZ）+ D50 决定 + E025 完成：E024（Hermes）给出 47 个外 B 场的四种政策：P1（遮展源）+12 个方向、P2（放宽到 8–10 ks）累计 +31、P3（加 9 个 Virgo 重提取）+40、P4（加单相机）+43；全部在 b < 0；NGC 147 不可用（仅剩 22% 视场）。PI 决定 D50（2026-10-15）：P1+P2 立即拟合、6 个真正外围场 Virgo 重提取、跳过 P4、弃 NGC 147。E025（luna）照此执行：P1 12/12 掩膜审计未过 → 全部转重提取；P2 18/19 接受；干净外样本（13 场）0.846 ± 0.059，外/内 = 0.877 ± 0.072；correction 1 把 10 个 P2 a+b 场也判为掩膜不足 → 重提取清单 6 + 12 + 10 = 28 场。产出 newangels3_discussion/E024_hermes/、newangels3_discussion/E025_luna/。 E024 done (2026-10-15 ~13:2xZ) + the D50 decision + E025 done: E024 (Hermes) laid out four policies for the 47 outer class-B fields: P1 (mask extended targets) +12 directions, P2 (relax to 8–10 ks) +31 cumulative, P3 (add 9 Virgo re-extractions) +40, P4 (add one-camera fields) +43; all at b < 0; NGC 147 unusable (22% of the field of view left). The PI decided (D50, 2026-10-15): fit P1+P2 now, re-extract 6 genuinely-outer fields on Virgo, skip P4, drop NGC 147. E025 (luna) executed it: all 12 P1 fields fail the mask audit → all to re-extraction; 18/19 P2 fits accepted; the clean outer sample (13 fields) 0.846 ± 0.059, outer/inner = 0.877 ± 0.072; correction 1 found the 10 P2 a+b fields also under-masked → the re-extraction list is 6 + 12 + 10 = 28 fields. Outputs newangels3_discussion/E024_hermes/, newangels3_discussion/E025_luna/.
  11. A1d 完成 + A2 基线在它上面重算（2026-10-16 13:2xZ；临时，阈值未校准）：A1d（codexnim，18:50Z → ~11:20Z，65.7 worker-小时）14/14 场 × 117 点 = 1638/1638，0 个拟合失败，v19 与 Model-A 恢复门检 14/14。A2 基线（同一估计器：MW 常数、kT 在最佳共温 ±1 行内 profile、Δstat = 2.71、插值）给出 20 kpc 处单侧 95% 上限：核半径 ≤ 8 kpc 时 0.12–0.16（β 0.4）／0.075–0.10（β 0.5）／0.049–0.071（β 0.6）；r_c 30 为 0.30–0.56，r_c 50 为 0.68–1.26。插值收敛（RT-02 的 O9）：线性/样条上限比 0.77–0.82（A1 上 0.43–0.50），仍差于 10% 目标，取较松的样条值。Zhang 名义中值 0.8285 / 上限（r_c ≤ 8）= 5.2–6.7（β 0.4）、8.1–11.1（β 0.5）、11.7–17.1（β 0.6）。红队 RT-13 已分诊：算术与输入成立、2.71 是「参数不能为负」边界情形下正确的分位数；上限主要靠一个视场（0800730601）——去掉它放宽 +46…+62%，措辞改为「全样本 ≳4 倍；去掉约束最强的视场仍 ≥3 倍」；报出的自然样条值是四种插值里最松的（其余紧 2–23%）。红队 RT-08 已分诊：toy-MC 的位置族校准只在简化模型下有效，「上限已校准良好」不成立——阈值改用能谱级模拟校准（E028）。产出 newangels3_discussion/E013_A2_claude/（a2_ul_vs_rc_interp_A1d.csv、a2_ul_vs_rc_interp_A1d_linear.csv、a2_ul_vs_rc_A1d.png）。 A1d done + the A2 baseline recomputed on it (2026-10-16 13:2xZ; provisional, threshold NOT calibrated): A1d (codexnim, 18:50Z → ~11:20Z, 65.7 worker-hours) gives 14/14 fields × 117 points = 1638/1638, 0 failed fits, with the v19 and Model-A restore gates 14/14. The A2 baseline (the same estimator: MW constant, kT profiled over the best shared kT ±1 row, Δstat = 2.71, interpolated) gives one-sided 95% ULs at 20 kpc: for core radius ≤ 8 kpc, 0.12–0.16 (β 0.4) / 0.075–0.10 (β 0.5) / 0.049–0.071 (β 0.6); r_c 30 gives 0.30–0.56 and r_c 50 gives 0.68–1.26. Interpolation convergence (RT-02's O9): the linear/spline UL ratio is 0.77–0.82 (0.43–0.50 on A1), still outside the 10 % target, so the looser spline values are quoted. Zhang's nominal median 0.8285 / UL for r_c ≤ 8 = 5.2–6.7 (β 0.4), 8.1–11.1 (β 0.5), 11.7–17.1 (β 0.6). Red team RT-13 triaged: the arithmetic and inputs hold and 2.71 is the correct quantile for a parameter bounded at zero; the limit leans on one field (0800730601) — leaving it out loosens by +46…+62 %, so the wording becomes "≳4× for the full sample; still ≥3× without the most constraining field"; the quoted natural-spline UL is the loosest of four interpolants (the others 2–23 % tighter). Red team RT-08 triaged: the toy-MC location-family calibration holds only under its simplified model, so "the limits are well calibrated" does not survive — the threshold is to be calibrated by spectrum-level simulation (E028). Outputs newangels3_discussion/E013_A2_claude/ (a2_ul_vs_rc_interp_A1d.csv, a2_ul_vs_rc_interp_A1d_linear.csv, a2_ul_vs_rc_A1d.png).
  12. E027 stage 0/0b + 1a 完成（2026-10-16 06:0xZ–07:10Z）；1b/1c 待定：PI 2026-10-16 两次要求——先问 GATE/FLR_z3/MS_z3.5_core 对我们结果的改进，随后把范围改为全部 31 个 LP 080073 观测（保留多少有价值时间、降低多少误差、如何反馈 ParetoGTI 标准设计）。Stage 0/0b（灯光曲线级）：GATE 不可用（丢 5–6 个主场）；FLR_z3 与 MS_z3.5_core 清洗相同、代价不同（2–3% vs ~10% 曝光）；主要价值是覆盖（RT-10 复核后：约 4 个有用 + 5 个边缘非 v19 指向）。红队 RT-10 分诊 + 预注册修订 v2（任何谱数据之前）：S2 场间 τ 只作描述；泳道不嵌套 → S4 改为集合差 + 公共核分解；S1 的 CGMsum 与 SP 简并 → 主通道改为软质子归一化与不依赖模型的软波段面亮度。Stage 1a 完成并验收（codexnim，06:26Z → ~07:10Z）：203 个 GTI FITS 带 sha256 manifest，kept live 与 ParetoGTICal per_obs K 逐秒相等。1b Virgo 提取等 E026 修正后的 runner；1c 拟合；1d 反馈备忘录。RT-04-T1 已并入。谱级结果待定。 E027 stages 0/0b + 1a done (2026-10-16 06:0xZ–07:10Z); 1b/1c Pending: the PI's two 2026-10-16 requests — first about GATE/FLR_z3/MS_z3.5_core for the improvement to our results, then rescoped to all 31 LP 080073 observations (how much valuable time is kept, how much error is reduced, and how to feed back into ParetoGTI criterion design). Stages 0/0b (light-curve level): GATE is unusable (it drops 5–6 primaries); FLR_z3 and MS_z3.5_core clean identically at different cost (2–3% vs ~10% of the exposure); the main value is coverage (after RT-10's review: about 4 useful and 5 marginal non-v19 pointings). Red team RT-10 triaged + pre-registration amendment v2 (before any spectral data): S2's field-to-field τ becomes descriptive only; the lanes are not nested → S4 becomes the set difference plus a common-core decomposition; S1 on CGMsum is degenerate with the SP term → the primary channel becomes the soft-proton norm and a model-independent soft-band surface brightness. Stage 1a done and accepted (codexnim, 06:26Z → ~07:10Z): 203 GTI FITS with a sha256 manifest, kept live equal to ParetoGTICal's per_obs K to the second. Stage 1b's Virgo extraction waits for E026's corrected runner; then 1c fits and 1d the feedback memo. RT-04-T1 is merged in. Spectral results Pending.
  13. 运行中，待定：E028（stage 0：用能谱模拟校准 A2 的 Δstat = 2.71 阈值，codexnim 13:21Z 启动）与红队 RT-15（审 E028 设计，codexnim 15:12Z）、RT-14-T1（只用响应文件的 CCF 审计，codexnim 14:17Z）、RT-16（A2 为何如此依赖 0601 场，Hermes 15:47Z）；E026（28 场 Virgo 重提取——stage 0 已完成，Virgo 无 ODF 需 NXSA；红队 RT-11 已分诊：runner 会让全部 28 个场在审计这一步失败，修正单 CORRECTION_01_RT11.md 正在试点场上应用，通过前不批处理）；E027 stage 1b/1c（谱级提取与拟合，等 E026 修正后的 runner）；RT-09-T3（早/晚配对重提取，排队）。红队 RT-01…RT-14 已全部分诊完毕（见组会页）。结果出来前不引用任何数字。 Running, Pending: E028 (stage 0: calibrating the A2 Δstat = 2.71 threshold with spectral simulation, codexnim, launched 13:21Z) and red teams RT-15 (reviewing the E028 design, codexnim 15:12Z), RT-14-T1 (a response-only CCF audit, codexnim 14:17Z) and RT-16 (why the A2 limit leans so hard on field 0601, Hermes 15:47Z); E026 (28-field Virgo re-extraction — stage 0 done, no ODFs on Virgo so NXSA is needed; red team RT-11 is triaged: the runner would fail all 28 fields at the audit step, the correction list CORRECTION_01_RT11.md is being applied on the pilot field and no batch starts before it passes); E027 stage 1b/1c (spectral extraction and fits, waiting for E026's corrected runner); RT-09-T3 (the early/late paired re-extraction, queued). Red teams RT-01…RT-14 are all triaged (see the meeting page). No numbers quoted until results land.
  14. 其后：等 E028 用能谱模拟校准 Δstat = 2.71 阈值，再决定插值终值（PCHIP 或显式节点合并，RT-13 的 O3）并引用校准后的收敛上限；在谱曲面上以流量空间重做 E023（加 MW 纬度项与外场系统项）；E026 修正通过后批处理 28 场，按 v19 规则拟合并在扩大的内外样本上做联合拟合；E027 stage 1 的谱结果与 ParetoGTI 反馈备忘录；RT-14-T1/RT-09-T3 判定外场年代混杂的原因；E015（RASS 25° 原生计数图）；E018b 保持可选（I034：预期变化不大）。 After that: wait for E028 to calibrate the Δstat = 2.71 threshold with spectral simulation, then settle the interpolant (PCHIP or an explicit node-merge rule, RT-13's O3) and quote the calibrated, converged UL; redo E023 in flux space on spectral surfaces (with an MW latitude term and an outer systematic); once E026's correction passes, batch the 28 fields, fit them under v19 rules and run the joint fit on the enlarged inner+outer sample; E027 stage 1's spectral results and the ParetoGTI feedback memo; RT-14-T1/RT-09-T3 decide the cause of the outer epoch confound; E015 (the 25° native-count RASS map); E018b stays optional (I034: expected to change little).
  15. E015（排队）：M31 周围 25° 的 RASS 原生计数图（MW 纬度模型 + M31 中心项 + LHB 模板），为本地 MW 模型与 Local Hot Bridge 检验（O4）铺路。 E015 (queued): a 25° native-count RASS map around M31 (MW latitude model + an M31-centred term + an LHB template), paving the way for the local MW model and the Local Hot Bridge test (O4).
  16. E017b（低优先，排队）：只用 Paper I 星表重建源掩膜（需 Virgo 重提取），对比 v19 的 skyarea 与通量——E017a 的交叉匹配已把掩膜出处缺口闭合，此项仅在审稿人要求时做。 E017b (low priority, queued): rebuild the source mask from the Paper I catalogue only (needs re-extraction on Virgo) and compare skyarea and flux with v19 — E017a's cross-match has closed the mask-provenance gap, so this runs only if a referee asks.
  17. 论文讨论：把 I033 的科学线（总和、轮廓形状——含大核限制、NW 起源）与 E016/E019/E020（RT-01 后作敏感性变体）/E022/E023/E025/E017a 的新结果写进 New-ANGELS III 草稿的讨论部分；提交门槛仍按 I013。 Paper discussion: write the I033 science lines (total; profile shape — now including the large-core limit; NW origin) plus the new E016/E019/E020 (a sensitivity variant after RT-01)/E022/E023/E025/E017a results into the New-ANGELS III draft discussion; submission gates remain those of I013.

来源：cognition-os/00_AI_BRIEFING.md（2026-10-14 D48 与 2026-10-15 13Z update）；cognition-os/experiments/E013（A2 v1、toy-MC 与「A2 v2 final」节）、E019、E020（「Red-team RT-01 triage」节）、E021、E014、E016、E018a、E022（correction 1 节）、E023、E024、E025（correction 1 节）、E017a；cognition-os/insights/I013、I034；.agent/jobs/2026-10-15-v01-verify-e020-e022/REPORT.luna.md；cognition-os/04_DECISION_LOG.md D49（update 节已撤回）与 D50；cognition-os/REDTEAM_QUEUE.md；.agent/LOOP_STATE.md（A1d/RT-02/RT-03/E026 运行与队列）。

Source: cognition-os/00_AI_BRIEFING.md (2026-10-14 D48 and the 2026-10-15 13Z update); cognition-os/experiments/E013 (the A2 v1, toy-MC and "A2 v2 final" sections), E019, E020 (the "Red-team RT-01 triage" section), E021, E014, E016, E018a, E022 (the correction-1 section), E023, E024, E025 (the correction-1 section), E017a; cognition-os/insights/I013, I034; .agent/jobs/2026-10-15-v01-verify-e020-e022/REPORT.luna.md; cognition-os/04_DECISION_LOG.md D49 (its update section withdrawn) and D50; cognition-os/REDTEAM_QUEUE.md; .agent/LOOP_STATE.md (the A1d/RT-02/RT-03/E026 runs and the queue).

本次更新What's new

来源：cognition-os/experiments/E013（「A1d accepted and A2 baseline on A1d」「Red-team RT-08 triage」「Red-team RT-13 triage」节；历史条目的 A2 v1、toy-MC 与「A2 v2 final」节）、E019、E020（「Red-team RT-01 triage」节）、E021、E018a、E016、E022（correction 1 节）、E023、E024、E025（correction 1 节、RT-09/RT-09-T1/RT-12/RT-09-T2/RT-14 与 Hermes 臂修订节）、E027（stage 1a、RT-10 修订 v2 节）、E028、E017a；.agent/jobs/2026-10-15-e026-virgo-reextract-stage0/CORRECTION_01_RT11.md；.agent/jobs/2026-10-15-v01-verify-e020-e022/REPORT.luna.md；cognition-os/04_DECISION_LOG.md D48、D49（update 节已撤回）与 D50；cognition-os/insights/I033、I034；cognition-os/REDTEAM_QUEUE.md；cognition-os/06_AGENT_LOG.md 181–196 行；handbook/live_updates.jsonl L001–L008。

Source: the "A1d accepted and A2 baseline on A1d", "Red-team RT-08 triage" and "Red-team RT-13 triage" sections of cognition-os/experiments/E013 (plus the A2 v1, toy-MC and "A2 v2 final" sections for the historical entries), E019, E020 (the "Red-team RT-01 triage" section), E021, E018a, E016, E022 (the correction-1 section), E023, E024, E025 (the correction-1 section and the RT-09/RT-09-T1/RT-12/RT-09-T2/RT-14 sections plus the Hermes-arm amendment), E027 (the stage-1a and RT-10-amendment-v2 sections), E028, E017a; .agent/jobs/2026-10-15-e026-virgo-reextract-stage0/CORRECTION_01_RT11.md; .agent/jobs/2026-10-15-v01-verify-e020-e022/REPORT.luna.md; cognition-os/04_DECISION_LOG.md D48, D49 (its update section withdrawn) and D50; cognition-os/insights/I033, I034; cognition-os/REDTEAM_QUEUE.md; rows 181–196 of cognition-os/06_AGENT_LOG.md; handbook/live_updates.jsonl L001–L008.

数据截至Data as of

2026-10-16T16:56Z（最新读到的来源：cognition-os/06_AGENT_LOG.md 181–196 行（最后一条为 16:56Z「handbook-stage9-failed, stage10-relaunched」；本站刷新即其后续）；cognition-os/experiments/E013_joint_spatial_profile_likelihood.md（「A1d accepted and A2 baseline on A1d」「Red-team RT-08 triage」「Red-team RT-13 triage」节）；cognition-os/experiments/E025_outer_classB_fits.md（RT-09、RT-09-T1、RT-12、RT-09-T2、RT-14 与 15:4xZ 的 Hermes 臂修订节）；cognition-os/experiments/E027_gti_policy_screen.md（stage 1a、RT-10 修订 v2 节）；cognition-os/experiments/E028_a2_spectral_calibration.md；cognition-os/insights/I033 末两段；cognition-os/03_TRUST_TABLE.md 第 73、86、87、89 行；cognition-os/REDTEAM_QUEUE.md 第 15–30 行；.agent/jobs/2026-10-15-e026-virgo-reextract-stage0/CORRECTION_01_RT11.md；handbook/live_updates.jsonl L001–L008newest source read: rows 181–196 of cognition-os/06_AGENT_LOG.md (the last being 16:56Z "handbook-stage9-failed, stage10-relaunched"; this site's refresh is its continuation); the "A1d accepted and A2 baseline on A1d", "Red-team RT-08 triage" and "Red-team RT-13 triage" sections of cognition-os/experiments/E013_joint_spatial_profile_likelihood.md; the RT-09, RT-09-T1, RT-12, RT-09-T2 and RT-14 sections plus the 15:4xZ Hermes-arm amendment of cognition-os/experiments/E025_outer_classB_fits.md; the stage-1a and RT-10-amendment-v2 sections of cognition-os/experiments/E027_gti_policy_screen.md; cognition-os/experiments/E028_a2_spectral_calibration.md; the last two paragraphs of cognition-os/insights/I033; rows 73, 86, 87 and 89 of cognition-os/03_TRUST_TABLE.md; rows 15–30 of cognition-os/REDTEAM_QUEUE.md; .agent/jobs/2026-10-15-e026-virgo-reextract-stage0/CORRECTION_01_RT11.md; handbook/live_updates.jsonl L001–L008）