已经确立的

What is established

下面每一条都有已验证的记录支撑（执行者跑完、Claude 复算或红队复核过）。"确立"不等于"解决"：它们是可引用的测量，不是项目的终点。

Every item below rests on a verified record (executor ran it, Claude recomputed it, or a red team checked it). "Established" does not mean "solved": these are quotable measurements, not the project's finish line.

1. 考试集建成并冻结：7 个 problem（固定 4′ 天域）、18 个 ObsID、45 个 case，40 PASS、5 EXCLUDED（干净曝光 < 20 ks 被排除，D020）。这是后续所有分数的基准面。The exam set is built and frozen: 7 problems (fixed 4′ cells), 18 ObsIDs, 45 cases, 40 PASS, 5 EXCLUDED (clean exposure < 20 ks, D020). Every later score is computed on it.
2. 边缘变亮是真实存在的：CDFS P0 基线 \(\tau/\bar{I}=0.246\)（own 掩膜）、M31 P0 \(\tau/\bar{I}=0.14\)；CDFS 径向诊断里边缘/内部净面亮度中位数 \(\times1.53\)（63 个 case）。Edge brightening is real: CDFS P0 baseline \(\tau/\bar{I}=0.246\) (own mask), M31 P0 \(\tau/\bar{I}=0.14\); in the CDFS radial diagnostic the median edge/inner net surface brightness is \(\times1.53\) (63 cases).
3. M31 的两个"变亮"测量都站得住：配对斜率（仅 QPB）\(+1.51\pm0.16\ \%/\text{arcmin}\)（逐观测剔除后 \(+1.19\) 到 \(+2.12\)，全部 \(\ge\) 6\(\sigma\)）；平衡边缘步长（8–10′ 对 12–14′）\(+9.9\pm2.0\ \%\)（结构化天空零假设偏置 \(-0.01\ \%\)，MC sd/\(\sigma=0.98\)）。Both M31 "brightening" measurements stand: the pair slope (QPB only) \(+1.51\pm0.16\ \%/\text{arcmin}\) (leave-one-observation-out \(+1.19\) to \(+2.12\), all \(\ge\) 6\(\sigma\)); the balanced edge step (8–10′ vs 12–14′) \(+9.9\pm2.0\ \%\) (structured-sky null bias \(-0.01\ \%\), MC sd/\(\sigma=0.98\)).
4. SP 模板决定答案：同一 M31 数据，无 SP 扣除 \(+9.49\pm1.20\ \%\)，flat SP \(+0.10\ \%\)，光子 vignetting SP \(+15.20\ \%\)，ESAS flare 图（SP 自己的模板）\(+6.69\pm1.72\ \%\)。The SP template decides the answer: the same M31 data gives no-SP \(+9.49\pm1.20\ \%\), flat SP \(+0.10\ \%\), photon-vignetted SP \(+15.20\ \%\), ESAS flare map (SP's own template) \(+6.69\pm1.72\ \%\).
5. SP 自己的空间分布（硬带 2–5 keV，真角区示踪）：原始 SP 计数每像素从 \(2'–6'\) 到 \(12'–14'\) 下降 \(0.88\pm0.13\)；ESAS det 图给出 0.76（拟合最好，\(\chi^2_\nu=1.74\)），均匀 2.41，光子 vignetting 4.98（被排除）。SP's own spatial distribution (hard band 2–5 keV, true out-of-FoV corner tracer): raw SP counts per pixel fall \(0.88\pm0.13\) from \(2'–6'\) to \(12'–14'\); the ESAS det map gives 0.76 (best fit, \(\chi^2_\nu=1.74\)), flat 2.41, photon vignetting 4.98 (excluded).
6. MOS1/MOS2 的 SP 归一比在 2012 年出现台阶：\(0.68\pm0.03\)（2002–2012）\(\to 1.47\pm0.11\)（2015–2019）；面积解释被排除；台阶与 MOS1 CCD3 丢失（2012-12-11）完全退化，也与 QPB 模板年代（库止于 2012-01-23）退化。红队把 7\(\sigma\) 修正为含场偏移的 3.1\(\sigma\)。The MOS1/MOS2 SP-normalisation ratio steps in 2012: \(0.68\pm0.03\) (2002–2012) \(\to\) \(1.47\pm0.11\) (2015–2019); the area explanation is refuted; the step is perfectly degenerate with the MOS1 CCD3 loss (2012-12-11) and with the QPB-template epoch (library ends 2012-01-23). The red team corrected 7\(\sigma\) to 3.1\(\sigma\) with field offsets.
7. LMC 0690744801 的边缘超额是仪器/处理，不是天空也不是 SP：\(+21.8\pm4.3\ \%\)；预算排除了曝光/vignetting 图（~1 % 对所需的 21.8 %）、QPB 角区形状（需 +151 %）、SP（需 +168/+273 %）；机制候选只剩掩膜/CCD 边缘（2.9 % 的 MOS2-only 边缘像素，+44 %）。The LMC 0690744801 rim excess is instrumental/processing, not sky and not SP: \(+21.8\pm4.3\ \%\); the budget rules out the exposure/vignetting map (~1 % against the needed 21.8 %), the QPB rim shape (would need +151 %), SP (+168/+273 %); the surviving mechanism candidate is mask/CCD-edge (2.9 % MOS2-only rim pixels, +44 %).
8. 重复指向测不出 SP 的绝对水平：最好的重复对（0784000101/0201，相距 0.04′）只是上限（2\(\sigma\) 到达 0.99/0.55 \(\times10^{-4}\) cts/s/arcmin\(^2\)，超过 M31 谱 SP 总体散布 0.73/0.45）；测量受 QPB 协方差限制，不受计数限制。Repeat pointings cannot measure the absolute SP level: the best repeat pair (0784000101/0201, 0.04′ apart) is an upper limit only (2\(\sigma\) reach 0.99/0.55 \(\times10^{-4}\) cts/s/arcmin\(^2\), exceeding the M31 spectral-SP population spread 0.73/0.45); the test is limited by the QPB covariance, not by counts.
9. 叠加规则：2–7 keV 天空可以叠加谱（\(\le\) 3 % rms）；软带天空只能来自逐观测联合拟合——叠加软带天空 = 联合天空 + SP 折衷残差，且没有安全的 SP 散布阈值。The stacking rule: the 2–7 keV sky can be stacked (\(\le\) 3 % rms); the soft-band sky comes only from joint per-observation fits — a stacked soft-band sky is the joint sky plus an SP-compromise residual, and no safe SP-spread threshold exists.
10. 相机一致的边缘估计：用同相机 LOO 参考后，M31 边缘步长与"SP 在谱率水平"一致（samecam \(-1.16\pm1.67\)、expmatch \(+1.90\)，两者符号相反、都不显著）——这是一个条件性非探测（在假设的 SP 归一下），不是"不存在"的证明（DL 上限 +6.1/+9.8 %）。The camera-consistent edge estimate: with same-camera LOO references, the M31 edge step is consistent with SP at its spectral rate (samecam \(-1.16\pm1.67\), expmatch \(+1.90\), opposite signs, neither significant) — a conditional non-detection at the assumed SP normalisation, not a proof of absence (DL upper limits +6.1/+9.8 %).
11. 逐成分的全局相机标度不够：三个全局 MOS1/MOS2 天空标度（LB 1.108、MW halo 1.050、热晕 0.728）给出 \(\Delta\chi^2=368\)（3 参数），但对相机无关天空仍差 \(\Delta\chi^2=603\)（8/15 观测 > 7.8）；热晕比非常数（\(\chi^2=114/13\) dof）；残差集中在 0.7–2 keV，主要在 MOS2。Per-component global camera scales are not enough: three global MOS1/MOS2 sky scales (LB 1.108, MW halo 1.050, hot halo 0.728) give \(\Delta\chi^2=368\) for 3 parameters, but still \(\Delta\chi^2=603\) against camera-independent sky (8/15 obs > 7.8); the hot-halo ratio is non-constant (\(\chi^2=114/13\) dof); residuals peak at 0.7–2 keV, mostly MOS2.

来源：Source: MosaicSelfCal/results/exam_v4_r240/SUMMARY.md, MosaicSelfCal/results/cdfs_pipeline_comparison_v1.csv, cognition-os/PITFALLS.md （RT-M29/RT-M31/RT-M32/RT-M33/R17c 条目）(the RT-M29/RT-M31/RT-M32/RT-M33/R17c entries), MosaicSelfCal/RESEARCH_PLAN.md （R18c/R18e/R18b2/R20i/R20n/R20c 行）(the R18c/R18e/R18b2/R20i/R20n/R20c rows), cognition-os/03_TRUST_TABLE.md

仍然开放的

What is still open

开放问题按"离 PI 三步目标多远"排序；每条给出下一个检验（详见编号计划）。

Open problems, ordered by distance from the PI's three-step goal; each names its next test (see the numbered plan).

1. 边缘变亮的归因：SP、QPB、仪器响应各占多少？SP 模板选择就能让 M31 步长在 \(+0.1\) 到 \(+15.2\ \%\) 之间移动；SP 的绝对水平又测不出来（见上）。这是三步目标的第 3 步，也是最差观感所在。下一个检验：把 SP 水平、QPB 协方差、SP 模板合并成每观测 SP 产品（计划第 4 步）。Attribution of the edge brightening: how much is SP, QPB, instrument response? The SP template choice alone moves the M31 step between \(+0.1\) and \(+15.2\ \%\); and the absolute SP level is not measurable (above). This is step 3 of the PI goal and the worst visual failure. Next test: merge the SP level, QPB covariance and SP template into one per-observation SP product (plan step 4).
2. SP 软带空间形状：边缘变亮所在的 0.4–1.25 keV 里，SP 的分布测不到（硬度示踪在软带无信号；QPB 只占软带计数的 18 %）。硬带（2–5 keV）已测：接近 ESAS det 图。SP's soft-band spatial shape: in 0.4–1.25 keV, where the edge brightening lives, SP's distribution is not measurable (hardness tracers have no soft-band signal; the QPB is only 18 % of the soft rate). The hard band (2–5 keV) is measured: close to the ESAS det map.
3. 2012 相机-SP 台阶的原因：MOS1 CCD3 丢失与 QPB 模板年代完全退化，现有数据分不开；R20m 检验了 QPB 库年代（早期 vs 晚期库），R 移动 \(\le\) 0.06，解释不了 0.5–0.7 的台阶；sky linking 是更大的敏感度来源（LMC \(\Delta\chi^2=769\)）。The cause of the 2012 camera-SP step: the MOS1 CCD3 loss and the QPB-template epoch are perfectly degenerate and the data cannot separate them; R20m tested the QPB library epoch (early- vs late-only rows), R moves \(\le\) 0.06, which does not explain the 0.5–0.7 step; sky linking is the larger sensitivity (LMC \(\Delta\chi^2=769\)).
4. QPB 的软带 \(\theta\) 形状：硬带（9.5–12 keV）obj/QPB 对 \(\theta\) 平或下降（每个场/相机 \(-0.9\) 到 \(0.0\ \%/\text{arcmin}\)），排除了硬带 QPB \(\theta\) 标度；但软带里 QPB 项只占计数的 \(\le3\ \%/\text{arcmin}\)，藏在 vignetting 主导的比值下，至今未测。The QPB's soft-band \(\theta\) shape: in the hard band (9.5–12 keV) obj/QPB is flat or falling with \(\theta\) (\(-0.9\) to \(0.0\ \%/\text{arcmin}\) in every field/camera), excluding a hard-band QPB \(\theta\)-scale; but in the soft band a QPB term is only \(\le 3\ \%/\text{arcmin}\) of the counts, hidden under a vignetting-dominated ratio — untested.
5. 两台相机为什么不同意：全局逐成分标度不够（见上）；R20o 正在检验"一个每观测的、sky 与 SP 共享的 MOS1/MOS2 因子"。Why the two cameras disagree: global per-component scales are not enough (above); R20o is testing one per-observation MOS1/MOS2 factor shared by sky and SP.
6. pn 能否用：QPB 高估 2–29 %（I014），修正候选是 FoV 12–14 keV 归一化；R5b 的 pn 减数据 6/18 同步，仍慢。Can pn be used: its QPB is over-normalised by 2–29 % (I014); the correction candidate is the FoV 12–14 keV normalisation; R5b's pn reduction has synced 6/18 and is still slow.
7. SAS 22.1 能否用于生产：v1.8.0-dev.5 代码已过 91 测试 + 6 子测试（commit 431a054），但 SAS22 的 -back.pi STAT_ERR/TUNIT 缺陷（tool_iteration 45）修复后还需在 Virgo 重导出受影响臂并做 apple-to-apple（D026）。Can SAS 22.1 be used in production: the v1.8.0-dev.5 code passed 91 tests + 6 subtests (commit 431a054), but after the SAS22 -back.pi STAT_ERR/TUNIT defects (tool_iteration 45) the affected arms still need re-derivation on Virgo and the apple-to-apple comparison (D026).
8. 0690744801 的机制：候选 a（曝光/vignetting）、b（QPB 角区）、c（姿态）、e（SP）已关闭；d（掩膜/CCD 边缘）只在 2.9 % 的 MOS2-only 边缘像素里存活（+44 %）；自/LOO 剖面是 7–14′ 上的缓升，不是纯边缘环。The 0690744801 mechanism: candidates a (exposure/vignetting), b (QPB corner shape), c (attitude), e (SP) are closed; d (mask/CCD-edge) survives only in 2.9 % MOS2-only rim pixels (+44 %); the self/LOO profile is a gradual rise over 7–14′, not a rim-only ring.

来源：Source: cognition-os/08_RUNNING.md §7, MosaicSelfCal/RESEARCH_PLAN.md （R18b2/R12b/R20m/R20o/R5b/R15 行）(the R18b2/R12b/R20m/R20o/R5b/R15 rows), cognition-os/PITFALLS.md （RT-M31 条目）(the RT-M31 entry), cognition-os/tool_iteration.md 45

我们对每一条有多确信

How sure we are of each

信任分级沿用项目的定义（cognition-os/00_AI_BRIEFING.md）：Trusted（规则或已验证的事实）/ Provisional（有记录但未过红队或只有单臂）/ Questioned（红队提出实质异议）/ Deprecated（被推翻）。下表按此定义逐行核对（D27）。

Trust follows the project's definitions (cognition-os/00_AI_BRIEFING.md): Trusted (rules or verified facts) / Provisional (recorded but not red-teamed, or single-arm) / Questioned (a red team raised a material objection) / Deprecated (overturned). Each row below was re-checked against its definition (D27).

来源：Source: cognition-os/00_AI_BRIEFING.md （分级定义）(tier definitions), cognition-os/REDTEAM_QUEUE.md （各 RT 行）(the RT rows), cognition-os/08_RUNNING.md §7 item 1

数字总表

The number table

本页与工作台页出现过的每个数字，附来源路径与出处行。复算过的标注"复算"。

Every number that appears on this page and the workbench page, with its source path and the line it came from. "Recomputed" marks numbers this handbook recalculated.

来源：Source: 各行的来源列。the Source column of each row.