第一阶段：立项与继承（2026-10-05 起）

Stage 1: scoping and inheritance (from 2026-10-05)

项目从两个已有成果上起步（D001）：MosaicSelfCal（重叠自校准）与 MaskingMethodDemo（掩膜方法演示）。同日 D003 定下方法论基调：拟合/减数据契约 T01–T07 原样继承，本项目检验它们而不是重新推导——校准必须测量生产流水线本身，不是它的变体。D006 定下分工：执行交给 Codex，Claude 规划与评审。

The project started on two existing assets (D001): MosaicSelfCal (overlap self-calibration) and MaskingMethodDemo. The same day, D003 set the methodological tone: the fitting/reduction contracts T01–T07 are inherited as-is; this project tests them rather than re-derives them — the calibration must measure the production pipeline, not a variant. D006 set the division of labour: execution to Codex, planning and review to Claude.

来源：Source: cognition-os/04_DECISION_LOG.md D001/D003/D006

第二阶段：把诊断变成考试（2026-10-06）

Stage 2: from diagnosis to an exam (2026-10-06)

第一轮诊断（E001-E005）确认了症状并排除了几个候选：CDFS 63 个 case 全部向外变亮，边缘/内部净面亮度中位数 1.54\(\times\)（12–14′ 对 0–4′）；超额跟着探测器不跟天空（LOO 残差 +0.19/10′）；候选 QPB 欠归一或残余 SP。I010 关键转折：边缘超额在四次排除后依然存在——不是 SP 单一解释、不是仪器线、不是未分辨 CXB（每个候选都有对应检验）。这把问题从"找到元凶"变成"建一个可打分的考试"。

The first diagnostic round (E001–E005) confirmed the symptom and eliminated candidates: in CDFS all 63 cases brighten outward, edge/inner median net SB 1.54\(\times\) (12–14′ vs 0–4′); the excess follows the detector, not the sky (LOO residual +0.19/10′); candidates were QPB under-normalisation or residual SP. The I010 turning point: the edge excess survives four eliminations — not a single-SP explanation, not the instrumental lines, not the unresolved CXB (each candidate had its test). This turned the question from "find the culprit" into "build a scorable exam".

同日 D005/D007 冻结了基准测试：固定天域 + 覆盖它的 ObsID 集合，谱级答案只给"必须相同的天空成分"打分（CXB 每例自由，吸收掩膜深度差异）。

The same day, D005/D007 froze the benchmark: a fixed sky cell plus the ObsIDs covering it, with a spectral answer that scores only the sky components that must be the same (CXB free per case, absorbing mask-depth differences).

来源：Source: MosaicSelfCal/cognition-os/insights/I008, I010, MosaicSelfCal/BENCHMARK.md, cognition-os/04_DECISION_LOG.md D005/D007

第三阶段：工具与子项目（2026-10-10 至 13）

Stage 3: tools and sub-projects (2026-10-10 to 13)

D011 注册了 XMM–EP-FXT 交叉校准子项目（FXT 低粒子背景做参考）。D014 注册 ParetoGTICal。10-13 是结构性的三天：D015 把 xsherpa 与 esas_data_reduction 从只读兄弟变成被本项目迭代的工具（发现→issue→版本化发布→DXC 固定版本重测）；D016 固定 xsherpa dev3、pn QPB 标度只许 test-only；D017 给每个子项目独立的认知 OS，伞形问题定为"处理后的光子信号对年代、离轴角、姿态角、SP/QPB 状态不变"。ObsConsistency 就位（NGC 3221/5775 逐观测分解）。

D011 registered the XMM–EP-FXT cross-calibration sub-project (FXT's low particle background as reference). D014 registered ParetoGTICal. The 13th was structural: D015 turned xsherpa and esas_data_reduction from read-only siblings into tools this project iterates (finding → issue → versioned release → DXC re-test pinned); D016 pinned xsherpa dev3 and made the pn QPB scale test-only; D017 gave every sub-project its own cognition OS, with the umbrella question set to "the processed photon signal is invariant to epoch, off-axis angle, roll angle, SP/QPB state". ObsConsistency was in place (per-observation decomposition of NGC 3221/5775).

来源：Source: cognition-os/04_DECISION_LOG.md D011/D014/D015/D016/D017, cognition-os/02_CURRENT_QUESTION.md

第四阶段：考试集与分数（2026-10-13 至 16）

Stage 4: the exam set and the scores (2026-10-13 to 16)

考试集 v4 r240 冻结（D019/D020）：7 problem、18 ObsID、45 case。逐观测拟合（R3）给出第一批结构化结论：WCDFS/M31 的天空在 A 变体下一致且平；LMC 不一致——后来发现是 0690744801 的孔径选择（RT-M15）。红队密集期：RT-M01 发现仪器线（Al-K/Si-K）主导暗场计数级面亮度；RT-M08 发现离轴角用错了参考点（档案目标 vs 指向，16/40 换 bin）——\(\theta\) 体系全部重算；RT-M09 证明了 ARF-BACKSCAL 组合是对的，防住一个"修复"倒退。D022/D023/D024/D025 密集决策：Virgo 拟合环境、ciao-4.18、xsherpa dev6（G2 分组默认）、dev7（cstat 模式）。

The exam set v4 r240 was frozen (D019/D020): 7 problems, 18 ObsIDs, 45 cases. The per-observation fits (R3) gave the first structured verdicts: WCDFS/M31 skies consistent and flat under variant A; LMC inconsistent — later traced to 0690744801's aperture selection (RT-M15). The dense red-team period: RT-M01 found the instrumental lines (Al-K/Si-K) dominate faint-field count-level SB; RT-M08 found the wrong off-axis reference (archive target vs pointing, 16/40 cases change bin) — the whole \(\theta\) system was recomputed; RT-M09 proved the ARF-BACKSCAL combination right, blocking a "fix" that would have double-counted. Dense decisions D022/D023/D024/D025: the Virgo fit environment, ciao-4.18, xsherpa dev6 (G2 grouping default), dev7 (cstat mode).

来源：Source: MosaicSelfCal/RESEARCH_PLAN.md R3 rows, cognition-os/REDTEAM_QUEUE.md RT-M01/M08/M09/M15, cognition-os/04_DECISION_LOG.md D019–D025

第五阶段：SP 主导（2026-10-17 起）

Stage 5: SP takes over (from 2026-10-17)

PI 判断边缘变亮以 SP 为主（"SP dominate the edge brightening"），给出七步路线图（见 MosaicSelfCal/RESEARCH_PLAN.md）。R18a/b/c 测出 SP 自己的空间分布（硬带接近 ESAS det 图，光子 vignetting 被排除）；R18c 找到 proton 真正用的图（flare 图 = CCF XMM_FLARE_0001）与正确的每相机旋转（MOS2 是 PA+270，不是 PA+180——R18a 的 +12.9 % 因此撤回）。R18e/f 把相机一致的边缘步长做成 edgestep 包（v1/v2，RT-M52/M55 把校准问题摊开：\(\sigma\) 未校准，s 不能当天线读数）。R16/R16b/R16c 排除了 common-CCD 与 CCD4 重纳。R17 系列定下叠加规则。R19 系列把"SP 检测下限"倒过来：M31 的 SP 在下限之下，"探测不到"不等于"无害"（RT-M40）。

The PI judged the edge brightening SP-dominated ("SP dominate the edge brightening") and gave the seven-step roadmap (see MosaicSelfCal/RESEARCH_PLAN.md). R18a/b/c measured SP's own spatial distribution (hard band close to the ESAS det map, photon vignetting excluded); R18c found the maps proton really uses (flare maps = CCF XMM_FLARE_0001) and the right per-camera rotation (MOS2 is PA+270, not PA+180 — R18a's +12.9 % was withdrawn for this). R18e/f built the camera-consistent edge step into the edgestep package (v1/v2; RT-M52/M55 exposed the calibration problems: \(\sigma\) not calibrated, s not readable as an antenna). R16/R16b/R16c excluded common-CCD and CCD4 re-admission. The R17 series set the stacking rules. The R19 series inverted the "SP detection floor": M31's SP lies below its floor, so "undetectable" is not "harmless" (RT-M40).

来源：Source: MosaicSelfCal/RESEARCH_PLAN.md （SP 路线图与 R16–R19 行）(the SP roadmap and the R16-R19 rows), cognition-os/PITFALLS.md （RT-M33/M36/M40/M41/M45/M52/M55 条目）(the RT-M33/M36/M40/M41/M45/M52/M55 entries)

第六阶段：年代与协方差（2026-10-18 至 19）

Stage 6: epochs and covariances (2026-10-18 to 19)

R20 家族把 SP 水平问题拆成十几个可检验的小问题：重复指向（R20b/c）只是上限，受 QPB 协方差限制（R20d/e/f/g 把这个协方差本身量到构造层面）；R20h/i 发现 MOS1/MOS2 SP 率比在 2012 年台阶（0.68→1.47），与 CCD3 丢失完全退化；R20k 排除了掩膜足迹与角区模板陈旧；R20l 量了硬带 QPB 水平（最多解释 21\(\pm\)10 %）；R20m（QPB 库年代）\(\Delta R\le0.06\)，解释不了台阶；R20n 排除了逐成分全局相机标度；R20o（在跑）检验每观测公共因子。同一时间 SAS 22.1 后端过了代码审（v1.8.0-dev.5，91+6 测试），RT-M74 抓出 -back.pi STAT_ERR/TUNIT 缺陷。

The R20 family split the SP-level question into a dozen testable pieces: repeat pointings (R20b/c) give upper limits only, limited by the QPB covariance (R20d/e/f/g measured that covariance down to the construction); R20h/i found the MOS1/MOS2 SP-rate ratio steps in 2012 (0.68→1.47), perfectly degenerate with the CCD3 loss; R20k excluded the mask footprint and stale corner templates; R20l measured the hard-band QPB level (explains at most 21\(\pm\)10 %); R20m (QPB library epoch) \(\Delta R\le0.06\), not explaining the step; R20n excluded per-component global camera scales; R20o (running) tests the per-observation common factor. In parallel the SAS 22.1 backend passed code review (v1.8.0-dev.5, 91+6 tests), and RT-M74 caught the -back.pi STAT_ERR/TUNIT defects.

来源：Source: MosaicSelfCal/RESEARCH_PLAN.md （R20b–R20o 行）(the R20b-R20o rows), cognition-os/08_RUNNING.md §3/§7, cognition-os/tool_iteration.md 45

死胡同与教训

Dead ends and their lessons

项目记录死胡同与记录进展同样认真；每条都变成了规则（PITFALLS）。

The project records dead ends as carefully as progress; each became a rule (PITFALLS).

1. "SP 计数看起来像光子一样 vignetting"（R6 早期）：flat SP 扣除让 WCDFS 斜率变负、LMC 不动——没有单一模板同时满足两个场。教训：SP 需要自己的空间模板（PI 2026-10-17），后来 R18 系列测出它。"SP counts look vignetted like photons" (early R6): flat-SP subtraction sent WCDFS negative and left LMC untouched — no single template fits both fields. Lesson: SP needs its own spatial template (PI 2026-10-17), which the R18 series then measured.
2. R18a 的 +12.9 % 撤回：MOS2 det→sky 旋转用错（PA+180 而不是 PA+270），自检通过但仍是错的。教训：每个投影图都要对着 SAS 的 det/sky 图像对逐相机验证。R18a's +12.9 % withdrawn: the MOS2 det→sky rotation was wrong (PA+180 instead of PA+270) and passed its own reproduction checks anyway. Lesson: validate every projected map against the SAS det/sky image pair, per camera.
3. "R 的年代台阶"：RT-M75（1.25\(\sigma\)）、R20m（\(\Delta R\le0.05\)）、R20n（0.918/1.599）逐步把一个诱人的读法拆掉。教训：先解开天空再读年代；一个 7\(\sigma\) 在含场偏移的误差模型下可以是 3.1\(\sigma\)。"The epoch step of R": RT-M75 (1.25\(\sigma\)), R20m (\(\Delta R\le0.05\)), R20n (0.918/1.599) dismantled a tempting reading step by step. Lesson: unlink the sky before reading epochs; a 7\(\sigma\) can be 3.1\(\sigma\) under a field-offset error model.
4. cheese 掩膜读反（RT-M36）：把 0 当可用，报成"移除 48–92 %"，实际 8–52 %。教训：用之前对着 mask-im 检查约定。The cheese mask read backwards (RT-M36): treating 0 as usable reported "48–92 % removed" where the truth is 8–52 %. Lesson: check the convention against mask-im before use.
5. R17 注入检验的假偏差：inject_test.py 把所有 case 折过 cases[0] 的 ARF，制造了 −29…+32 % 的"叠加偏差"。教训：真实与注入必须用同一个 builder。R17's fake stacking bias: inject_test.py folded every case through cases[0]'s ARF, manufacturing a −29…+32 % "stacking bias". Lesson: real and injected data must go through the same builder.
6. 模拟器把抽样写回模板（RT-M42）：QPB 抽样链每代叠加，方差 1.03→5.5。教训：每代从冻结的模板抽样，检验 var/mean\(\approx1\)。A simulator wrote its draw back into the template (RT-M42): the QPB draw chain compounded, variance 1.03→5.5. Lesson: draw every replicate from a frozen template and test var/mean\(\approx1\).
7. common-CCD 不是解药（R16）：相交 CCD 集把全场缩到最残缺观测的足迹，反常观测原封不动（+21.8→+22.2 %），一致观测反而丢面积。教训：保留各观测自己的足迹，公共足迹只做诊断。Common-CCD is not a cure (R16): intersecting CCD sets shrinks the field to the most-reduced observation's footprint, leaving the anomaly untouched (+21.8→+22.2 %) while the consistent observations lose area. Lesson: keep each observation's own footprint; report the common footprint only as a diagnostic.
8. R20i 的 QPB 率 \(\times\)5（R20k 发现）：把计数 \(\times\) eV 重叠而不除以通道宽。教训：手工波段积分要对照拟合自己的波段行率核对。R20i's QPB rates \(\times\)5 (found by R20k): multiplying counts by the eV overlap without dividing by the channel width. Lesson: check any hand-made band rate against the fit's own band-row rates.

来源：Source: cognition-os/PITFALLS.md, cognition-os/REDTEAM_QUEUE.md （RT-M36/M42 行）(the RT-M36/M42 rows), MosaicSelfCal/RESEARCH_PLAN.md （R16/R17/R18a 行）(the R16/R17/R18a rows)