目录Contents
前世今生：时间线与教训Past and present: the timeline and its lessons
项目史不只是成功史——每个死胡同都有教训。日期与结论均来自 cognition-os 的决策日志（D 编号）、洞察（I 编号）与实验（E 编号）。
This history is not just a list of successes — every dead end has a lesson. Dates and conclusions come from the cognition-os decision log (D numbers), insights (I numbers) and experiments (E numbers).
时间线Timeline
| 时间When | 事件What happened | 教训Lesson |
|---|---|---|
| 2026-05-30 | S01 开始：101 个 ObsID 的「盘 0.7 keV 超出 vs 恒星质量」研究。D01：把 080073 晕场当「纯银河系天空背景」用。S01 begins: "disc 0.7 keV excess vs stellar mass" over 101 ObsIDs. D01: the 080073 halo fields are used as a "pure MW sky background". | 起点合理，但前景假设会在 S02 被推翻——永远先问「背景里还混着什么」。A reasonable start, but the foreground assumption will be overturned in S02 — always ask "what else is mixed into my background". |
| 2026-05-31 | D17–D25：单 ObsID 拟合 + 加权平均天空背景；发现 SP 在单场拟合里失控（disk SP 跑到 0.86），用 21 场平均解决。D20 冻结 MWhothalo=0 得盘 0.7 keV 上限。D22 用 delta_stat F 检验分类。D17–D25: per-ObsID fits plus a weighted-average sky background; SP runs away in single-field fits (disc SP to 0.86), solved by averaging 21 fields. D20 freezes MWhothalo=0 for a disc 0.7 keV upper limit. D22 classifies detections by a delta_stat F-test. | D22 的 F 检验后来被判 Questioned：非负边界上 F 检验无效（见 8 节）。方法要匹配参数空间。D22's F-test is later Questioned: invalid at the non-negative boundary (see section 8). Match the method to the parameter space. |
| 2026-07-10 | 转折点 S02（D26）：080073 的 22 个指向重新定性——它们是 M31 内晕科学视线（中心半径 10.5–30.8 kpc），不是纯银河系背景。D01 作废；「MWhalo」分量实为 MW+M31 之和（I006）。D30/D32：论文改为「先测总和 CGMsum，条件场景再分解」。The S02 turn (D26): the 22 080073 pointings are reclassified — they are M31 inner-halo science sightlines (centre radii 10.5–30.8 kpc), not pure MW background. D01 is retired; the "MWhalo" component is really the MW+M31 sum (I006). D30/D32: the paper becomes "measure the sum CGMsum first; decompose only in conditional scenarios". | 项目最大的一次转向：一个「背景」定性错误会把整篇论文的科学问题改写。改对了才有后面的测量。The project's biggest turn: one mis-classification of a "background" rewrites the paper's whole scientific question. Getting it right made the measurement possible. |
| 2026-07-10 | E003：v18 → v19 生产基线。B003 揭示 v18 的有效滤波把已剔除通道带回（44/44 会话含禁带），v19 修正：notice 先行、强分组边界、Al/Si 线冻结为 0、逐指向 HI4PI N_H。22 场全部跑完 → 18 技术通过 → 14 主样本。共温 kT=0.17531±0.00436 keV；南北差 0.212±0.210（1.01σ，一致性）。E003: the v18 → v19 production baseline. B003 reveals that v18's effective filter reintroduced excluded channels (44/44 sessions included the forbidden band); v19 fixes it: notice first, forced group boundaries, Al/Si lines frozen at 0, per-pointing HI4PI N_H. All 22 fields run → 18 technical passes → 14 primary. Common kT = 0.17531±0.00436 keV; North-minus-South 0.212±0.210 (1.01σ, consistency). | 「源码里写了什么」不等于「会话里保存了什么」——B003 靠恢复会话审计才抓到。审计实际保存的产品，别信代码注释。"What the source says" is not "what the session saved" — B003 was caught only by restoring sessions. Audit the saved products, not the code comments. |
| 2026-07-10–15 | D34–D47：16 页 ApJ 内部稿（AASTeX）+ 公开镜像 + MW/M31 剖面探索器。Figure 3 文献一致性检验建立：Locatelli 1.290 > 测量 0.965（紧张）；Ueda 0.312、HaloSat 0.565（域外外推）、eFEDS 0.291/0.407、Henley 中位 0.385（非本地）。E008 图谱、E007 外部比较器审计相继完成。D34–D47: the 16-page ApJ internal draft (AASTeX) plus the public mirror and the MW/M31 profile explorer. The Figure-3 literature consistency test is built: Locatelli 1.290 > measured 0.965 (a tension); Ueda 0.312, HaloSat 0.565 (outside-domain extrapolation), eFEDS 0.291/0.407, Henley median 0.385 (non-local). The E008 atlas and E007 external-comparator audit follow. | 每个文献值都要带「域」标签：本地方向模型、域外外推、非本地人群——不可比的东西不放在同一根轴上假装可比。Every literature value carries a "domain" label: local-directional model, outside-domain extrapolation, non-local population — do not put incomparable things on one axis and pretend they compare. |
| 2026-07-16 – 10-13 | 死胡同 E009：M31 中心 R8-12/R10-14 环的 3B vs 4 分支温度比较。v1–v6 合同/审计迭代（0C/7I/3M → REQUEST_CHANGES → APPROVED 0C/0I/0M），但从未执行：历史回放正控制被 REJECTED（0.5–2 keV 源计数相对差 0.2614，阈值 0.01），冻结阈值没有放宽。Dead end E009: the M31 central R8-12/R10-14 annulus 3B-vs-4 branch temperature comparison. v1–v6 contract/audit iterations (0C/7I/3M → REQUEST_CHANGES → APPROVED 0C/0I/0M), but never executed: the historical replay positive control was REJECTED (0.5–2 keV source counts relative difference 0.2614 against a 0.01 threshold), and the frozen threshold was not relaxed. | 教训：正控制失败时不要放松阈值硬跑；合同再完美，也救不了无法回放的数据。D48 把 E009 整体冻结，退出论文路径。Lesson: when the positive control fails, do not relax the threshold to force a run; a perfect contract cannot rescue data that cannot be replayed. D48 freezes E009 entirely, off the paper path. |
| 2026-10-13 | E012：Zhang/He & Li 叠加剖面与 14 场 CGMsum 的「垂直比较」页面（探索器 stack-comparison 路由）。确认 Zhang v3 MW/M31 模板与测量的条件比较（非排除）。E012: the Zhang/He & Li stack-profile "vertical comparison" page against the 14-field CGMsum (the explorer stack-comparison route). Confirms the Zhang v3 MW/M31 templates as a conditional comparison (not an exclusion). | 比较页先于似然页：先「量级对不对」，再「统计上排不排除」。Comparison pages precede likelihood pages: first "is the order of magnitude right", then "does statistics exclude it". |
| 2026-10-14 | D49（今天）：E014 交叉检验完成（luna 主结果 + Hermes 校正轮）。外 NW 四场偏亮不是大尺度天空结构（RASS 比值 0.91 ± 0.12 vs XMM 1.28，无纬度梯度），不是 OMNI 追踪的太阳风电荷交换（无窗口超 4e8 cm⁻² s⁻¹）；软质子水平偏高（比值 3.1）但与亮度相关性弱（Spearman +0.35，p=0.22），未确立。D49 决定：在专门检验之前，外 NW 四场不解释为 MW 或 M31 结构，系统误差预算记为未解释场级项 ~0.25。排队检验：(a) lang 现有谱、软质子谱形自由重拟合（含同期对照 0701/0801/1501/1601）；(b) Virgo ODF 级更严格耀斑过滤后重处理。D49 (today): the E014 cross-check is complete (luna main result plus the Hermes correction round). The bright outer-NW four are not large-scale sky structure (RASS ratio 0.91 ± 0.12 vs XMM 1.28, no latitude gradient) and not OMNI-traced solar-wind charge exchange (no window above 4e8 cm⁻² s⁻¹); the soft-proton level is high (ratio 3.1) but correlates weakly with brightness (Spearman +0.35, p=0.22), not established. D49: until a dedicated test exists, the four fields are not interpreted as MW or M31 structure, and the systematic budget carries an unexplained field-level term of ~0.25. Queued tests: (a) a refit from the existing spectra on lang with the soft-proton spectral shape freed (plus same-epoch controls 0701/0801/1501/1601); (b) reprocessing on Virgo with a stricter ODF-level flare filter. | 教训：一个 2.5σ 的「纬度梯度」在独立历元（RASS）与独立物理通道（OMNI）下都不复现——先交叉验证，再写进论文解释。Lesson: a 2.5σ "latitude gradient" did not reproduce under an independent epoch (RASS) or an independent physical channel (OMNI) — cross-check first, then write it into the paper's interpretation. |
| 2026-10-14 | D48（同日）：v19 测量冻结；论文定名 New-ANGELS III；Kavanagh 2020 更正（其区域是盘旁背景，非晕视线）；I033 快速形状检验（轮廓平、M31 β 上限 0.05–0.19、NW 纬度梯度 Δχ²=6.2）；E013 谱联合似然与 E014 RASS/OMNI 交叉检验派发；进入科学讨论阶段。D48 (same day): the v19 measurement is frozen; the paper is named New-ANGELS III; the Kavanagh 2020 correction (their regions are disc-adjacent backgrounds, not halo sightlines); the I033 quick shape check (flat profile, M31 beta limit 0.05–0.19, NW latitude gradient Δχ²=6.2); E013 (spectral joint likelihood) and E014 (RASS/OMNI cross-check) dispatched; the project enters the science-discussion phase. | 「测量冻结 + 解释继续」让论文与实验解耦：E013/E014 不会推翻 0.965±0.044，只会改变解释。"Freeze the measurement, keep interpreting" decouples the paper from the experiments: E013/E014 cannot overturn 0.965±0.044, only its interpretation. |
来源：cognition-os/04_DECISION_LOG.md（D01, D17–D25, D26, D30–D47, D48, D49）；cognition-os/experiments/E003、E009、E012、E014；cognition-os/insights/I006、I033；cognition-os/bug_impacts/B003。
Source: cognition-os/04_DECISION_LOG.md (D01, D17–D25, D26, D30–D47, D48, D49); cognition-os/experiments/E003, E009, E012, E014; cognition-os/insights/I006, I033; cognition-os/bug_impacts/B003.
死胡同清单（不止 E009）The dead-end list (not just E009)
- S01 的 0.7 keV 盘分支：相对结论（盘减晕）可能活下来，但物理归属依赖「晕=纯背景」——D26 推翻后整支变 Provisional，需在前景重解释下重跑（03_TRUST_TABLE）。S01's 0.7 keV disc branch: the relative (disc minus halo) result may survive, but its physical attribution depended on "halo = pure background" — after D26 the whole branch is Provisional and needs a re-run under the revised foreground (03_TRUST_TABLE).
- RASS v4/rates 分支：rates-PI 拟合三次滤带失败、收敛到差 162.27 的局部极小，还覆盖了保存的 .fitting 会话（B001）。教训：失败分支会破坏好产品——会话要只读，或先备份。The RASS v4/rates branch: the rates-PI fit failed three filter checks, converged to a minimum worse by 162.27, and overwrote the saved .fitting session (B001). Lesson: a failing branch destroys good products — sessions must be read-only, or backed up first.
- 旧宽场 0.5–0.7 keV「剖面」：130 ObsID 的计数空间表混了 3B/4 与不同滤片，48 场里只有 27 个过 20% 一致性门，且无径向趋势（ρ=+0.009）——它不是 M31 剖面，只是再处理的候选控制场（I026/I028）。The old wide-field 0.5–0.7 keV "profile": the 130-ObsID count-space table mixes 3B/4 and different filters; only 27 of 48 fields pass a 20% consistency gate and show no radial trend (ρ=+0.009) — it is not an M31 profile, only a candidate control sample for reprocessing (I026/I028).
- X-LEAP 作为第六个宽带参考：M31 4.4° 内无干净行，源的公式与图自相矛盾——只能作验证，不作先验（I017/I023）。X-LEAP as a sixth broadband reference: no clean row within 4.4° of M31 and the source's own equation conflicts with its figure — validation only, never a prior (I017/I023).
来源：03_TRUST_TABLE.md 对应行（Disk 0.7 keV、RASS v4、Legacy wide-field、X-LEAP）；cognition-os/insights/I026、I028、I017；cognition-os/bug_impacts/B001。
Source: the matching rows of 03_TRUST_TABLE.md (Disk 0.7 keV, RASS v4, Legacy wide-field, X-LEAP); cognition-os/insights/I026, I028, I017; cognition-os/bug_impacts/B001.