前世今生：项目史The story so far: a project history

这条时间线按项目自己的记录（cognition-os 的 session、决策日志、实验与红队记录，以及 git log）写成，死胡同也讲。每一条都有来源；被推翻的结论标“已撤回”。This timeline follows the project's own records (the cognition-os sessions, the decision log, the experiment and red-team records, and git log), dead ends included. Every entry carries a source; overturned claims are marked "withdrawn".

1. 2026 年 7–9 月：老服务器时代1. 2026, months 7-9: the old-server era

• 2026-07-29（git）：项目在老服务器上成型——Capella 点源演示（v1 到 v4）、N132D \(3 \times 3\) 试点（v1 拒绝：背景矩形被截断；v2 拒绝：VAPEC 下界低于硬下限；v3 通过）、NEI 重跑（v4）。2026-07-29 (git): the project took shape on the old server: the Capella point-source demos (v1 to v4), the N132D \(3 \times 3\) pilot (v1 rejected, the background rectangle truncated; v2 rejected, a VAPEC bound below the hard floor; v3 passed), and the NEI rerun (v4).
• 2026-08-25/29：背景策略系统化——SAS 用户指南与标定备忘录 XMM-SOC-CAL-TN-0058 都点名 N132D：标准局部背景含源发射线，会弄错线通量。由此确立模板背景（rgsbkgmodel）路线与“TOTAL COUNTS、永不减背景”的计数契约。2026-08-25/29: the background strategy was systematised. Both the SAS User Guide and the calibration note XMM-SOC-CAL-TN-0058 name N132D: the standard local background contains source emission lines and would corrupt line fluxes. This established the template-background (rgsbkgmodel) route and the "TOTAL COUNTS, never subtract" count contract.
• 2026-08-30（git 920e741、29fe88c）：仓库以私有 GitHub 发布；v0.3.0 公共 I/O 契约（schema v6）定稿。2026-08-30 (git 920e741, 29fe88c): the repository was published privately on GitHub; the v0.3.0 public I/O contract (schema v6) was finalised.
• 2026-09-29（git 7235a8d）：多 roll 进展归档，下一阶段计划（docs/n132d-next-stage-plan-2026-09-29.md）与 v17 基线移交。老服务器随后不可访问（SSH 超时）。2026-09-29 (git 7235a8d): multi-roll progress archived, with the next-stage plan (docs/n132d-next-stage-plan-2026-09-29.md) and the v17 baseline handed over. The old server then became inaccessible (SSH timeout).

来源：git log（2026-07-29 至 2026-09-29 的提交）；agent-memory/environment.md（Remote Defaults、Accepted Demo、N132D Pilot 各节）；docs/n132d-next-stage-plan-2026-09-29.md。Source: git log (the commits from 2026-07-29 to 2026-09-29); agent-memory/environment.md (the Remote Defaults, Accepted Demo and N132D Pilot sections); docs/n132d-next-stage-plan-2026-09-29.md.

2. 2026-10-01/02：搬到 lang，重建输入2. 2026-10-01/02: moving to lang, rebuilding the inputs

• 用户决定：不连接老服务器，缺的观测从 XSA 官方下载（D01）。缺的 0125100201 下载、校验、入档；三个观测的 prepare/extract 完成（E001，INPUTS_COMPLETE，39 个注册产品）。The user decided: do not connect to the old server; download the missing observation from the official XSA (D01). The missing 0125100201 was downloaded, verified and ingested; the three observations' prepare/extract completed (E001, INPUTS_COMPLETE, 39 registered products).
• 参考探针试点（E002）：0210681301 的中心、边界、中央条带外三个 XDSP 位置，21 个响应，预声明 0.001 阈值全过（PASS_REFERENCE_PROBES，2026-10-02T12:12:11Z）。两次失败保留为证据：一次 SAS 参数拼写，一次 FITS 字符串填充差异（B001）。The reference probe pilot (E002): three XDSP positions of 0210681301 (centre, edge, outside the central strip), 21 responses, all under the pre-declared 0.001 threshold (PASS_REFERENCE_PROBES, 2026-10-02T12:12:11Z). Two failures were kept as evidence: a SAS parameter typo and a FITS string-padding difference (B001).
• 项目认知 OS（cognition-os）从 Hermes research-cognition-os 框架初始化（D03）。The project cognition OS was initialised from the Hermes research-cognition-os framework (D03).

来源：cognition-os/sessions/2026-10-02.md；04_DECISION_LOG.md D01、D03；03_TRUST_TABLE.md E001、E002、B001 行。Source: cognition-os/sessions/2026-10-02.md; 04_DECISION_LOG.md D01 and D03; 03_TRUST_TABLE.md rows E001, E002 and B001.

3. 2026-10-03 到 10：27 路诊断与 PSF 对照3. 2026-10-03 to 10: the 27-path diagnostic and the PSF comparisons

• 10-03：27 路（9 源格到 3 条带）固定谱形诊断完成（E004）：CStat 301.110249 对同条带 9 路的 299.953830——完整响应没有改善统计描述（I006）。教训：响应数量不是科学标准，可恢复性才是。首轮优化器与注入门失败保留（B003）。10-03: the 27-path (9 source cells to 3 strips) fixed-shape diagnostic completed (E004): CStat 301.110249 against 299.953830 for the same-strip 9 paths; the full response did not improve the statistical description (I006). Lesson: response count is not a scientific criterion; recoverability is. First-round optimizer and injection-gate failures were kept (B003).
• 10-05：九格 r 线振幅之和差 8.50 %，但 CStat 几乎不动——模型敏感性，不是偏差结论（D06）。10-05: the nine-cell r-line amplitude sums differ by 8.50 % while CStat barely moves; model sensitivity, not a bias conclusion (D06).
• 10-06 到 10：四轮配对 Poisson mock（E005–E009）把“完整 PSF 是否更好”这个问题拆开回答：绝对通量上，遗漏跨条带路径高估 +8 到 +14 %（E008）；有限计数的线比 RMSE 上，完整 PSF 不保证更好（E009）；自由共同线宽不吸收遗漏路径（E011）。教训：收益取决于目标量，没有免费的午餐。10-06 to 10: four rounds of paired Poisson mocks (E005-E009) took apart "is the full PSF better": for absolute flux, omitting cross-strip paths overestimates by +8 to +14 % (E008); for finite-count line-ratio RMSE, the full PSF is not guaranteed better (E009); a free common width does not absorb the omitted paths (E011). Lesson: the gain depends on the target quantity; there is no free lunch.

来源：cognition-os/sessions/2026-10-05.md、2026-10-06.md；03_TRUST_TABLE.md E004、E008、E009、E011、B003 行；04_DECISION_LOG.md D06。Source: cognition-os/sessions/2026-10-05.md and 2026-10-06.md; 03_TRUST_TABLE.md rows E004, E008, E009, E011 and B003; 04_DECISION_LOG.md D06.

4. 2026-10-13/14：定位澄清与白盒仿真器4. 2026-10-13/14: the framing clarification and the white-box emulator

• 10-13：用户澄清项目定位（D17）：这是数据融合——用 RGS 不同 roll 的互补色散投影，重建高能量分辨率的谱；目标精度 XRISM 级。单 roll 内空间与波长简并，换 roll 只改空间投影，谱不变，可借此分离。10-13: the user clarified the framing (D17): this is data fusion, using the RGS's complementary dispersion projections at different rolls to rebuild a high-energy-resolution spectrum; the target precision is XRISM-grade. Within one roll, space and wavelength are degenerate; changing the roll changes only the spatial projection, not the spectrum, which separates them.
• 10-14：白盒仿真器系列（E010–E013）：色散偏移就是平移（0.1233 Å/′，几何预言的 0.995）；官方 angdist 展源响应与点源叠加差 1.3–2.4 %，平移仿真器 0.14–0.54 %；玩具速度梯度 mock 里 0°+180° 最利于 \(g_x\) 与 \(v_0\)、0°+90° 利于 \(g_y\)、三 roll 最好（E013）。教训：先造一个你能完全控制的仪器模型，再去碰真实数据。10-14: the white-box emulator series (E010-E013): a dispersion offset acts as a shift (0.1233 Angstrom per arcmin, 0.995 of the geometric prediction); the official angdist extended-source response deviates 1.3-2.4 % from a point-source superposition, the shift emulator 0.14-0.54 %; in a toy velocity-gradient mock, the 0+180 degree pair is best for \(g_x\) and \(v_0\), 0+90 for \(g_y\), and three rolls are best overall (E013). Lesson: build an instrument model you fully control before touching real data.

来源：00_AI_BRIEFING.md（定位澄清 2026-10-13 与 E010–E013 各节）；04_DECISION_LOG.md D17–D20；03_TRUST_TABLE.md E010、E012、E013、I015 行。Source: 00_AI_BRIEFING.md (the framing clarification of 2026-10-13 and the E010-E013 sections); 04_DECISION_LOG.md D17-D20; 03_TRUST_TABLE.md rows E010, E012, E013 and I015.

5. 2026-10-15 到 17：交叉色散闭合与红队5. 2026-10-15 to 17: cross-dispersion closure and the red teams

• 10-15：真实数据拟合前的闭合检验（E018a–g）发现一串仪器事实：前向模型把 N132D 放偏 \(-6.25\)″（I021）；LSF 宽度差 2 % 使 \(\sigma_v\) 偏 24 km/s（I020）；Chandra 三个波段的形态不同（I023）。用户同时启动常驻红队（D26）：RT01–RT23 至今，红队只提交发现，主 agent 复跑后才改信任。10-15: the closure checks before the real fit (E018a-g) turned up a chain of instrument facts: the forward model misplaces N132D by \(-6.25\) arcsec (I021); a 2 % LSF width error biases \(\sigma_v\) by 24 km/s (I020); the three Chandra bands differ in morphology (I023). The user also started the standing red teams (D26): RT01-RT23 so far; red teams submit findings only, and the main agent changes trust after replaying them.
• 10-16/17：标定源审计（E019a–e，20 个观测）：Capella 与 PKS 2155-304 都显示交叉色散平移随波长变化（\(-0.21\)″/Å，I026）；点源在条带内预测准到 0–5 %（I028）；但 \(w \approx -3\,\%\) 依然无解（RT14、RT19 都没能解释）。红队自己的错误也被记录：RT03 两次注入失败后由主 agent 精确重做；RT17 线性偏差符号反、放大 2 倍，主 agent 改正重算。10-16/17: the calibrator audit (E019a-e, 20 observations): both Capella and PKS 2155-304 show the cross-dispersion shift varying with wavelength (\(-0.21\) arcsec per Angstrom, I026); a point source inside a strip is predicted to 0-5 % (I028); yet \(w \approx -3\,\%\) remains unexplained (neither RT14 nor RT19 could explain it). The red teams' own errors were recorded too: RT03's injection failed twice and the main agent redid it exactly; RT17's linear bias had the sign reversed and doubled, and the main agent recomputed it.

来源：00_AI_BRIEFING.md（10-15 至 10-17 的各节）；REDTEAM_QUEUE.md（RT01–RT23 各行，含撤回记录）；03_TRUST_TABLE.md I020–I028 行。Source: 00_AI_BRIEFING.md (the sections from 10-15 to 10-17); REDTEAM_QUEUE.md (the RT01-RT23 rows, with the withdrawal records); 03_TRUST_TABLE.md rows I020-I028.

6. 2026-10-19：D29 与第一次真实速度拟合6. 2026-10-19: D29 and the first real velocity fit

• E018b 第一阶段（10-17 完成，10-19 双臂核对）：第一次用真实 RGS 数据拟合 N132D 的 O VII 速度场——气体相对 LMC 静止（\(v_0 = 12 \pm 21\) km/s），速度弥散 \(\sigma_v = 500 \pm 27\) km/s（远超 85 km/s 的 LSF 上限），向西梯度只有 \(1.9\sigma\)、方向不稳健（I029，Provisional）。E018b, the first stage (finished 10-17, both arms checked 10-19): the first real-RGS-data fit of N132D's O VII velocity field. The gas is at rest in the LMC frame (\(v_0 = 12 \pm 21\) km/s) with a dispersion \(\sigma_v = 500 \pm 27\) km/s (far above the 85 km/s LSF limit); the westward gradient is only 1.9 sigma and its direction is not robust (I029, Provisional).
• 红队 RT21 的双成分检验：固定分界把 \(\Delta C\) 改善 \(-101\)，但每半仍需 \(\sigma_c = 438\) km/s——宽度主要是内部弥散。梯度随谱线形状模型变化，尚不稳健。Red team RT21's two-component test: a fixed boundary improves \(\Delta C\) by \(-101\), but each half still needs \(\sigma_c = 438\) km/s; the width is mostly internal dispersion. The gradient moves with the line-shape model and is not yet robust.
• PI 设定 D29：O VII 与 O VIII 窄带的完整谱线建模，空间与速度分布；handbook 每 8 小时更新。E018c（O VII 第二阶段）、E020（O VIII 输入包）、E021（档案普查）、RT22、RT23 派出，全部在跑（待定）。The PI set D29: the complete line modelling of the O VII and O VIII narrow bands, with the spatial and velocity distributions; the handbook refreshes every 8 hours. E018c (O VII stage 2), E020 (the O VIII pack), E021 (the archive census), RT22 and RT23 were dispatched and are all running (Pending).

来源：04_DECISION_LOG.md D29；insights/I029_n132d-ovii-velocity-stage1.md（含 2026-10-19 修订）；sessions/2026-10-19.md；08_RUNNING.md。Source: 04_DECISION_LOG.md D29; insights/I029_n132d-ovii-velocity-stage1.md (with the 2026-10-19 amendment); sessions/2026-10-19.md; 08_RUNNING.md.