前世今生History

阶段 1:CDFS 起点（2026-08-30 – 09-02）Phase 1: CDFS beginnings (2026-08-30 - 09-02)

老项目（名字叫 CDFS,但 GTI 基准数据其实全是 M31 MOS2,ObsID 0800730101–0800733101,30 个指向的马赛克）从"配对实验"开始：不做没配对的 2.1/4 对比，先造配对产品。第一批结论很快否定了几个直觉：响应不随 GTI 改变（被驳倒：ARF/RMF 因掩膜和选择而异）；2.1.filter 用 FOV/corner 比选 GTI（被驳倒：它只看全探测器率，corner 曲线是事后诊断）；删掉更多时间就删掉了更多软质子（被驳倒：0800730101 MOS1 的历史 2.1 GTI 甚至不是按它自己标注的中心选的）。

The old project (named CDFS, though its GTI benchmark data are actually all M31 MOS2, ObsIDs 0800730101-0800733101, a 30-pointing mosaic) began with a "paired experiment" decision: no unpaired 2.1/4 comparison; build paired products first. The first conclusions quickly overturned several intuitions: the response does not change with the GTI (refuted: ARF/RMF vary with mask and selection); 2.1.filter selects GTIs by a FOV/corner comparison (refuted: it only sees the whole-detector rate; the corner curves are post-hoc diagnostics); cutting more time removes more soft protons (refuted: the historical 2.1 GTI of 0800730101 MOS1 was not even selected at the centre its own plot labels).

阶段 2:2.1/2.1c 线锚定设计（2026-09-03 – 09-04）Phase 2: the 2.1/2.1c line-anchored designs (2026-09-03 - 09-04)

核心想法变成：仪器荧光线是背景的锚。软质子经望远镜聚焦，不带 Al/Si 荧光线；粒子打在相机结构上才激发荧光线。所以"高/低率两段光谱共享同样的线"是选时间时的约束。2.1c 设计把验证变成两个坐标：1 s 泊松平稳性检验 + f_T–R_C 帕累托平面。第一份 CDFS 配对筛选（0108060601 MOS1 的 A/B0）立刻给出一个有用的负结果：B0 少 26.2% 时间，连续谱却没降（R_C = −0.037 ± 0.047）—— "丢时间没换来干净"。

The core idea became: the instrumental fluorescence lines anchor the background. Soft protons are focused by the telescope and carry no Al/Si fluorescence; the lines are excited by particles striking the camera structure. So "the high/low rate spectra must share the same lines" is a constraint when selecting time. The 2.1c design turned validation into two coordinates: a 1 s Poisson stationarity test and the f_T - R_C Pareto plane. The first CDFS paired screen (0108060601 MOS1, A/B0) immediately gave a useful negative result: B0 keeps 26.2% less time with no lower continuum (R_C = −0.037 ± 0.047) — "exposure lost with nothing gained".

阶段 3:策略网格 r7–r23 与教训（2026-09-05 – 09-07）Phase 3: the strategy grids r7-r23 and their lessons (2026-09-05 - 09-07)

在 0800730101 与近稳态控制 0800730201 上铺开策略网格（r7 直接到 r23,几百个候选 GTI）：对称 σ 截断、贝叶斯块、CUSUM、corner 标度……这一阶段产出三个持久教训：

The strategy grids were laid out on 0800730101 with the near-steady control 0800730201 (r7 through r23, hundreds of candidate GTIs): symmetric σ cuts, Bayesian blocks, CUSUM, corner scaling... This phase produced three durable lessons:

1. 负控制是必须的:在 0800730201 上，旧的连续谱约减坐标能到 +10.42% 而观测根本没有耀斑 —— 纯选择伪影。Al 线比坐标最大只有 +1.07% (0.66σ),经受住了这个检验。Negative controls are mandatory: on 0800730201 the old continuum-reduction coordinate reached +10.42% although the observation has no flares — a pure selection artifact. The Al line-ratio coordinate peaked at +1.07% (0.66σ) and survived the test.
2. 误差棒曾经是错的:Gehrels 上限被当作方差传播，σ 被膨胀中位数 3.52 倍。2026-09-07 全部重算（r3 重打分），结论反转：r22 不是惰性的（9/1,Z = 3.66），r16 才是不迁移的那个（10/8,Z = 1.33）。The error bars were wrong: the Gehrels upper limit was propagated as a variance, inflating σ by a median factor 3.52. On 2026-09-07 everything was recomputed (the r3 rescore) and conclusions reversed: r22 is not inert (9/1, Z = 3.66); r16 is the one that does not transfer (10/8, Z = 1.33).
3. 单观测前沿不是证据:r16 是 0800730101 前沿的真实成员，迁移后却不显著。由此定下项目最重要的方法论规则：只有预注册规则在未参与选择的观测上的应用才算证据。A single-observation frontier is not evidence: r16 was a genuine member of the 0800730101 frontier yet failed to transfer. Hence the project's most important methodological rule: only a pre-registered rule applied unchanged to observations that took no part in its selection counts as evidence.

同一时期还回答了"删掉的到底是什么":A1 测量移除超额的光谱 —— 79% 无线（Al −5.96σ,Si 独立一致 −2.58σ），软；B1 测量其空间形状 —— 延展、非紧凑，更像软质子花样，但不能排除许多微弱可变源（M31 盘的 X 射线双星）。

The same period answered "what was actually removed": A1 measured the removed excess's spectrum — 79% line-free (Al −5.96σ, with Si independently agreeing at −2.58σ), and soft; B1 measured its spatial shape — extended, not compact, closer to the soft-proton pattern, though many faint variable sources (X-ray binaries in the M31 disk) are not excluded.

阶段 4:I004 定稿与 g1 实现（2026-09-14 – 09-16）Phase 4: I004 finalised and the g1 implementation (2026-09-14 - 09-16)

I004 把帕累托坐标定稿：纵轴改为 R = C_all/C_K（参考固定为该观测的全时段连续谱，而不是随 f_T 移动的弃掉时间），约束 L_all = L_K 在不相交的 K、B 上评估。同日实现（g1）并冻结 19/11 训练/保留划分 —— 当天就消费了 11 个保留 ObsID（所有率排名候选的保留集 R 已在收据里）。这个"保留集测试"是近平凡的：任何耀斑切割都给 ln R > 0（Z = 124 只说明"这些观测有耀斑"）。2026-09-16 生成教学收据（0800731101 的 33 个候选）。

I004 finalised the Pareto coordinates: the ordinate became R = C_all/C_K (the reference is the observation's own all-time continuum, not the rejected time that moves with f_T), with the constraint L_all = L_K evaluated on disjoint K and B. The g1 implementation ran the same day and the 19/11 train/holdout split was frozen — the 11 holdout ObsIDs were consumed that same day (every rate-rank candidate's holdout R is in the receipt). That holdout test is near-trivial: any flare cut gives ln R > 0 (Z = 124 only says "these observations have flares"). The tutorial receipt (33 candidates for 0800731101) followed on 2026-09-16.

阶段 5:迁入 DXC 与 P001(2026-10-13)Phase 5: the move into DXC and P001 (2026-10-13)

一天之内五步：D001 迁移（legacy 快照 + 三个活动脚本，tools/verify_migration.py 校验 PASS）;D002 事实纠错（训练/保留集已消费、两个度量族、数据是 M31 不是 CDFS、Al 亏损只在反事实修正后出现）;D003 预备 M31 MOS1 输入 + 派发 CDFS 还原 + 起草 P001;D004 PI 批准 P001 与修正案 A1–A3、代码冻结、回归门 PASS(26,111 个数字，最大相对差 0);D005 MOS1 盲测完成 —— E1 通过、E2 通过、E3 失败(13/25,Z = 0.02),角落太吵。

Five steps in one day: D001 the migration (a legacy snapshot + three active scripts, verifier PASS); D002 the fact correction (the train/holdout split already consumed, two metric families, the data are M31 not CDFS, the Al deficit appears only after the counterfactual correction); D003 staging the M31 MOS1 inputs + dispatching the CDFS reduction + drafting P001; D004 the PI approving P001 and amendments A1-A3, the code freeze, the regression gate PASS (26,111 numbers, max relative difference 0); D005 the MOS1 blind run — E1 PASS, E2 PASS, E3 FAIL (13/25, Z = 0.02), corner too noisy.

阶段 6:CDFS —— 两个代理的还原、一次纠错、盲测与关闭（2026-10-13，D006）Phase 6: CDFS — a two-agent reduction, one correction, the blind run and the close (2026-10-13, D006)

CDFS 的事件列表在 Virgo 上已不存在（被当作可再生而清理），所以 34 个 MosaicSelfCal ObsID 从头做 stage-1/2 还原（esas_data_reduction v1.7.7,MOS1+MOS2）。还原拆给两个代理：codexnim 18 个、Hermes 17 个，0604960101 由两家都做、按 A2 条款 11 取 codexnim 副本直接比对 —— 两次还原因 emchain 帧内随机化而不同（2.filter GTI 差 MOS1 −420 s、MOS2 −60 s），所以比较要按统计意义、不按 sha256。执行中出了两个教训，都进了 PITFALLS:U002 纠错 —— Hermes 先把 0555780601 跳过，标它"CalClosed";其实它只有 U002 MOS 曝光，而 U 开头不等于定标曝光（XSA is_scientific=true，116.6 ks 科学曝光），纠正后重做;路径核查 —— Virgo 上的 XMM_Data_nvme 名字带 nvme、实际在 HDD（/home/huangrui 是到 HDD 的符号链接），用 readlink -f 与 df 核实后才写进简报。Claude 独立验收了全部 34 个（sha256 对清单、PrimeFullWindow/Thin1、n > 0）。然后冻结代码 46d8df3 原样盲测两台相机（M1 47 min、M2 45 min，各 34/34），终点一次算出（E003），D006 记录并关闭 P001：主问题否定（四个主集的 E3 全失败），角落按数据集回答，还留下一个未预注册的探索性张力 —— 角落的粒子水平变化不带入门内 Al 线比。

CDFS event lists no longer existed on Virgo (cleaned as reproducible), so the 34 MosaicSelfCal ObsIDs were reduced from scratch through stage-1/2 (esas_data_reduction v1.7.7, MOS1+MOS2). The reduction was split across two agents — codexnim 18, Hermes 17 — with 0604960101 done by both and taken from codexnim per A2 item 11; the two reductions differ because emchain randomises within frames (2.filter GTI differs by MOS1 −420 s, MOS2 −60 s), so reductions are compared statistically, never by sha256. Two execution lessons entered PITFALLS: the U002 correction — Hermes first skipped 0555780601 as "CalClosed"; in fact its only MOS exposures are U002, and a U prefix does not mean calibration (XSA is_scientific=true, a 116.6 ks science exposure), so it was corrected and redone; and the path check — Virgo's XMM_Data_nvme has "nvme" in its name but lives on an HDD (/home/huangrui is a symlink to an HDD), verified with readlink -f and df before it entered a brief. Claude independently accepted all 34 (sha256 vs manifest, PrimeFullWindow/Thin1, n > 0). Then the frozen code 46d8df3 ran blind on both cameras (M1 47 min, M2 45 min, 34/34 each), the endpoints came out in one pass (E003), and D006 recorded them and closed P001: negative on the primary question (E3 fails on all four primary sets), the corner answered per set, and one exploratory (not pre-registered) tension left open — the corner's particle-level change does not carry into the as-gated Al-line ratio.

死胡同清单（每个都教会一件事）The dead ends (each taught something)