评审更正（2026-10-19T16:35Z）：本版数据截至 D079。之后：D080 主列表改为 XJ.MS、JFR_z3、GATE0_f0.15、KFP_z3_a1、LG.MS；D082 功效区间：5 个中预期 Confirmed 1.8–2.5（开发替身），4.3–4.6（P006，上限）；D083（待 PI 批准）：增益部分来自丢弃整条观测，Confirmed 另需 p16(D̄_both) ≤ 0。Review correction (2026-10-19T16:35Z): this build reflects data up to D079. Since then: D080 primary list is XJ.MS, JFR_z3, GATE0_f0.15, KFP_z3_a1, LG.MS; D082 power bracket: expected Confirmed of 5 is 1.8–2.5 (development stand-ins), 4.3–4.6 (P006, upper bound); D083 (pending PI): part of the gain comes from discarding whole observations, so Confirmed also needs p16(D̄_both) ≤ 0.

光变曲线能替 GTI 吗？—— 一个冻结规则的保留集判决Can light curves replace GTIs? — a frozen-rule held-out verdict

报告人占位符 · 单位占位符 · 2026 年 · 45 分钟 + 10–15 分钟提问Speaker placeholder · Institution placeholder · 2026 · 45 min + 10-15 min questions

ParetoGTICal：XMM-Newton 弥散 X 射线观测的 GTI 筛选政策校准ParetoGTICal: calibrating GTI screening policies for XMM-Newton diffuse X-ray observations

11

同一天空的两次观测，背景水平可以差 20% —— 而你只有 5% 的科学预算Two observations of the same sky can differ by 20% in background level — and your science budget is 5%

0.5–2 keV 拼接光变曲线的观测间散布 \(\sigma_R\)，2.filter（归档标准）下 17–25%。The between-observation scatter \(\sigma_R\) of 0.5-2 keV concatenated light curves is 17-25% under 2.filter (the archived standard).

软质子随时涨落，GTI（好时间间隔）是每条 XMM 光谱背后的第一道筛选 —— 但它的规则从未被判决过。Soft protons flare at any time; GTIs (good time intervals) are the first screen behind every XMM spectrum — but their rules have never been put on trial.

承诺：45 分钟后你会知道 (1) 一个只用光变曲线的政策族能走多远，(2) 一次冻结规则的保留集判决怎么说，(3) 你的数据该不该换规则。Promise: in 45 minutes you will know (1) how far a light-curve-only policy family gets, (2) what a frozen-rule held-out verdict says, (3) whether your data should switch rules.

22

路线图Roadmap

  1. 1. 背景：EPIC、软质子、为什么弥散发射最受伤1. Background: EPIC, soft protons, why diffuse emission suffers most
  2. 2. 数据与方法：119 个保留观测、83 个政策、冻结的判决规则2. Data and method: 119 held-out observations, 83 policies, a frozen verdict rule
  3. 3. 结果 1：83 个里 79 个 Confirmed —— 但参照很弱3. Result 1: 79 of 83 Confirmed — but the reference is weak
  4. 4. 结果 2：等保留时间下优势未确立4. Result 2: at matched kept time the advantage is not established
  5. 5. 结果 3：判据的边界 —— 三次收窄5. Result 3: the criterion's boundaries — three narrowings
  6. 6. 这意味着什么、下一步6. What it means, and what follows
33

EPIC 把每 1000 秒的天空放进一张谱里 —— 而软质子随时都在敲门EPIC puts the sky into a spectrum every 1000 s — and soft protons knock at any moment

XMM-Newton 的 EPIC 相机（MOS1、MOS2、pn）每 1000 秒读出一帧；我们的打分把每条光变曲线切成 1000 s 的段。角外视野（"角落"）通常被当作粒子监控器，但角落本身也吃天光。XMM-Newton's EPIC cameras (MOS1, MOS2, pn) read out a frame every 1000 s; our scoring cuts each light curve into 1000 s pieces. The out-of-fields view ("corners") is often used as a particle monitor, but the corners themselves see some sky.

软质子：被磁层聚焦的低能带电粒子，落在探测器上的时间分布无规则，谱形与弥散热发射重叠。Soft protons: low-energy charged particles focused by the magnetosphere, arriving at irregular times, with a spectral shape that overlaps diffuse thermal emission.

44

点源可以减背景，弥散发射不能 —— 它就是背景Point sources can subtract a background; diffuse emission cannot — it is the background

一个 5% 的系统误差在点源工作里可以忽略，在弥散工作里是判决级的：观测间的水平差直接进入物理量。A 5% systematic that a point-source analysis can ignore is verdict-level for diffuse work: the between-observation level difference enters the physical quantity directly.

所以我们选的判据是观测间一致性 \(\sigma_R\)：同一片天空、同一批政策下，观测与观测之间 0.5–2 keV 拼接光变曲线水平的散布。So the criterion we chose is between-observation consistency \(\sigma_R\): the scatter of the 0.5-2 keV concatenated light-curve level from observation to observation, on the same sky and under the same policy.

55

判据：观测间散布 \(\sigma_R\)，不是"谁删得多"The criterion: between-observation scatter \(\sigma_R\), not "who cuts more"

\(\sigma_R\) 的定义：格内逐观测配对自举（2000 次抽样，种子 20261015）的 p16/p84。Definition of \(\sigma_R\): p16/p84 from a per-observation paired bootstrap within the cell (2000 draws, seed 20261015).

它不含观测间水平项，也不直接奖励多删 —— 但多删能降低散布，所以保留时间必须与判决同报。It carries no between-observation level term and does not directly reward cutting more — but cutting more lowers the scatter, so kept time must be reported with the verdict.

66

83 个政策、23 个脚本 —— 只用光变曲线，从不看角落83 policies, 23 scripts — light curves only, never the corners

政策族：对全视场 0.5–2 keV 光变曲线做阈值、斜率、多尺度形状的筛选，产出一个 GTI；83 个候选覆盖 23 个"家族"（同一规则的不同参数）。The policy family: thresholds, slopes and multi-scale shapes screened on the full-field 0.5-2 keV light curve, producing a GTI; 83 candidates cover 23 "families" (the same rule at different parameters).

全部政策只用光变曲线；角落数据从不进入筛选决策（这是与文献方法的核心区别）。All policies use light curves only; corner data never enter the screening decision (this is the core difference from the literature methods).

77

规则先冻结，分数后打开 —— 清单 sha 8598b555…Freeze the rule first, open the scores after — manifest sha 8598b555…

2026-10-18T15:33Z 冻结：83 个候选、23 个脚本、11 条命令；此后只有"可执行回归失败"能改代码，且每次改动带失败回归测试。Frozen at 2026-10-18T15:33Z: 83 candidates, 23 scripts, 11 commands; after that only a failing executable regression may change code, and every change carries a failing regression test.

一次冻结链运行在保留集观测 0302351401 的空天空坐标上停止；修复带 FAIL \(\to\) PASS 回归，发展数据逐位不变。One frozen-chain run stopped on null sky coordinates in held-out observation 0302351401; the fix carries a FAIL \(\to\) PASS regression and leaves development data bit-identical.

88

119 个从未参与设计的观测，六个天空格119 observations that took no part in the design, in six sky cells

六个天空格：两个开发格（设计与调参用）+ 两个复验格（收紧用）+ 两个新天空格（判决用）。119 个保留观测从未进入任何政策的参数选择。Six sky cells: two development (design and tuning), two confirmation (tightening), two new-sky (verdict). The 119 held-out observations never entered any policy's parameter choice.

主参照：归档 2.filter（SAS 标准好时间间隔筛选）；次参照：复现滤波器（espfilt + 硬段裁剪），204 个非候选政策里 50 个也通过同一条规则。Primary reference: the archived 2.filter (the SAS standard good-time screening); secondary: a reproduced filter (espfilt + a hard-band clip), which 50 of 204 non-candidate policies also pass.

99

判决规则：至少 4 个格增益、两个新天空格都增益、无伤害格The verdict rule: at least 4 cells gain, both new-sky cells gain, no harm cell

Confirmed 当且仅当：至少 4 个格增益（p84(\(\delta\)) \(<-0.5\) pp）、两个新天空格都增益、且没有格变差超过阈值。Ambiguous / Not confirmed 按冻结文本。Confirmed if and only if: at least 4 cells gain (p84(\(\delta\)) \(<-0.5\) pp), both new-sky cells gain, and no cell worsens beyond threshold. Ambiguous / Not confirmed follow the frozen text.

83 个候选逐一与两个参照对比；每个数字都可追溯到冻结清单。Each of the 83 candidates is compared against both references; every number traces to the frozen manifest.

1010

读数分三层：归档参照 → 复现参照 → 等保留时间The reading has three layers: archived reference \(\to\) reproduced reference \(\to\) matched kept time

层Layer 参照Reference 结果Result
1归档 2.filterarchived 2.filter79/83 Confirmed79/83 Confirmed
2复现滤波器reproduced filter11/8311/83
3等保留时间matched kept time0–1/83（配对自举）0-1/83 (paired bootstrap)
1111

三个结果，每个说三遍：预览、展示、总结Three results, each said three times: preview, show, summarise

  1. 1. 79/83 相对归档参照 Confirmed —— 判决是真的1. 79/83 Confirmed against the archived reference — the verdict is real
  2. 2. 只剩 11/83 相对复现滤波器 —— 参照很弱2. Only 11/83 survive the reproduced filter — the reference was weak
  3. 3. 等保留时间下 0–1/83 —— 优势未确立3. 0-1/83 at matched kept time — the advantage is not established
1212

第 1 层：79/83 降低观测间散布 —— 图 1aLayer 1: 79/83 lower the between-observation scatter — Figure 1a

The held-out verdict: 79/83 Confirmed against the archived 2.filter, 11/83 against the reproduced filter, 0-1/83 at matched kept time

横轴是 83 个候选（按家族排序），纵轴是配对差 \(\delta\)（百分点，pp）。蓝点：相对归档 2.filter；红点：相对复现滤波器。误差条是逐观测配对自举的 p16/p84。来源：E024；heldout_p006/verdict.json。The x-axis is the 83 candidates (ordered by family), the y-axis the paired difference \(\delta\) (percentage points, pp). Blue: vs the archived 2.filter; red: vs the reproduced filter. Error bars are the p16/p84 of a per-observation paired bootstrap. Source: E024; heldout_p006/verdict.json.

1313

增益大多是拼接层面的：73/79 "集中"The gain is mostly mosaic-level: 73/79 "concentrated"

"集中"：增益集中在少数格，不是均匀铺开。HO-M31 MOS1 的 \(-10.1\) pp 里约 4.6 pp 来自一个坏参照文件；单个观测的剔除能让等保留时间的计数在 0 到 21 之间摆动。"Concentrated": the gain sits in a few cells, not spread evenly. About 4.6 pp of the HO-M31 MOS1 \(-10.1\) pp is one broken reference file; removing a single observation swings the matched-time count between 0 and 21.

23 个家族全部至少有一个 Confirmed；中位改善 2.9–10.1 pp。All 23 families have at least one Confirmed; the median improvement is 2.9-10.1 pp.

1414

第 2 层：复现滤波器拿走一半以上的增益Layer 2: the reproduced filter takes more than half the gain

一个 espfilt 加硬段裁剪的复现滤波器 —— 标准做法的忠实复现 —— 已经拿到最优候选增益的一半以上。相对它只剩 11/83，且全部保留更少时间（0.32–0.76 对 0.81–0.93）。A reproduced filter — espfilt plus a hard-band clip, a faithful reproduction of standard practice — already delivers more than half of the best candidate's gain. Against it only 11/83 survive, all keeping less time (0.32-0.76 vs 0.81-0.93).

全部 11 个幸存者都靠删更多时间；这是下一层的伏笔。All 11 survivors cut more time; that sets up the next layer.

1515

第 3 层：等保留时间下配对自举只剩 0–1 个Layer 3: at matched kept time a paired bootstrap leaves 0-1

把每个候选与"保留至少同样多时间的最好复现滤波器设置"比较（48 个设置的阶梯，每次抽样重画），逐观测配对自举。点估计下 37/83 领先，配对自举下只剩 0–1 个 Confirmed（\(\delta_0=0.5\) pp）。Each candidate is compared with the best reproduced-filter setting that keeps at least as much time (a ladder of 48 settings, redrawn per draw), by per-observation paired bootstrap. Under the point estimate 37/83 lead; under the paired bootstrap only 0-1 remain Confirmed (\(\delta_0=0.5\) pp).

结论：优势未确立。这是预先注册的下一个测试（P007）要回答的问题。Conclusion: the advantage is not established. That is the question the pre-registered next test (P007) answers.

1616

四条局限，每条带数字Four limitations, each with its number

  1. 参照强度：204 个非候选里 50 个也过同一条规则。Reference strength: 50 of 204 non-candidates also pass the same rule.
  2. 保留时间不在判决里（PI 决定）；等保留时间读数功效接近零。Kept time is outside the verdict (the PI's decision); the matched-time reading has near-zero power.
  3. 增益构成：73/79 集中；\(-10.1\) pp 里 4.6 pp 是坏参照文件。Gain composition: 73/79 concentrated; 4.6 pp of the \(-10.1\) pp is a broken reference file.
  4. 未检验：Al-K 线门（0–3 个观测/格）；谱验证未做。Untested: the Al-K line gate (0-3 observations per cell); spectral validation not done.
1717

三个带回家的消息Three take-home messages

  1. 79/83 相对归档参照 Confirmed —— 光变曲线筛选值得继续。79/83 Confirmed against the archived reference — light-curve screening is worth continuing.
  2. 11/83 相对复现滤波器 —— 大半增益来自"删更多"，不是"删得准"。11/83 against the reproduced filter — most of the gain is "cut more", not "cut right".
  3. 0–1/83 等保留时间 —— 优势未确立；P007 预注册回答它。0-1/83 at matched kept time — the advantage is not established; the pre-registered P007 answers it.

数据截至 2026-10-19T10:38Z（D079）。致谢：XMM-Newton（ESA/NASA）；SAS/ESAS；本手稿用 AI 助手辅助编写，科学判断由作者负责。Data as of 2026-10-19T10:38Z (D079). Acknowledgments: XMM-Newton (ESA/NASA); SAS/ESAS; AI assistants helped prepare this manuscript; scientific judgement is the authors'.

1818

备用 1：三层读数的完整表Backup 1: the full three-layer reading table

层LayerConfirmedConfirmed来源Source
归档 2.filterarchived 2.filter79/83E024; D068
复现滤波器reproduced filter11/83E024; D070
等保留时间（配对自举）matched kept time (paired bootstrap)0–1/83E024; D073
B1B1

备用 2：冻结链与一次修复Backup 2: the freeze chain and one fix

冻结：2026-10-18T15:33Z；清单 sha 8598b555…；运行版 v2 sha c7aad3f8…。修复：保留集观测 0302351401 的空天空坐标；FAIL \(\to\) PASS 回归；发展数据逐位不变。Freeze: 2026-10-18T15:33Z; manifest sha 8598b555…; run version v2 sha c7aad3f8…. Fix: null sky coordinates in held-out observation 0302351401; FAIL \(\to\) PASS regression; development data bit-identical.

B2B2

备用 3：P007 预注册草案Backup 3: the P007 pre-registration draft

主估计量：六格等权平均 \(\bar D\) 对复现滤波器；\(\delta_0=0.5\) pp；伤害守卫；六格全有效；规模 \(\leq 3\%\)；判决代码已写好并在发展数据上自检通过（D079）。Primary estimand: the six-cell equal-weight mean \(\bar D\) against the reproduced filter; \(\delta_0=0.5\) pp; harm guard; all cells valid; size \(\leq 3\%\); the verdict code is written and self-checked on development data (D079).

B3B3

讲稿（每页的时间为累计）Talk script (times are cumulative)

【1，0:00–1:00】各位好。今天讲一个很实际的问题：XMM 的好时间间隔筛选能不能只用光变曲线来做，以及它的规则第一次被冻结判决的结果。45 分钟，最后留 10 到 15 分钟提问。[1, 0:00-1:00] Hello. Today: can XMM good-time-interval screening be done from light curves alone, and what the first frozen verdict on such rules says. 45 minutes, then 10-15 minutes of questions.

【2，1:00–3:00】先说痛点。同一片天空的两次观测，背景水平可以差 20%，而弥散工作的科学预算只有 5%。软质子随时涨落，GTI 是每条光谱背后的第一道筛选，但它的规则从来没被判决过。今天你们会带走三件事：政策族能走多远、冻结判决怎么说、你的数据该不该换规则。[2, 1:00-3:00] First the pain. Two observations of the same sky can differ by 20% in background level, while diffuse work has a 5% science budget. Soft protons flare at any time; the GTI is the first screen behind every spectrum, but its rules have never been put on trial. You will take away three things: how far the policy family gets, what the frozen verdict says, whether your data should switch rules.

【3，3:00–4:00】路线图：背景、数据与方法、三个结果、意义与下一步。[3, 3:00-4:00] Roadmap: background, data and method, three results, meaning and next steps.

【4，4:00–6:00】EPIC 每 1000 秒读出一帧，我们的打分也切成 1000 秒的段。角落常被当粒子监控器用，但角落也吃天光。软质子是磁层聚焦的低能带电粒子，时间上无规则，谱形与弥散热发射重叠 —— 这就是它难缠的原因。[4, 4:00-6:00] EPIC reads out a frame every 1000 s, and our scoring cuts light curves into 1000 s pieces. Corners are often used as particle monitors, but corners see sky too. Soft protons are low-energy charged particles focused by the magnetosphere, irregular in time, spectrally overlapping diffuse thermal emission — that is why they are hard.

【5，6:00–8:00】为什么弥散发射最受伤？点源可以减背景，弥散发射不能 —— 它就是背景。5% 的系统误差在点源工作里可以忽略，在弥散工作里是判决级的。所以我们选观测间一致性做判据。[5, 6:00-8:00] Why does diffuse emission suffer most? Point sources can subtract a background; diffuse emission cannot — it is the background. A 5% systematic a point-source analysis ignores is verdict-level for diffuse work. So our criterion is between-observation consistency.

【6，8:00–10:00】判据是 \(\sigma_R\)：同一片天空同一批政策下，观测与观测之间 0.5–2 keV 拼接光变曲线水平的散布，逐观测配对自举给 p16/p84。注意它不含观测间水平项，也不直接奖励多删 —— 但多删能降散布，所以保留时间必须与判决同报。[6, 8:00-10:00] The criterion is \(\sigma_R\): the scatter of the 0.5-2 keV concatenated light-curve level from observation to observation on the same sky under the same policy, with p16/p84 from a per-observation paired bootstrap. It carries no between-observation level term and does not directly reward cutting more — but cutting more lowers the scatter, so kept time must be reported with the verdict.

【7，10:00–12:00】政策族：83 个候选、23 个家族，只用全视场 0.5–2 keV 光变曲线做阈值、斜率、多尺度形状筛选。角落数据从不进入决策 —— 这是与文献方法的核心区别。[7, 10:00-12:00] The policy family: 83 candidates in 23 families, screening the full-field 0.5-2 keV light curve by thresholds, slopes and multi-scale shapes. Corner data never enter the decision — the core difference from the literature.

【8，12:00–14:00】规则先冻结：2026-10-18T15:33Z，83 个候选、23 个脚本、11 条命令，清单哈希存档。之后只有失败的可执行回归能改代码。一次运行在保留集观测的空天空坐标上停住，修复带回归测试，发展数据逐位不变。[8, 12:00-14:00] Freeze the rule first: 2026-10-18T15:33Z, 83 candidates, 23 scripts, 11 commands, manifest hashed. Afterwards only a failing executable regression may change code. One run stopped on null sky coordinates in a held-out observation; the fix carries a regression test and leaves development data bit-identical.

【9，14:00–16:00】保留集：119 个从未参与设计的观测，六个天空格 —— 两个开发、两个复验、两个新天空。主参照是归档 2.filter，次参照是复现滤波器。[9, 14:00-16:00] The held-out set: 119 observations that took no part in the design, in six sky cells — two development, two confirmation, two new-sky. Primary reference: the archived 2.filter; secondary: a reproduced filter.

【10，16:00–18:00】判决规则：至少 4 个格增益、两个新天空格都增益、无伤害格。增益的阈值是 p84 差 \(<-0.5\) 个百分点。[10, 16:00-18:00] The verdict rule: at least 4 cells gain, both new-sky cells gain, no harm cell. The gain threshold is a p84 difference \(<-0.5\) percentage points.

【11，18:00–20:00】读数分三层，这张表是全场的骨架：归档参照 79/83，复现参照 11/83，等保留时间 0–1/83。下面逐层展开。[11, 18:00-20:00] The reading has three layers; this table is the skeleton of the whole talk: 79/83 against the archived reference, 11/83 against the reproduced filter, 0-1/83 at matched kept time. Now each layer.

【12，20:00–21:00】三个结果，每个我都会说三遍：预览、展示、总结。[12, 20:00-21:00] Three results; I will say each three times: preview, show, summary.

【13，21:00–24:00】看这张图。横轴是 83 个候选，纵轴是配对差 \(\delta\)（百分点）。蓝点对归档 2.filter，红点对复现滤波器，误差条是配对自举的 p16/p84。大部分蓝点在零线以下 —— 79/83 Confirmed，中位改善 2.9 到 10.1 个百分点。[13, 21:00-24:00] Look at this figure. The x-axis is the 83 candidates; the y-axis is the paired difference \(\delta\) in percentage points. Blue: vs the archived 2.filter; red: vs the reproduced filter; error bars are the paired-bootstrap p16/p84. Most blue points sit below zero — 79/83 Confirmed, median improvement 2.9 to 10.1 percentage points.

【14，24:00–26:00】但增益大多是拼接层面的：73/79 集中在少数格。HO-M31 MOS1 的 \(-10.1\) pp 里约 4.6 pp 是一个坏参照文件。这提示我们：判决是真的，但增益的构成要拆开看。[14, 24:00-26:00] But the gain is mostly mosaic-level: 73/79 concentrated in a few cells. About 4.6 pp of the HO-M31 MOS1 \(-10.1\) pp is one broken reference file. The verdict is real, but the composition of the gain must be taken apart.

【15，26:00–29:00】第 2 层：一个 espfilt 加硬段裁剪的复现滤波器 —— 标准做法的忠实复现 —— 已经拿到最优候选增益的一半以上。相对它只剩 11/83，而且全部保留更少时间：0.32–0.76 对 0.81–0.93。[15, 26:00-29:00] Layer 2: a reproduced filter — espfilt plus a hard-band clip, a faithful reproduction of standard practice — already delivers more than half of the best candidate's gain. Against it only 11/83 survive, all keeping less time: 0.32-0.76 vs 0.81-0.93.

【16，29:00–33:00】第 3 层最关键：把保留时间配平再比。每个候选与"保留至少同样多时间的最好复现滤波器设置"比较，48 个设置的阶梯，逐观测配对自举。点估计下 37/83 领先，配对自举下只剩 0–1 个。结论：优势未确立。这是预先注册的下一个测试要回答的问题。[16, 29:00-33:00] Layer 3 is the crux: match the kept time, then compare. Each candidate faces the best reproduced-filter setting that keeps at least as much time — a ladder of 48 settings — under a per-observation paired bootstrap. Under the point estimate 37/83 lead; under the bootstrap only 0-1 remain. Conclusion: the advantage is not established. The pre-registered next test answers exactly this.

【17，33:00–36:00】四条局限：参照强度（204 个非候选里 50 个也过同一条规则）；保留时间不在判决里；增益构成（73/79 集中）；未检验的部分（Al-K 线门只在 0–3 个观测上触发，谱验证未做）。[17, 33:00-36:00] Four limitations: reference strength (50 of 204 non-candidates also pass the same rule); kept time outside the verdict; gain composition (73/79 concentrated); the untested parts (the Al-K line gate fires on 0-3 observations per cell; spectral validation not done).

【18，36:00–45:00】总结。79/83 相对归档参照 —— 光变曲线筛选值得继续。11/83 相对复现滤波器 —— 大半增益来自删更多。0–1/83 等保留时间 —— 优势未确立，P007 预注册回答它。判据、规则、代码全部冻结可查。谢谢。[18, 36:00-45:00] Summary. 79/83 against the archived reference — light-curve screening is worth continuing. 11/83 against the reproduced filter — most of the gain is cutting more. 0-1/83 at matched kept time — the advantage is not established; the pre-registered P007 answers it. The criterion, the rule and the code are all frozen and checkable. Thank you.

预期问答Anticipated Q&A

问：为什么不直接用 ESAS？答：ESAS 是建模背景，不是筛选时间；两者互补，我们比较的是筛选规则本身（备用 1）。Q: Why not just use ESAS? A: ESAS models the background rather than screening time; the two are complementary, and we compare the screening rules themselves (Backup 1).

问：等保留时间下功效够吗？答：接近零，所以那里是描述性读数；P007 提高功效（备用 3）。Q: Is the matched-kept-time comparison well powered? A: Near zero, so there it is descriptive; P007 raises the power (Backup 3).

问：角落为什么不用？答：角落也吃天光，且我们要求政策可迁移到任何观测（备用 1）。Q: Why not use the corners? A: Corners see sky too, and we require policies transferable to any observation (Backup 1).

问：软质子事件会不会与真实变源混淆？答：会，所以判据是观测间一致性，不是单观测内的形状（备用 1）。Q: Could soft-proton flares be confused with real variability? A: Yes, which is why the criterion is between-observation consistency, not within-observation shape (Backup 1).

问：\(\sigma_R\) 为什么不含水平项？答：冻结时的选择；水平项在拼接图上另测（备用 2）。Q: Why does \(\sigma_R\) carry no level term? A: A freeze-time choice; the level term is measured separately on mosaics (Backup 2).

问：坏参照文件那 4.6 pp 会不会推翻判决？答：不会，它是 79 个中的一个格，且已从增益中拆出（备用 2）。Q: Does the 4.6 pp from the broken reference file overturn the verdict? A: No; it is one cell of the 79 and has been separated from the gain (Backup 2).

问：结果对别的望远镜通用吗？答：方法通用（光变曲线 + 冻结判决），数值不通用（备用 3）。Q: Do the results generalise to other telescopes? A: The method does (light curves + a frozen verdict); the numbers do not (Backup 3).

问：为什么不用深度学习？答：83 个可解释规则已经足够暴露判据的结构，且可冻结（备用 3）。Q: Why not deep learning? A: 83 interpretable rules already expose the structure of the criterion, and they can be frozen (Backup 3).

总时长 45:00（目标 45 分钟，偏差 0%）。Total 45:00 (target 45 min, deviation 0%).

00:00