论文草稿（未投稿；数据截至 2026-10-19T10:38Z）Paper draft (not submitted; data as of 2026-10-19T10:38Z)

标题与作者Title and authors

标题：用光变曲线筛选 XMM-Newton EPIC 好时间段：一次保留集判决，以及它相对复现标准滤波器的限度

Title: Screening XMM-Newton EPIC good time intervals from light curves alone: a held-out verdict and its limits against a reproduced standard filter

Running head: Held-out test of XMM GTI screening

Running head: Held-out test of XMM GTI screening

作者：<Author list — PI to decide>^[s0]

Authors: <Author list — PI to decide>^[s0]

0. 摘要0. Abstract

XMM-Newton EPIC 的观测被软质子耀斑污染，标准做法是用好时间段（good time interval, GTI）把耀斑时间删掉。删多删少是一对矛盾：删得多，剩下的数据更干净但更少。我们把这个取舍当作单次观测内的帕累托问题，用保留活时间与连续谱约减因子两个坐标给每个候选 GTI 打分，并要求它只从光变曲线构造。在 119 个从未参与任何设计或调参的观测上，我们做了第一次冻结的保留集判决：83 个候选中 79 个相对归档 2.filter 降低了观测间散布（中位改善 \(2.9\text{–}10.1\) 个百分点，23 个策略族，无伤害格）。但这个参照本身很弱 —— 204 个非候选策略里也有 50 个通过同一条规则；相对一个从同一批光变曲线复现的标准光变曲线滤波器（espfilt 加硬段裁剪），只剩 11 个候选通过，且全部保留更少时间；在等保留时间下用配对自举比较只剩 0\(\text{–}\)1 个，优势未确立。最大的表面增益（M31 一台相机上 10.1 个百分点）约 4.6 个百分点来自一个损坏的归档筛选器文件。下一个测试的规则与判决代码已经写好，参照是那个复现的标准滤波器。

XMM-Newton EPIC observations are polluted by soft-proton flares, and the standard practice is to remove the flaring time with good time intervals (GTIs). How much to cut is a genuine conflict: cut more and the retained data are cleaner but fewer. We treat the trade-off as a Pareto problem inside a single observation, scoring every candidate GTI by retained live time and a continuum reduction factor, and requiring that it be built from light curves alone. On 119 observations that took no part in any design or tuning, we ran the first frozen held-out verdict: 79 of 83 candidates lower the between-observation scatter relative to the archived 2.filter (median 2.9\(\text{–}\)10.1 percentage points, 23 families, no harm cell). That reference is weak: 50 of 204 non-candidate policies also pass the same rule; against a standard light-curve filter reproduced from the same curves (espfilt plus a hard-band clip) only 11 candidates survive, all keeping less time; at matched kept time a paired bootstrap leaves 0\(\text{–}\)1, so the advantage is not established. The largest apparent gain (10.1 percentage points on one M31 camera) is about 4.6 points from one broken archived filter file. The rule and verdict code for the next test, with the reproduced filter as reference, are written.

关键词Keywords

X 射线天文学；弥散 X 射线背景；天文数据还原；仪器效应；软质子；预注册检验

X-ray astronomy; diffuse X-ray background; astronomy data reduction; instrumentation; soft protons; pre-registered testing

1. 引言1. Introduction

每当 XMM-Newton 望远镜的 EPIC 相机观测一个延展的 X 射线源，它同时也在记录粒子背景。软质子 —— 被地球磁层聚焦、能量低到能穿过望远镜的低能质子 —— 随时间起伏，最坏的时候会把观测数据完全淹没。任何测量弥散发射的人都要先把这段时间删掉，这靠"好时间段"（good time interval, GTI）文件完成：保留干净的时间，删除耀斑的时间。这个决定直接进入最后的误差预算，因为删掉的每一秒都是丢失的曝光。

Every time the EPIC cameras aboard XMM-Newton observe an extended X-ray source, they also record the particle background. Soft protons — low-energy protons focused by the Earth's magnetosphere, energetic enough to pass through the optics — rise and fall on all timescales, and at their worst they swamp the observation entirely. Anyone measuring diffuse emission must first remove that time, and the removal is done with good time intervals (GTIs): keep the clean time, cut the flaring time. The decision enters the final error budget directly, because every second cut is exposure lost.

标准做法（"2.filter"）在 ESAS 软件包里：对 2.5–8 keV 的全视场光变曲线做迭代 \(2\sigma\) 剔除，把耀斑时间删掉。这个做法有一个众所周知的代价：它会把耀斑的"静息水平"也当作基线保留下来，而且在一个完全被耀斑覆盖的观测里，它可能锁到耀斑的峰值上、几乎删不掉任何东西。文献里的替代方案各有取舍：IN/OUT 比值法用未曝光的角落做粒子监测器；ESAS 的粒子背景模型用闭门 (filter-wheel-closed) 数据与角落归一化；Nevalainen 等人的双带监控器用视场外圈的两个能段做剔除。我们在此要回答的问题更窄也更基本：一个只看光变曲线、不读归档筛选器、不读角落的筛选规则，在从未参与它设计的观测上，到底比标准做法好多少？

The standard practice ("2.filter") lives in the ESAS package: an iterative \(2\sigma\) clipping of the 2.5–8 keV field-of-view light curve. Its costs are well known: it keeps the flare's quiescent level as a baseline, and in an observation fully covered by a flare it can lock onto the flare peak and cut almost nothing. The alternatives in the literature each trade something away: the IN/OUT ratio uses the unexposed corner as a particle monitor; the ESAS particle-background model uses filter-wheel-closed data normalised by the corner; the two-band monitor of Nevalainen et al. screens with two energy bands in the outer field of view. The question we ask here is narrower and more basic: a screening rule that reads only light curves — not the archived filter, not the corner — beats the standard practice by how much, on observations that took no part in its design?

这个问题之所以还没有答案，是因为"更好"需要一个判据，而判据本身就是研究的一部分。单次观测内部，删多删少可以用帕累托前沿描述：保留活时间 \(f_T\) 对连续谱约减因子 \(R\)。但 \(R\) 不能跨观测比较 —— 同一个候选在 30 次观测里 \(R\) 从 1.0 到 11.8^[s1]。跨观测的成功判据必须回答"一组观测放在一起时，结果一致吗"，这正是弥散发射测量的实际需求：把多次观测拼接成一个天空的地图时，观测间的不一致直接成为系统误差。本项目的误差预算测量给出：在 2.filter 下，0.5–2 keV 拼接光变曲线的系统散布 \(s_{\rm sys}\) 是 17–25%，而统计误差只有 0.04–0.2%^[s2] —— 系统误差约大 \(100\) 倍，且它不随观测数下降。所以"更好"的第一判据是：降低观测间的散布 \(\sigma_R\)，而不是提高单次观测的信噪比。

The question has no answer yet because "better" needs a criterion, and the criterion is itself part of the research. Inside one observation the cut-more-or-less trade-off is described by a Pareto front: retained live time \(f_T\) against the continuum reduction factor \(R\). But \(R\) cannot be compared across observations — the same candidate spans \(R\) from 1.0 to 11.8 over 30 observations^[s1]. A cross-observation success criterion must answer "when several observations are put together, do the results agree", which is the practical need of diffuse-emission work: when observations are mosaicked into one sky map, disagreement between them becomes a systematic error directly. This project's error-budget measurement gives: under 2.filter, the systematic scatter \(s_{\rm sys}\) of the 0.5–2 keV concatenated light curves is 17–25%, against a statistical error of only 0.04–0.2%^[s2] — the systematic is about \(100\times\) larger, and it does not fall with the number of observations. So the first criterion for "better" is: lower the between-observation scatter \(\sigma_R\), not raise the single-observation signal-to-noise.

本文做三件事。第一，我们给出这个判据下的第一次冻结的保留集判决：83 个只用光变曲线构造的候选政策，在 119 个从未参与任何设计或调参的观测上，与归档 2.filter 比较。第二，我们把这个判决收窄三次：同一个参照本身有多弱、对复现的标准滤波器还剩多少、以及在等保留时间下优势是否还能确立。第三，我们给出下一个测试的预先注册规则，把"比一个复现的标准滤波器更好"作为主问题，并写出它的判决代码。本文的结构：§2 给数据与还原；§3 按顺序给出判决与三次收窄；§4 讨论它的含义、局限与对同行的建议；§5 小结。

This paper does three things. First, it reports the first frozen held-out verdict under that criterion: 83 candidate policies built from light curves alone, compared with the archived 2.filter on 119 observations that took no part in any design or tuning. Second, it narrows that verdict three times: how weak the reference itself is, how much survives against a reproduced standard filter, and whether the advantage stands at matched kept time. Third, it gives the pre-registered rule for the next test, which makes "better than a reproduced standard filter" the primary question, and its verdict code. Roadmap: §2 gives the data and reduction; §3 reports the verdict and its three narrowings in order; §4 discusses what they mean, their limits, and the practical advice for others; §5 summarises.

2. 观测与数据还原2. Observations and data reduction

2.1 数据2.1 Data

我们用 XMM-Newton EPIC MOS1 与 MOS2（pn 只作交叉检验）。发展集是四个已知的天空集：M31 晕（31 个指向）、CDFS 空白场（34 个 ObsID）、Lockman Hole（13 个）与一个 Virgo 归档镜像集（142 个指向）。保留集由规则固定、与任何发展观测无重叠：119 个观测，分三类 —— 11 个 M31 指向（与发展的 M31 天空相邻但从未用过）、48 个深空白场指向、60 个"混合"指向（EXO 0748-676、A3266 等），共 130 个候选观测中 11 个因管道失败被记录并排除^[s3]。下载约 29 GB ODF；还原在专用服务器上用 SAS/ESAS 完成，本机只做分析与打分。

We used XMM-Newton EPIC MOS1 and MOS2 (pn only as a cross-check). The development data were four known sky sets: the M31 halo (31 pointings), the CDFS blank field (34 ObsIDs), the Lockman Hole (13), and a Virgo archive mirror (142 pointings). The held-out set was fixed by rule and shares no observation with any development data: 119 observations in three classes — 11 M31 pointings (adjacent to the development M31 sky but never used), 48 deep blank-field pointings, and 60 "mixed" pointings (EXO 0748-676, A3266 and others), out of 130 candidate observations with 11 excluded for documented pipeline failures^[s3]. The download was about 29 GB of ODF; the reduction ran with SAS/ESAS on a dedicated server, and all analysis and scoring ran locally.

2.2 候选政策的构造规则2.2 How the candidate policies were built

每个候选政策只读光变曲线：2.5–8 keV 的全视场率（硬带 \(h\)）、8–12 keV 的率（\(u\)）、活时间与中点。它不读归档 2.filter 的任何信息、不读未曝光角落的任何计数、不读被评分的 0.5–2 keV 能段。83 个候选来自三次政策搜索（两代机器搜索加一次更广的第三代），每个都由一个发展集筛选选出：在与 2.filter 相同的观测上、按逐观测配对的 \(\delta=\sigma_R(\text{policy})-\sigma_R(\text{2.filter})\) 计，8 个"天空集 \(\times\) 相机"格中至少 6 个格 \(\delta<0\) 且无伤害格^[s4]。独立性由两种置换检验复核：把角落计数置平、把被评分能段的计数置乱，83 个候选的掩模一次都不变（36,520 个"观测 \(\times\) 相机"行）^[s5]。

Every candidate policy reads only light curves: the 2.5–8 keV field-of-view rate (the hard band \(h\)), the 8–12 keV rate (\(u\)), live time and midpoints. It reads no information from the archived 2.filter, no counts from the unexposed corner, and no counts from the 0.5–2 keV band being scored. The 83 candidates came from three policy searches (two machine-search generations plus a broader third), each selected by a development screen: on the same observations as 2.filter, by the per-observation paired \(\delta=\sigma_R(\text{policy})-\sigma_R(\text{2.filter})\), at least 6 of the 8 sky-set \(\times\) camera cells with \(\delta<0\) and no harm cell^[s4]. Independence was re-checked by two permutation tests — flattening the corner counts and permuting the scored band's counts — and none of the 83 candidates' masks changed once (36,520 observation \(\times\) camera rows)^[s5].

2.3 打分与判决2.3 Scoring and the verdict

打分量是 \(\sigma_R\)：一个"天空集 \(\times\) 相机"格内、观测间 0.5–2 keV 水平的散布，按逐观测配对自举（2000 次抽样，种子固定）取 p16/p84。打分用 1000 s 的分段；分段长度的敏感性作为次要读数。判决规则在看到任何保留集分数之前冻结：一个候选"Confirmed"当且仅当至少 4 个格增益（\(\delta\) 的 p84 \(<-\delta_0\)，\(\delta_0=0.5\) 个百分点）、两个新天空格都增益、且没有任何格伤害；83 个候选、23 个脚本与 11 条命令全部哈希入清单，此后只有"可执行回归失败"能改代码^[s6]。拟合细节与打分器的完整定义在附录 A。

The scored quantity is \(\sigma_R\): the scatter of the 0.5–2 keV level between observations within a sky-set \(\times\) camera cell, with p16/p84 from a per-observation paired bootstrap (2000 draws, fixed seed). Scoring used 1000 s pieces; the piece-length sensitivity is a secondary reading. The verdict rule was frozen before any held-out score was seen: a candidate is "Confirmed" if and only if at least 4 cells gain (p84 of \(\delta\) \(<-\delta_0\), \(\delta_0=0.5\) percentage points), both new-sky cells gain, and no cell is harmed; the 83 candidates, 23 scripts and 11 commands were all hashed into a manifest, after which only a failing executable regression could change code^[s6]. Fitting details and the full scorer definition are in Appendix A.

3. 结果3. Results

本节按论证顺序给出四个结果：判决本身（§3.1）、参照有多弱（§3.2）、对复现滤波器还剩多少（§3.3）、等保留时间下是否确立（§3.4）。图 1 把四个读法放在一起。

This section reports four findings in the order that builds the answer: the verdict itself (§3.1), how weak the reference is (§3.2), how much survives against a reproduced filter (§3.3), and whether the advantage stands at matched kept time (§3.4). Figure 1 puts the four readings side by side.

3.1 判决：79 of 83 相对归档 2.filter Confirmed3.1 The verdict: 79 of 83 Confirmed against the archived 2.filter

我们首先问最直接的问题：在从未参与发展的观测上，这些只用光变曲线的政策是否降低了观测间的散布。答案由冻结规则给出：83 个候选中 79 个 Confirmed，4 个 Ambiguous，0 个 Not confirmed，分属 23 个策略族；没有一个 Confirmed 候选有任何伤害格^[s7]。中位 \(\delta\) 从 HO-M31 MOS1 的 \(-10.1\) pp 到 HO-mixed 的 \(-2.9\) / \(-3.5\) pp（图 1a）。尺度参照：HO-blank MOS1 上 2.filter 的 \(\sigma_R\) 是 8.58%，候选是 3.87–11.72%，完全不筛选是 47.3%^[s8]。

We first asked the most direct question: on observations that took no part in development, do these light-curve-only policies lower the between-observation scatter? The frozen rule answers: 79 of 83 candidates Confirmed, 4 Ambiguous, 0 Not confirmed, in 23 policy families; no Confirmed candidate has any harm cell^[s7]. The median \(\delta\) runs from \(-10.1\) pp on HO-M31 MOS1 to \(-2.9\) / \(-3.5\) pp on HO-mixed (Figure 1a). For scale: on HO-blank MOS1 the archived 2.filter has \(\sigma_R=8.58\%\), the candidates 3.87–11.72%, and no filtering at all 47.3%^[s8].

这个增益大多是拼接层面的。按预先采纳的逐观测读数，79 个 Confirmed 里只有 6 个"宽"（在每个增益格里至少三分之二的观测改善），73 个"集中"（只改善约一半到三分之二的观测）；移到新天空的证据是 44 个候选、14 个族^[s9]。

The gain is mostly a mosaic-level effect. Under the per-observation reading adopted in advance, only 6 of the 79 Confirmed are "broad" (at least two thirds of the observations improve in every gaining cell) and 73 are "concentrated" (improving about half to two thirds of the observations); the evidence for transfer to new sky is 44 candidates in 14 families^[s9].

3.2 参照有多弱：204 个非候选里也有 50 个通过3.2 How weak the reference is: 50 of 204 non-candidates also pass

79/83 听起来像一次完胜，所以我们接着问：这条规则本身有多容易通过？把同一判决规则用到 204 个没有被选为候选的策略上（包括从 2.filter 派生的复现滤波器、角落参照和更早的搜索代），其中 50 个也通过（23 个在全部 6 个格通过）^[s10]。也就是说，"降低相对归档 2.filter 的散布"这件事，一个简单的 espfilt 加硬段裁剪也能做到它的一半以上：EH_B1000_n2 单独就能拿到最优候选增益的中位 56%（范围 41–64%）^[s11]。所以 79/83 不能读成"我们的方法比一个标准滤波器更好"。

79/83 sounds like a clean win, so we next asked: how easy is this rule to pass? Applying the same verdict rule to 204 policies that were not selected as candidates (including 2.filter-derived reproduced filters, corner references and earlier search generations), 50 of them also pass (23 in all 6 cells)^[s10]. That is, "lower the scatter relative to the archived 2.filter" is something a simple espfilt plus a hard-band clip already achieves more than half of: EH_B1000_n2 alone delivers a median 56% (range 41–64%) of the best candidate's gain per cell^[s11]. So 79/83 cannot be read as "our method beats a standard filter".

3.3 对复现滤波器：11 of 83，全部保留更少时间3.3 Against the reproduced filter: 11 of 83, all keeping less time

于是我们把参照换成那个从同一批光变曲线复现的标准滤波器（EH_B1000_n2：espfilt 之后加一个 1000 s 硬段裁剪）。同一判决规则下只剩 11 个候选（6 个族）Confirmed，37 个 Ambiguous，35 个 Not confirmed^[s12]。这 11 个的保留时间在每个格都严格更短：0.32–0.76 对参照的 0.81–0.93（图 1c）。它们不是靠"同样的时间删得更准"赢的，而是靠"删得更多"。

We therefore swapped the reference for the standard light-curve filter reproduced from the same curves (EH_B1000_n2: espfilt followed by a 1000 s hard-band clip). Under the same verdict rule only 11 candidates (6 families) remain Confirmed, 37 Ambiguous, 35 Not confirmed^[s12]. All 11 keep strictly less time in every cell: 0.32–0.76 against the reference's 0.81–0.93 (Figure 1c). They do not win by "cutting the same time more accurately" but by "cutting more".

3.4 等保留时间：0–1 of 83，优势未确立3.4 At matched kept time: 0–1 of 83, the advantage is not established

最后我们问：如果把"删得更多"这个差别去掉，优势还在吗？我们把每个候选与"保留至少同样多时间的最好复现滤波器设置"比较（一个 48 设置的阶梯，每次抽样重画），用逐观测配对自举。点估计下 37/83 领先，但配对自举下只剩 0–1 个 Confirmed（\(\delta_0=0.5\) pp；在 \(\delta_0=0\) 是 16 个）^[s13]。唯一的幸存者贴着阈值（p84 = \(-0.502\) pp），换一个种子就消失，去掉单个观测 0202230301 也降到 0。所以诚实的结论是：在等保留时间下，候选相对最好复现滤波器的优势未确立 —— 候选可能真有优势，但现有数据不足以确立它。

Finally we asked: remove the "cuts more" difference — does the advantage survive? We compared each candidate with the best reproduced-filter setting that keeps at least as much time (a 48-setting staircase, redrawn in every draw), with a per-observation paired bootstrap. By point estimate 37 of 83 lead, but under the bootstrap only 0–1 is Confirmed (\(\delta_0=0.5\) pp; 16 at \(\delta_0=0\))^[s13]. The single survivor sits on the threshold (p84 = \(-0.502\) pp), vanishes under another seed, and goes to 0 without the single observation 0202230301. The honest conclusion is therefore: at matched kept time, the candidates' advantage over the best reproduced filter is not established — they may well be better, but the data so far do not establish it.

3.5 两个必须同报的限定3.5 Two qualifiers that must travel with the result

两个限定必须与判决同报。第一，HO-M31 MOS1 上最大的表面增益（\(-10.1\) pp）里，约 4.6 pp 来自一个损坏的归档 2.filter 文件（观测 0202230301 的归档 GTI 退化了：MOS1 保留了包括一次耀斑的全部 24.1 ks，而 espfilt 复现只保留 21.8 ks），约 5.5 pp 是候选在更难环境下的自身性质（去一观测法 \(-3.2\) 到 \(-6.1\) pp）^[s14]。第二，Al-K 荧光线门 —— 防止"挑噪声伪装成删耀斑"的保险 —— 在保留集上每个格只在 0–3 个观测上触发，所以本次判决没有检验它；两个线门候选的确认实际检验的是它们的基底政策^[s15]。

Two qualifiers must travel with the verdict. First, of the largest apparent gain on HO-M31 MOS1 (\(-10.1\) pp), about 4.6 pp come from one broken archived 2.filter file (the archived GTI of observation 0202230301 is degenerate: MOS1 keeps the whole 24.1 ks including a flare, while the espfilt reproduction keeps 21.8 ks), and about 5.5 pp are a candidate property in a harder regime (leave-one-out \(-3.2\) to \(-6.1\) pp)^[s14]. Second, the Al-K fluorescence-line gate — the insurance against "selecting noise masquerading as cutting flares" — fires on only 0–3 observations per cell on held-out data, so this verdict does not test it; the two line-gate candidates are confirmed on their base policies^[s15].

4. 讨论4. Discussion

4.1 回到问题4.1 Back to the question

我们问的是：一个只看光变曲线的筛选规则，在从未参与它设计的观测上，比标准做法好多少？答案分三层。相对归档的 2.filter，83 个候选里 79 个降低了观测间散布 —— 这是本次判决确立的。相对一个复现的标准光变曲线滤波器，只剩 11 个，且全部靠删更多时间。在等保留时间下，配对自举后优势未确立。所以对弥散发射工作者的实际建议是：单靠"比归档 GTI 更好"不足以换掉标准流程；换掉它的理由必须建立在"等保留时间下也更好"上，而这正是下一个测试预先注册的问题。本节小结：判决是真的，但它证明的是"相对归档筛选器的改进"，不是"相对标准方法的优势"。

We asked: by how much does a light-curve-only screening rule beat the standard practice, on observations that took no part in its design? The answer has three layers. Relative to the archived 2.filter, 79 of 83 candidates lower the between-observation scatter — that is what this verdict establishes. Relative to a reproduced standard light-curve filter, only 11 survive, and all by cutting more time. At matched kept time, the advantage is not established under a paired bootstrap. The practical advice for diffuse-emission work is therefore: "better than the archived GTI" alone does not justify replacing the standard pipeline; the justification must rest on "better at matched kept time", which is exactly the question the next test pre-registers. This subsection's point: the verdict is real, but what it establishes is an improvement relative to the archived filter, not an advantage over the standard method.

4.2 与文献的比较4.2 Comparison with the literature

文献里没有已发表的"完全不用角落的跨观测静息水平估计器"（R4 文献检查，三臂复核）^[s16]，所以本项目的政策族是新的地面。已有方法给出的是系统误差的锚点而非判据：严格双带监控后 Nevalainen 等人报告 0.8–1 keV 处 20%、4–7 keV 处 5%（90% 置信）的背景模型不确定度^[s17]；ESAS 的粒子背景建模给出打分带之外的系统误差锚点^[s18]。我们测到的 2.filter 下 17–25% 的 \(s_{\rm sys}\)（0.5–2 keV 拼接光变曲线）与这些锚点相容，且大得多 —— 这正是"观测间一致性优先"这个判据的根据。本节小结：文献给出锚点，本文给出判据与第一次保留集检验。

No published "fully corner-free cross-observation quiescent-level estimator" exists (the R4 literature check, three arms)^[s16], so this project's policy family is new ground. What existing methods give are anchors for the systematic, not a criterion: after strict two-band monitoring Nevalainen et al. report background-model uncertainties of 20% at 0.8–1 keV and 5% at 4–7 keV (90% confidence)^[s17]; ESAS particle-background modelling gives anchors for systematics outside the score band^[s18]. Our measured \(s_{\rm sys}\) of 17–25% under 2.filter (0.5–2 keV concatenated light curves) is consistent with those anchors and much larger — which is exactly the basis for the "consistency between observations first" criterion. This subsection's point: the literature gives anchors; this paper gives a criterion and the first held-out test of it.

4.3 局限，定量地4.3 Limitations, quantitatively

四条局限，每条带数字。(i) 参照强度：主判决相对的是归档 2.filter，204 个非候选里也有 50 个通过同一条规则（§3.2）。(ii) 保留时间：判决不比较保留时间（PI 决定把它放在判决之外）；11 个幸存者保留 0.32–0.76 对 0.81–0.93，等保留时间下的读数功效在下一个测试的尺寸下接近零，所以那里它是描述性的。(iii) 增益的构成：79 个 Confirmed 里 73 个是"集中"的；HO-M31 MOS1 的 \(-10.1\) pp 里约 4.6 pp 是一个坏参照文件；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。(iv) 未检验的部分：Al-K 线门每格只在 0–3 个观测上触发（本次判决没有检验它）；谱验证（把快速判据落到光谱一致性）还没做；\(\sigma_R\) 是观测内的拼接风险，不含观测间的水平项。本节小结：判决的每一层都有明确的边界，边界都写在结果里。

Four limitations, each with its number. (i) Reference strength: the primary verdict is relative to the archived 2.filter, which 50 of 204 non-candidates also pass (§3.2). (ii) Kept time: the verdict does not compare kept time (the PI set it outside the verdict); the 11 survivors keep 0.32–0.76 vs 0.81–0.93, and the matched-kept-time reading has near-zero power at the next test's size, so there it is descriptive. (iii) Composition of the gain: 73 of the 79 Confirmed are "concentrated"; 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. (iv) What is untested: the Al-K line gate fires on 0–3 observations per cell (this verdict does not test it); spectral validation (landing the quick criterion on spectral consistency) has not been done; \(\sigma_R\) is the within-observation mosaic risk and carries no between-observation level term. This subsection's point: every layer of the verdict has an explicit boundary, and the boundaries are written into the results.

4.4 对同行的建议与下一步4.4 Advice for others, and what follows

对做弥散发射的人，三条可操作的建议。第一，报"相对归档 GTI 的改进"时，把复现滤波器也报出来 —— 一个 espfilt 加硬段裁剪就能拿到最优候选增益的一半以上，不报它会让改进显得比实际大。第二，报保留时间：删多删少是这个问题的核心坐标，只报散布不报时间会误导。第三，把判决规则冻结在看到分数之前，并把每一步的读数收窄如实写出 —— 本次判决的三次收窄全部来自冻结后的红队检查，它们没有推翻判决，但把"判决证明了什么"的边界画清楚了。下一步是预先注册的 P007：主估计量是六个格的等权平均 \(\bar D\) 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^[s19]。本节小结：方法与判据都可以被直接借用，借用时必须带上等保留时间这一条。

For those doing diffuse-emission work, three actionable recommendations. First, when reporting "improvement over the archived GTI", also report the reproduced filter — a simple espfilt plus a hard-band clip already delivers more than half of the best candidate's gain, and omitting it makes the improvement look larger than it is. Second, report kept time: cut-more-or-cut-less is a core coordinate of this problem, and reporting scatter without time misleads. Third, freeze the verdict rule before any score is seen, and write each narrowing of the reading honestly — all three narrowings of this verdict came from post-freeze red-team checks; they did not overturn it, but they drew the boundary of what it establishes. What follows is the pre-registered P007: the primary estimand is the equal-weight six-cell mean \(\bar D\) against the reproduced filter (\(\delta_0=0.5\) pp, harm guard, all cells valid), size \(\leq 3\%\), with the verdict code written and self-checked on development data^[s19]. This subsection's point: the method and the criterion can be borrowed directly, and the matched-kept-time condition must travel with them.

5. 小结5. Summary

我们用 119 个从未参与设计的 XMM-Newton EPIC 观测，对 83 个只用光变曲线构造的 GTI 筛选政策做了第一次冻结的保留集判决，判据是观测间散布 \(\sigma_R\)。主要结果：

We ran the first frozen held-out verdict on 83 GTI screening policies built from light curves alone, using 119 XMM-Newton EPIC observations that took no part in their design, with the between-observation scatter \(\sigma_R\) as the criterion. The main results:

1. 79 of 83 相对归档 2.filter Confirmed（23 个族，无伤害格），中位改善 \(2.9\text{–}10.1\) pp；增益大多是拼接层面的（73 of 79 "集中"；图 1a）。79 of 83 Confirmed against the archived 2.filter (23 families, no harm cell), median improvement \(2.9\text{–}10.1\) pp; the gain is mostly mosaic-level (73 of 79 "concentrated"; Figure 1a).
2. 参照很弱：204 个非候选政策里也有 50 个通过同一条规则；相对复现的标准滤波器只剩 11 of 83，全部保留更少时间（0.32–0.76 对 0.81–0.93；图 1b–c）。The reference is weak: 50 of 204 non-candidate policies also pass the same rule; against the reproduced standard filter only 11 of 83 survive, all keeping less time (0.32–0.76 vs 0.81–0.93; Figure 1b–c).
3. 等保留时间下配对自举只剩 0–1 of 83 —— 优势未确立；这是下一个测试预先注册要回答的问题。At matched kept time a paired bootstrap leaves 0–1 of 83 — the advantage is not established; that is the question the next test pre-registers.

判据、规则与判决代码全部冻结并哈希；每一次读数收窄都来自冻结后的独立复核并如实报告。光变曲线筛选是一个值得继续的路线，而"等保留时间下也更好"是它必须跨过的下一道门槛。

The criterion, the rule and the verdict code were all frozen and hashed; every narrowing of the reading came from independent post-freeze checks and is reported as such. Light-curve screening is a road worth continuing, and "better at matched kept time" is the threshold it must cross next.

致谢、设施与软件、数据可用性、附录、参考文献、来源对照表Acknowledgments, facilities and software, data availability, appendices, references, source map

致谢Acknowledgments

本工作基于 XMM-Newton（ESA 科学卫星，其仪器与贡献由 ESA 成员国与 NASA 直接资助）获得的观测数据。感谢 XXX（资助占位符，由 PI 填写）。本手稿的编写与代码核验使用了 AI 助手（语言模型辅助的文本起草与自动核验脚本），全部科学判断与最终文本由作者负责。

Based on observations obtained with XMM-Newton, an ESA science mission with instruments and contributions directly funded by ESA Member States and NASA. We thank XXX (funding placeholder, for the PI to fill in). AI assistants (large-language-model help with text drafting and automated verification scripts) were used in preparing this manuscript and checking its code; all scientific judgement and the final text are the authors'.

设施与软件Facilities and software

设施：XMM(Newton)。软件：SAS 与 ESAS（Kuntz & Snowden 2008 及其引用的 SAS/ESAS 正式文献）；astropy、numpy、scipy、matplotlib 与本项目自有的分析脚本（仓库内，见数据可用性）。

Facilities: XMM(Newton). Software: SAS and ESAS (Kuntz & Snowden 2008 and the canonical SAS/ESAS papers cited therein); astropy, numpy, scipy, matplotlib, and this project's own analysis scripts (in the repository; see Data availability).

数据可用性Data availability

原始数据在 XMM-Newton 科学档案（ObsID 清单见 §2.1 与记录 [s3]）。产品与代码的发布位置是占位符（由 PI 决定）。

The raw data are in the XMM-Newton Science Archive (ObsID lists in §2.1 and record [s3]). The release location of the products and code is a placeholder (for the PI to decide).

附录 A：打分器与判决规则（技术细节）Appendix A: the scorer and the verdict rule (technical detail)

\(\sigma_R\) 的定义：一个格内观测间 0.5–2 keV 水平的散布，由逐观测配对自举（2000 次抽样，种子 20261015）给出 p16/p84。打分分段 1000 s；分段长度的敏感性（500/2000 s、300 s、无长度规则）作为次要读数，领先者的符号在全部规则下不变。判决：一个候选 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、无伤害格；Ambiguous / Not confirmed 按冻结文本。83 个候选、23 个脚本与 11 条命令在 2026-10-18T15:33Z 冻结（清单 sha 8598b555…，运行版 v2 sha c7aad3f8…）；此后只有"可执行回归失败"能改代码，且每次改动都带失败回归测试。冻结链运行一次因保留集观测 0302351401 的空天空坐标在命令 0 停止；修复（丢弃空天空事件）带 FAIL \(\to\) PASS 回归，发展数据逐位不变。

Definition of \(\sigma_R\): the scatter of the 0.5–2 keV level between observations within a cell, with p16/p84 from a per-observation paired bootstrap (2000 draws, seed 20261015). Scoring used 1000 s pieces; the piece-length sensitivity (500/2000 s, 300 s, a length-free rule) is a secondary reading, and the leaders' sign is unchanged under every rule. Verdict: a candidate is Confirmed if and only if at least 4 cells gain (p84(\(\delta\)) \(<-0.5\) pp), both new-sky cells gain, and no cell is harmed; Ambiguous / Not confirmed follow the frozen text. The 83 candidates, 23 scripts and 11 commands were frozen at 2026-10-18T15:33Z (manifest sha 8598b555…, run version v2 sha c7aad3f8…); after that only a failing executable regression may change code, and every change carries a failing regression test. One frozen-chain run stopped at command 0 on null sky coordinates in held-out observation 0302351401; the fix (dropping null-sky events) carries a FAIL \(\to\) PASS regression and leaves development data bit-identical.

参考文献（经 ADS/Crossref 核实的真实文献）References (real, verified against ADS/Crossref)

1. De Luca, A., & Molendi, S. 2004, A&A, 419, 837（doi:10.1051/0004-6361:20034421）De Luca, A., & Molendi, S. 2004, A&A, 419, 837 (doi:10.1051/0004-6361:20034421)
2. Kuntz, K. D., & Snowden, S. L. 2008, A&A, 478, 575（doi:10.1051/0004-6361:20077912）Kuntz, K. D., & Snowden, S. L. 2008, A&A, 478, 575 (doi:10.1051/0004-6361:20077912)
3. Carter, J. A., & Sembay, S. 2008, A&A, 489, 837（doi:10.1051/0004-6361:200809997）Carter, J. A., & Sembay, S. 2008, A&A, 489, 837 (doi:10.1051/0004-6361:200809997)
4. Nevalainen, J., Markevitch, M., & Lumb, D. 2005, ApJ, 629, 172（arXiv:astro-ph/0504362; doi:10.1086/431198）Nevalainen, J., Markevitch, M., & Lumb, D. 2005, ApJ, 629, 172 (arXiv:astro-ph/0504362; doi:10.1086/431198)
5. Ghizzardi, S., Marelli, M., Salvetti, D., & Gastaldello, F. 2017, Exp. Astron., 44, 273（doi:10.1007/s10686-017-9554-1）Ghizzardi, S., Marelli, M., Salvetti, D., & Gastaldello, F. 2017, Exp. Astron., 44, 273 (doi:10.1007/s10686-017-9554-1)

来源对照表（本手册专用附录，不属于投稿件）Source map (a handbook-only appendix, not part of the submission)