前沿策略对比（GTI 政策的图文展示）Frontier policies (the GTI policies, shown)

1. 一段话回答1. The one-paragraph answer

答：在 31 个 M31 halo 观测、2 台 MOS 相机上，与归档 2.filter 相比，前沿政策确实降低了"每观测总不确定度 σ_o"：以 gen2:hermes.FLR_z3 为例，25 个测试观测（没有任何设计者看过）的叠加 mosaic 期望风险 σ_R 从 5.15% 降到 3.81%（MOS1，配对差 −1.34 pp，自举离散 [−1.98, −0.65]）与从 5.84% 降到 3.93%（MOS2，−1.92 [−2.58, −1.18]）；代价是保留 0.85–0.87 的活时间。3 个政策（FLR_z3、MS_z3.5_core、RT3_h2.5_l1）在 4 个集 × 2 台相机共 8 种组合上都优于 2.filter（R9，离散均不含 0）。但这个改进主要是修好 2.filter 失效的观测（RT-16 F-01）：在 M31 MOS1 上，去掉 2 个最差的 2.filter 观测后 FLR_z3 的增益从 −1.77 降到 −0.79 pp，MS_z3.5_core 变成 +0.33 pp（持平）。观测间剖面水平偏移（RT-12 F-02 的池化统计）在 MOS2 上 FLR_z3 降低 −2.0 pp（95% [−3.3, −0.7]），Q_d0.05_n2 在两台相机都降（MOS1 −3.7 [−5.9, −0.1]，MOS2 −2.3 [−3.3, −0.9]）——但该统计量是事后选择的，待预注册（R7）。竖直剖面的中位 τ_b 没有显著改善：观测到的 −0.5 到 −0.8 pp 落在 0 与预测（−0.8…−1.9 pp）的 1σ 之内，检验功效只有 1–2σ（RT-12 F-01）。而"覆盖代价"这一分数本身已经不可靠：RT-18 发现 σ_RC 给每个被丢指向收同一个中位 V_miss，而实际被丢指向的插值残差大一个量级，且 V_miss 混入了 2.filter 自己的失效；结论（GATE 掉到 2.filter 之后）随 V_miss 估计量而翻转。因此现在只用 σ_R 排名，同时显式报告每个政策丢掉了哪些观测和什么类型的目标（E022/R13），直到有一个干净参考的覆盖价格。主候选正在被重新询问：FLR_z3 的 M31/LH 优势来自它的绝对硬度门（r0 = 1.8），而同一个门在 Virgo-server 集上丢掉 51/142（MOS1，2.filter 时间的 34%）与 54/142（MOS2，38%）个观测——亮源或硬目标场的天空比 1.8 更硬（E022）。

Answer: over the 31 M31-halo observations and 2 MOS cameras the frontier policies do lower the per-observation total uncertainty σ_o relative to the archived 2.filter: taking gen2:hermes.FLR_z3, on the 25 test observations (no designer looked at them) the stacked mosaic expected risk σ_R falls from 5.15% to 3.81% (MOS1; paired difference −1.34 pp, bootstrap spread [−1.98, −0.65]) and from 5.84% to 3.93% (MOS2; −1.92 [−2.58, −1.18]), for 0.85–0.87 of the live time. 3 policies (FLR_z3, MS_z3.5_core, RT3_h2.5_l1) beat 2.filter in all 8 set × camera cases (4 sets × 2 cameras; R9; every spread excludes 0). But the win is mostly the repair of 2.filter's failed observations (RT-16 F-01): on M31 MOS1, removing the 2 worst 2.filter observations takes FLR_z3's gain from −1.77 to −0.79 pp and turns MS_z3.5_core into +0.33 pp (no significant difference). On the pooled profile-level offset statistic (RT-12 F-02), FLR_z3 lowers MOS2 by −2.0 pp (95% [−3.3, −0.7]) and Q_d0.05_n2 lowers both cameras (MOS1 −3.7 [−5.9, −0.1], MOS2 −2.3 [−3.3, −0.9]) — but that statistic was chosen after the fact and awaits pre-registration (R7). The profile median τ_b shows no significant improvement: the observed −0.5 to −0.8 pp lies within 1σ of both 0 and the prediction (−0.8…−1.9 pp); the test has only 1–2σ power (RT-12 F-01). And the coverage-cost score itself is not reliable: RT-18 found that σ_RC charges every dropped pointing the same median V_miss, while the interpolation residuals of the pointings actually dropped are an order of magnitude larger, and V_miss mixes in 2.filter's own failures; the conclusion (GATE falling below 2.filter) flips with the V_miss estimator. So the ranking now uses σ_R only, with an explicit report of which observations and which target types each policy drops (E022/R13), until a clean-reference coverage price exists. The main candidate is being re-asked: FLR_z3's M31/LH edge comes from its absolute hardness gate (r0 = 1.8), and the same gate drops 51/142 observations (MOS1, 34% of 2.filter's time) and 54/142 (MOS2, 38%) on the Virgo-server set — the sky of bright or hard targets is harder than 1.8 (E022).

2. 展示的政策（十个）2. The policies on show (ten)

本页每张图都用同一套颜色：2.filter 永远黑色；用角落的政策（Q_d0.05_n2）画空心点。T/2.filter 是全部 31 个 M31 halo 观测叠加后的保留活时间比（MOS1 / MOS2）。所有政策只用光变曲线（PI 规则 D013），不依赖 2.filter 构建。2026-10-16 新增 merged:sol.MS_z3.5_core：R9 判定它是 3 个跨集稳健政策里的最佳折衷（在每个集都排前 7–17，保留 0.74–0.88 的时间），且 E022 发现它每个集最多只丢 1 个观测——这正是 σ_RC 不可靠后（RT-18）必须显式报告的那个量。它的详细说明在 policy_ms.html。

Every figure on this page uses one colour set: 2.filter is always black; the corner-using policy (Q_d0.05_n2) is drawn hollow. T/2.filter is the stacked kept-live-time ratio over all 31 M31-halo observations (MOS1 / MOS2). All policies use light curves only (the PI's rule, D013) and are built independently of 2.filter. New on 2026-10-16: merged:sol.MS_z3.5_core — R9 judged it the best compromise among the 3 cross-set robust policies (top 7–17 everywhere, keeping 0.74–0.88 of the time), and E022 found it drops at most 1 observation per set, which is exactly the quantity that must now be reported explicitly once σ_RC is unreliable (RT-18). Its full walk-through is on policy_ms.html.

3. 光变曲线与保留时间3. Light curves and kept time

每个基本案例（D015 的角色划分）一张图：上排是 300 s bin 的 0.5–2 keV（30″ 源掩蔽，打分带）与 2.5–8 keV FOV（选择带）率，MOS1 与 MOS2；下排是各政策每个 300 s bin 的保留比例（色条与第 2 节同色）。要看的是：2.filter 留下了什么（红/蓝曲线在黑条之上仍升高的时段），以及领先政策是否把它删掉。每张图的说明里给出该观测、该相机的保留时间 K 与每观测总不确定度 σ_o（来源列）。

One figure per basic case (the roles fixed by D015): the top row is the 0.5–2 keV (30″ source-masked, the score band) and 2.5–8 keV FOV (the selection band) rates in 300 s bins, MOS1 and MOS2; the bottom row is each policy's kept fraction per 300 s bin (same colours as section 2). What to look for: what 2.filter leaves behind (time where the curves stay high above the black strip) and whether the leaders cut it. Each caption quotes the kept time K and the per-observation total uncertainty σ_o for that observation and camera (source column).

4. 单次观测的帕累托前沿4. Pareto front for one observation

两次观测 × 两台相机：横轴是保留活时间 K / K(2.filter)，纵轴是每观测总不确定度 σ_o = √(τ² + 1/N)（v3.1 收据 per_obs.json）。灰色小点是池里其余政策；绿色阶梯是非支配前沿；九个展示政策用第 2 节的彩色圆点（用角落的 Q 为空心）；每个彩色政策还画一个同色"×"表示扣掉信息地板后的 σ_o_net = √(τ²_net + 1/N)；箭头指向 σ_o 最小的点。

Two observations × two cameras: x is the kept live time K / K(2.filter), y is the per-observation total uncertainty σ_o = √(τ² + 1/N) (the v3.1 receipt per_obs.json). Small grey dots are the other policies of the pool; the green staircase is the non-dominated front; the nine shown policies are the coloured circles of section 2 (Q, the corner user, hollow); each coloured policy also carries a same-colour "×" for the floor-subtracted σ_o_net = √(τ²_net + 1/N); the arrow marks the minimum-σ_o point.

5. 叠加全部观测的帕累托前沿（马赛克）5. Pareto front for all observations stacked (the mosaic)

把 31 个 M31 halo 观测当作一个拼接观测（E019 的 σ_R 判据）：横轴是 T/2.filter（叠加保留活时间比），纵轴是马赛克期望风险。三列分别是三个分数：σ_R（报告的）、σ_R_net（扣掉信息地板）、σ_RC（加上覆盖代价 V_miss）。但 σ_RC 已在 2026-10-16 被 RT-18 降级：它给每个被丢指向收同一个中位 V_miss，而实际被丢指向的插值残差大一个量级（FLR_z3 丢的 1701/1801/3101 均值 0.044，中位只有 0.0045），且 V_miss 混入了 2.filter 自己的失效；换一个估计量，结论就翻转。所以σ_R 是唯一用于排名的分数，σ_RC 一列仅作为"曾经怎么算"的记录保留，被丢观测数另列报告（E022/R13）。主图用 test 子集（25 个观测，任何设计者都没看过）；all 31 的数字在下拉表里给出。灰色小点是池里其余政策；绿色阶梯是前沿；箭头指向每列的最小值。

Treating the 31 M31-halo observations as one concatenated observation (E019's σ_R criterion): x is T/2.filter (the stacked kept-live-time ratio), y is the mosaic expected risk. The three columns are the three scores: σ_R (as reported), σ_R_net (floor-subtracted) and σ_RC (with the coverage cost V_miss). But σ_RC was demoted on 2026-10-16 by RT-18: it charges every dropped pointing the same median V_miss, while the interpolation residuals of the pointings actually dropped are an order of magnitude larger (FLR_z3's dropped 1701/1801/3101 average 0.044 against a median of 0.0045), and V_miss mixes in 2.filter's own failures; a different estimator flips the conclusion. So σ_R is the only score used for ranking; the σ_RC column is kept only as a record of how it was once computed, and the number of dropped observations is reported separately (E022/R13). The main figure uses the test subset (25 observations no designer looked at); the all-31 numbers are in the dropdown table below. Small grey dots are the other policies; the green staircase is the front; the arrow marks each column's minimum.

"某政策在前沿的什么位置？""Where does policy X sit?"

选一个政策（与相机）：上表给出它在 210 个政策里的排名、到前沿的距离（同 T 处前沿分数减去它的分数，负值 = 在前沿上/更优）、对 2.filter 的配对差与自举离散 p16–p84，以及 T/2.filter。数字来自 risk.json 的 test 子集（25 个观测）。同时下面两张前沿图会换成该政策被高亮的版本（更大的点、粗体标签、箭头）：单观测前沿（第 4 节）与叠加前沿（第 5 节）。相机选择只改表格与叠加图——单观测图固定画 MOS1 与 MOS2 两列。

Pick a policy (and camera): the table gives its rank among the 210 policies, its distance to the front (the front's score at the same T minus its own; negative = on/below the front), the paired difference against 2.filter with its bootstrap spread p16–p84, and T/2.filter. The numbers come from the test subset (25 observations) of risk.json. The two front figures below also switch to a version with that policy highlighted (bigger marker, bold label, arrow): the per-observation front (section 4) and the stacked front (section 5). The camera choice changes only the table and the stacked figure — the per-observation figure always draws both MOS1 and MOS2.

选一个政策：Pick a policy: 相机camera MOS1 MOS2

排名按该分数在 210 个政策中的位置；距离为 0 表示该政策就在前沿上；负值表示它比同 T 处的前沿更优（前沿阶梯未覆盖该点之外）。σ_RC 一列按 RT-18 降级处理，不作排名。

The rank is the policy's position among the 210 policies under that score; a distance of 0 means it lies on the front; a negative value means it is better than the front at the same T (the staircase does not extend beyond that point). The σ_RC column is demoted per RT-18 and is not a ranking.

6. 观测之间的一致性6. Consistency between observations

第一张图是配对哑铃图：每个观测一行，黑点 = 2.filter 的 σ_o，蓝点 = FLR_z3 的 σ_o（同一观测、同一相机、同一收据）。它回答"哪些观测赢了"——每一条蓝点在黑点左边的行就是 σ_o 更低的那一行。

The first figure is a paired dumbbell plot: one row per observation, black = 2.filter's σ_o, blue = FLR_z3's σ_o (same observation, same camera, same receipt). It answers "which observations gain" — every row where the blue dot sits left of the black one is a row where σ_o is lower.

池化的每观测水平偏移（RT-12 F-02）。这是比中位 τ_b 更有功效的统计量：把每个观测在同一 z 仓的（泊松误差 < 25% 的）亮度偏差取平均，得到每观测偏移；其跨观测 rms（扣除平均泊松噪声）就是"观测间散布"。2.filter 下它是 10.5%（MOS1）与 7.9%（MOS2）。配对自举差（1000 次重采样，95% 区间）：

The pooled per-observation offset (RT-12 F-02). This is a more powerful statistic than the median τ_b: for each observation, average its fractional deviations from the same-z mean over its well-measured bins (Poisson error < 25%); the rms of these per-observation offsets across observations, corrected for the mean Poisson noise, is the between-observation scatter. Under 2.filter it is 10.5% (MOS1) and 7.9% (MOS2). Paired bootstrap differences (1000 resamples, 95% intervals):

Lockman Hole：一个观测承担的观测间项The Lockman Hole: one observation carrying the between term

在 Lockman Hole（13 个观测）上，2.filter 的观测间项本来就是 14.6%（MOS1）/ 23.3%（MOS2）（角落修正后；CDFS 同法只有 0.0%）。RT-06 找到原因：0022740101 一个观测（DATE-OBS 2001-10-25）承担了几乎全部——espfilt 的直方图峰锁在它的耀斑上（μ_H = 3.0 / 2.6 c/s，集合中位数的 6.8× / 6.0×），2.filter 在这个水平上保留 13.5 ks / 7.4 ks，保留的软率是中位的 1.8× / 2.8×。单独删掉它，between 降到 9.3% / 8.9%；其余每个观测的刀切检验都留在 22.6–25.5%（MOS2）。角落在 LH 上几乎不起作用（只移走 0.3 / 0.0 点，CDFS 上是 11.4 / 13.5 点），因为粒子超额在 FOV 内、角落看不到。这一失效类（espfilt 把静息峰锁到耀斑上）正是运行中的 R2（跨观测静息参考）的目标。

On the Lockman Hole (13 observations), the between term under 2.filter is already 14.6% (MOS1) / 23.3% (MOS2) (corner-corrected; the same pipeline gives 0.0% on CDFS). RT-06 found the cause: the single observation 0022740101 (DATE-OBS 2001-10-25) carries almost all of it — espfilt's histogram peak locks onto its flare (μ_H = 3.0 / 2.6 c/s, 6.8× / 6.0× the set median), 2.filter keeps 13.5 ks / 7.4 ks at that level, and the kept soft rate is 1.8× / 2.8× the median. Dropping it alone takes the between term to 9.3% / 8.9%; every other observation's jackknife leaves 22.6–25.5% (MOS2). The corner does almost nothing on the LH (it removes 0.3 / 0.0 points against 11.4 / 13.5 on CDFS) because the particle excess sits inside the FOV, where the corner cannot see it. This failure class — espfilt locking its quiescent peak onto a flare — is what the running R2 (a cross-observation quiescent reference) targets.

7. 竖直剖面（M31 短轴）7. The vertical profile (M31 minor axis)

2.filter 与三个领先政策（FLR_z3、JFR_z3.5、MG_CP_k0.5）的 0.5–2 keV 竖直剖面（4′ 仓，对数亮度；左 MOS1、右 MOS2；sol run2，CCD 足迹修正后的产品）。四条曲线几乎重合——这本身就是结果：剖面的形状对政策不敏感，一致性问题出在"同一 z 仓内不同观测的散布"上，而不是形状上。

The 0.5–2 keV vertical profiles of 2.filter and three leaders (FLR_z3, JFR_z3.5, MG_CP_k0.5; 4′ bins, log brightness; MOS1 left, MOS2 right; the sol run2, CCD-footprint-corrected products). The four curves nearly coincide — that is itself the result: the profile's shape is insensitive to the policy; the consistency problem lives in the scatter of different observations within the same z bin, not in the shape.

每观测偏移的图在第 6 节（fig_front_offsets.png）。近盘仓（|z| < 60′）的 τ_b 比远盘仓多约 3 pp（RT-12 F-06：MOS1 中位 11.6% 对 8.7%，MOS2 13.6% 对 10.7%）——R7 的预注册将限制在远盘仓上。

The per-observation-offset figure is in section 6 (fig_front_offsets.png). The near-disk bins (|z| < 60′) carry about 3 pp more τ_b than the far bins (RT-12 F-06: MOS1 median 11.6% vs 8.7%, MOS2 13.6% vs 10.7%) — R7's pre-registration will restrict to the far-disk bins.

8. 还没有展示的（诚实清单）8. What is not yet shown (the honest list)

• 覆盖代价的价格还没有定论，σ_RC 已被降级。RT-18（2026-10-16）发现 σ_RC 给每个被丢指向收同一个中位 V_miss（MOS1 0.0045），而实际被丢指向的插值残差大一个量级（FLR_z3 丢的 1701/1801/3101 均值 0.044），且 V_miss 混入了 2.filter 自己的失效（9/31 个残差的邻居是 2.filter 失效观测）；估计量的自举离散 5.8–11.2%，蒙特卡洛偏差 −30% / +40%。一个正确的价格需要被丢指向的干净参考水平，还没有。所以本页只用 σ_R 排名，另列报告每个政策丢掉的观测数（E022/R13）。 The coverage price is not settled, and σ_RC has been demoted. RT-18 (2026-10-16) found that σ_RC charges every dropped pointing the same median V_miss (MOS1 0.0045), while the interpolation residuals of the pointings actually dropped are an order of magnitude larger (FLR_z3's dropped 1701/1801/3101 average 0.044), and V_miss mixes in 2.filter's own failures (9 of 31 residuals have a 2.filter-failed neighbour); the estimator's bootstrap spread is 5.8–11.2% with a −30% / +40% Monte-Carlo bias. A correct price needs a clean reference level for the dropped pointing, which does not exist yet. So this page ranks on σ_R only and reports the dropped-observation count separately (E022/R13).
• 主候选 FLR_z3 正在被重新询问（E022，2026-10-16）。它的 M31/LH 优势来自绝对硬度门（静息 h/u 与 M31 halo 常数 r0 = 1.8 比较）：关掉硬度部分后（FLRr_z3，只留率偏移），M31 MOS1 变成 −0.67 pp [−1.64, +0.35] 持平，LH 从 −5.53 降到 −2.91。同一个门在 Virgo-server 集上丢掉 51/142（MOS1，2.filter 时间的 34%）与 54/142（MOS2，38%）个观测——M31 Core、NGC 1365、M32、NGC 1399 的热气团：亮源或硬目标场的天空比 1.8 更硬，一个弥漫发射分析会因此丢掉全部数据。下一步 R13：预注册 MS_z3.5_core 与 FLRr_z3，各配 (a) 场参照拒绝门 XJ_g1.5（E021）、(b) R11 线门、(c) 两者，在四个集上打分，由 PI 在 FLR_z3 与加了门的 MS 之间决定。 The main candidate FLR_z3 is being re-asked (E022, 2026-10-16). Its M31/LH edge comes from the absolute hardness gate (the quiescent h/u against the M31-halo constant r0 = 1.8): with the hardness parts switched off (FLRr_z3, rate excursion only), M31 MOS1 becomes −0.67 pp [−1.64, +0.35] (no significant difference) and the LH falls from −5.53 to −2.91. The same gate drops 51/142 observations (MOS1, 34% of 2.filter's time) and 54/142 (MOS2, 38%) on the Virgo-server set — M31 Core, NGC 1365, M32, NGC 1399's hot gas: the sky of bright or hard targets is harder than 1.8, and a diffuse-emission analysis would lose all its data. Next, R13: pre-register MS_z3.5_core and FLRr_z3, each with (a) the field-referenced rejection gate XJ_g1.5 (E021), (b) the R11 line gate, (c) both, scored on all four sets, with the PI deciding between FLR_z3 and the gated MS.
• 跨观测静息参考（R2 / E021）已经跑完，没有产生新的领先政策。两个执行臂（codexnim 与 Hermes）共 8 个家族 42 个掩蔽，全部无角落。结论：在 M31 上场参照不打败单观测领先政策（codexnim HSH_q20_z3 MOS1 −0.21 [−1.18, +0.78] 持平、MOS2 −2.75；Hermes HRt_z4 排名 15 / 12，不是领先者——报告里"MOS2 全场最好"的说法已被 QA 更正）。它的价值是失效探测器：XJ_g1.5 在两侧相机都拒绝 M31 3101、1701 与 LH 0101（开发集误拒 5/45），而 FFC_q20_z4 完整拒绝 LH 0101 但只把 M31 3101 切到 0.35–0.42。Hermes 臂的 XQ 家族仍在 QA 中（待复核 / under review）。R11 的线锚门（P003 已冻结）只拒绝一个测试观测 0674210601——一个 2.filter 漏掉的真实高电平观测，而两个领先政策已经修好它，所以门被保留为 D017 的拒绝规则（安全网），不作为 σ_R 改进。 The cross-observation quiescent reference (R2 / E021) has finished and produced no new leader. Two executor arms (codexnim and Hermes), 8 families and 42 masks in total, all corner-free. Conclusion: on M31 a field reference does not beat the single-observation leaders (codexnim HSH_q20_z3 MOS1 −0.21 [−1.18, +0.78] no significant difference, MOS2 −2.75; Hermes HRt_z4 ranks 15 / 12, not a leader — the report's "best in the whole pool on MOS2" was corrected by QA). Its value is as a failure detector: XJ_g1.5 rejects M31 3101, 1701 and LH 0101 on both cameras (5/45 false drops on development), while FFC_q20_z4 rejects LH 0101 completely but only cuts M31 3101 to 0.35–0.42. The Hermes XQ families are still being QA'd (under review). R11's line-anchored gate (P003 frozen) rejects exactly one test observation, 0674210601 — a real high-level observation 2.filter missed, which both leaders already repair, so the gate is kept as the D017 rejection rule (a safety net), not as a σ_R improvement.
• "复现 2.filter" 的定义已改为文件级（PI 2026-10-16；R10 在跑）。R3 的预算规则被 RT-17 修正为"水平预算 AND 活时间在 ±10% 内"（E_k2.5 403/446，E_k1 13/446），但它只是 E_k2.5 的水平一致性诊断；真正判据是我们的链喂进 2.filter 的区间后写出同样的 gti/allevc/lightcurve 文件。R10 的 Layer A 已在 6 个 CDFS 相机案例上通过（30/30 个文件 0 数据差、0 结构关键字差）；espfilt 直方图拟合的 PRIMARY 关键字（CENTER、WIDTH、NORM…）在复现文件里缺失，是否被下游读取要等 Task 4 冒烟测试。Layer B 与 codexnim 臂仍在跑。 The definition of "reproducing 2.filter" is now file-level (the PI, 2026-10-16; R10 running). R3's budget rule was corrected by RT-17 to "level budget AND live time within ±10%" (E_k2.5 403/446, E_k1 13/446), but it is only a level-agreement diagnostic of E_k2.5; the real criterion is that our chain, fed 2.filter's intervals, writes the same gti/allevc/lightcurve files. R10's Layer A has passed on the 6 CDFS camera cases (30/30 files with 0 data and 0 structural-keyword differences); the espfilt histogram-fit PRIMARY keywords (CENTER, WIDTH, NORM…) are absent from the reproduced files, and whether any downstream task reads them is what the Task-4 smoke test must show. Layer B and the codexnim arm are still running.
• 政策是在 6 个开发案例上选的。本页的领先政策（FLR_z3、JFR_z3、RT3_h2.5_l1、MG_CP_k0.5、MS_z3.5_core、GATE）出自 E012/E015 在六个基本案例上的设计；25 个测试观测的数字是"没被设计者看过"的，但案例本身是开发数据。R9 已把三个稳健政策对 2.filter 的优势复制到 CDFS、LH 与 Virgo-server 集（8 个集×相机全部显著）；GATE 在 CDFS/LH 上丢光时间、EH_B1000_n2 不修复 espfilt 的失效（M31、LH）但在 espfilt 正常的地方有帮助（E019 R9 与 RT-16 F-02 的措辞更正）。RT-16 F-01：优势主要是修好 2.filter 失效的观测——M31 MOS1 去掉两个最差观测后 FLR_z3 只剩 −0.79 pp、MS_z3.5_core 变成 +0.33 pp；三个领先政策的活时间重叠 0.79–0.93，是同一个设计家族而非三个独立确认（F-04）。任何政策在被推出去之前仍需一次新鲜场检验。 The policies were chosen on 6 development cases. The leaders on this page (FLR_z3, JFR_z3, RT3_h2.5_l1, MG_CP_k0.5, MS_z3.5_core, GATE) were designed on the six basic cases in E012/E015; the 25 test observations were never seen by a designer, but the cases themselves are development data. R9 has replicated the three robust policies' advantage over 2.filter on CDFS, the LH and the Virgo-server set (significant in all 8 set × camera cases); GATE drops nearly all time on CDFS/LH, and EH_B1000_n2 does not repair espfilt's failures (M31, LH) but helps where espfilt works (the E019 R9 wording corrected by RT-16 F-02). RT-16 F-01: the advantage is mostly the repair of 2.filter's failed observations — on M31 MOS1, removing the two worst observations leaves FLR_z3 at only −0.79 pp and turns MS_z3.5_core into +0.33 pp; the three leaders' live time overlaps 0.79–0.93, so they are one design family, not three independent confirmations (F-04). Any policy still needs one fresh-field test before being put forward.
• 谱验证（C2）还没做。C001 的光变曲线判据最终要用光谱一致性检验（MosaicSelfCal CDFS 金/银 cell 上的 τ 排名）；PI 决定 C2 在 P002 之后（D008）。 Spectral validation (C2) has not been done. C001's light-curve criterion must finally be checked against spectral consistency (the τ ranking on the MosaicSelfCal CDFS golden/silver cells); the PI decided C2 waits for P002 (D008).
• QPB 没有减。本页所有剖面与偏移都是毛计数：天空 + QPB + 残余软质子混在一起。QPB（约 10% 量级）的差异会以"每观测偏移"的形式出现在第 6/7 节里；分开它需要 MosaicSelfCal 的背景建模。 QPB is not subtracted. Every profile and offset on this page is a gross count: sky + QPB + residual soft protons together. QPB differences (of order 10%) appear as "per-observation offsets" in sections 6 and 7; separating them needs MosaicSelfCal's background modelling.
• R2（跨观测静息参考）已跑完，但没有产生新的领先政策（E021，2026-10-16；详见上文）。它针对的正是本页展示的两类 2.filter 失效（0800733101 与 0022740101——espfilt 把静息峰锁到耀斑上）：场参照能完整拒绝 LH 0101，却只能把 M31 3101 切到 0.35–0.42。它的下一步不是当更好的裁剪，而是当拒绝规则：R13 预注册"领先政策 + 场参照拒绝门 / 线门"的组合，PI 在 FLR_z3 与加了门的 MS_z3.5_core 之间决定。 R2 (the cross-observation quiescent reference) has finished and produced no new leader (E021, 2026-10-16; see above). It targets exactly the two 2.filter failure classes shown on this page (0800733101 and 0022740101 — espfilt locking its quiescent peak onto a flare): a field reference rejects LH 0101 completely but only cuts M31 3101 to 0.35–0.42. Its next step is not as a better clip but as a rejection rule: R13 pre-registers "leader + field-referenced rejection gate / line gate", with the PI deciding between FLR_z3 and the gated MS_z3.5_core.