策略详解：MS_z3.5_corePolicy: MS_z3.5_core

1. 一段话回答1. One-paragraph answer

MS_z3.5_core 是一条"只看高能段、逐段检验正超出"的 GTI 规则。它只读每次观测自己的 2.5–8 keV 视场光变曲线（不看未曝光角落、不读归档 2.filter、不按 ObsID 分支），把耀斑定义成"一段持续的正超出"：先估一个安静参考率，再找出所有超出参考 0.7 个泊松标准差的连续段，只把"整段累计超出超过 z² 个泊松偏差且至少含 200 s 活时间"的段连同一格保护带一起删掉。它来自 2026-10-14 的 E012 政策搜索（执行者 sol，六个已消费设计案例），模块此后冻结未再调（SHA-256 fb2e0451…）。在 E019 的 R9 跨集检验里，它是三个在 25 观测的 M31 test 子集与其余三个集上全部 8 个"集 × 相机"格子里都优于 2.filter 的政策之一，保留 0.74–0.88 倍 2.filter 的时间，每个集里排名前 6–25。但 RT-16（2026-10-16 已裁定）指出：这个胜利大部分是"修好了 2.filter 自己失败的观测" —— 在 M31 MOS1 上去掉 2.filter 最差的两个观测（0301、2601）后，MS 的配对差从 −0.80 变成 +0.33，即持平。RT-16 当时提到的 FLR_z3 领先，到 E022/R13 已查明来源：它靠一条绝对硬度门把 Virgo-server 集 34–38% 的曝光整段丢掉；而 MS 加上 E021 的场门（XJ.MS）在 M31 上是 −1.53 [−2.31, −0.73] / −3.84，会否掉 0800733101 与 0800731701，但在 P004 的三条冻结读数条件里失手两条（LH MOS2 上对 FLR_z3 的 Δσ_R 的 p84 为 +0.81，要求 ≤ +0.5 pp；Virgo-server 上丢 8 个观测，而 FLR_z3 丢 51–54，要求 ≤ 它所丢的十分之一即 5.1–5.4）。主候选留给 PI 决定。还悬着的事：R7（剖面预注册）排队中；谱验证 C2 未做；R2/E021 已完成，其跨观测参考没能超过单观测领先政策。

MS_z3.5_core is a "high-band-only, run-by-run positive-excess" GTI rule. It reads only each observation's own 2.5–8 keV FOV light curve (no unexposed corner, no archived 2.filter, no ObsID branching) and defines a flare as a coherent positive excursion: estimate a quiet reference rate, find every contiguous run above it by more than 0.7 Poisson standard errors, and cut a run only when its integrated positive Poisson deviance exceeds z² and it holds at least 200 s of live time — plus one cell of padding. It comes from the E012 policy search of 2026-10-14 (executor sol, six consumed design cases), and the module has been frozen since (SHA-256 fb2e0451…). In E019's R9 cross-set test it is one of the three policies that beat 2.filter in all 8 set × camera cases on the 25-observation M31 test subset and the other three sets, keeping 0.74-0.88 of 2.filter's time and ranking in the top 6-25 of each set. But RT-16 (triaged 2026-10-16) finds that most of this win is "repairing 2.filter's own failed observations" — on M31 MOS1, dropping 2.filter's two worst observations (0301, 2601) takes MS's paired difference from -0.80 to +0.33, i.e. no significant difference. The FLR_z3 lead that RT-16 mentioned was traced by E022/R13 to an absolute hardness gate that drops 34-38% of the Virgo-server exposure outright; MS plus the E021 field gate (XJ.MS) gives -1.53 [-2.31, -0.73] / -3.84 on M31 and rejects 0800733101 and 0800731701, but it fails two of P004's three frozen reading conditions (on LH MOS2 its delta sigma_R against FLR_z3 has p84 = +0.81, required ≤ +0.5 pp; it drops 8 Virgo-server observations, required ≤ one tenth of FLR_z3's 51–54, i.e. 5.1–5.4). The main candidate awaits the PI. Still open: R7 (the pre-registered profile test) is queued; the spectral check C2 has not been done; R2/E021 is complete and its cross-observation reference did not beat the single-observation leaders.

2. 它用什么、不用什么2. What it uses and what it does not

3. 算法，从零一步步讲3. The algorithm, step by step, from zero

下面每一步都对应冻结模块 policies_sol.py 里 scan_core(t, c, z) 的一段代码，z = 3.5 就是名字里的 3.5。先说清楚：z = 3.5 是设计参数，不是标定过的假阳性率。搜索里同时试过 3.5/4/4.5/5/6/7 六个值，报告明确写了"检测阈值是启发式设计参数，不是标定的假阳性概率"。

Each step below maps to one block of scan_core(t, c, z) in the frozen module policies_sol.py, and z = 3.5 is the 3.5 in the name. State plainly: z = 3.5 is a design parameter, not a calibrated false-positive rate. The search tried 3.5/4/4.5/5/6/7 together, and the report says "detection thresholds are heuristic design parameters, not calibrated false-positive probabilities".

第 0 步：把观测切成 100 s 的格子Step 0: cut the observation into 100 s cells

每个 20 s 的原始 cell 按时间归入 100 s 的格子（blocks 函数）。每格得到两个数：活时间 t_i（秒，来自 EXPOSU，不是"有事件的格子数"）和高能段计数 c_i = h_i + u_i（2.5–8 keV 加 8–12 keV）。

Every raw 20 s cell is binned into 100 s cells (the blocks function). Each cell gets two numbers: live time t_i (seconds, from EXPOSU, not "cells with events") and high-band count c_i = h_i + u_i (2.5–8 keV plus 8–12 keV).

第 1 步：估一个安静参考率 μStep 1: estimate a quiet reference rate mu

这一步是整条规则的"锚"。把格子按 5 个一组并成 500 s 的块，取这些块的率的下四分位数，再按泊松噪声修正：μ ← μ + 0.67449·√(μ/中位活时间)。然后做 4 轮"只信低的那部分"的迭代：凡是计数超过 μ·t + 3.5√(μ·t) 的块都算被污染，剔除后重新加权平均，重复 4 次。为什么不能直接用中位数：当大部分时间都在耀发时，中位数本身就坐在耀斑里（PITFALLS 记了这个失败）。

This is the rule's anchor. Group the cells into 500 s blocks of 5, take the lower quartile of their rates, then correct for Poisson noise: mu ← mu + 0.67449·√(mu/median live time). Then run 4 rounds of "trust only the low part": any block with counts above mu·t + 3.5√(mu·t) is treated as contaminated, and mu is re-averaged over the rest, 4 times. Why not the plain median: when most of the time is flaring, the median itself sits inside the flare (PITFALLS records this failure).

第 2 步：找"支撑段"（support runs）Step 2: find the support runs

对每个 100 s 格子，把计数和期望值都做 3 格（300 s）的滑动平均，然后问：平滑后的超出 (c − e) 是否大于 0.7 个泊松标准差 √e？是就是 1，不是就是 0。再用 3 格的结构元素做一次闭运算（把 ≤ 2 格的短缝隙填上），得到一串连续的 1 段 —— 这些就是候选耀斑段。观测的首尾格子单独判断，不让边界把一段耀斑切成两半。

For each 100 s cell, smooth both the counts and the expectation over 3 cells (300 s) and ask: is the smoothed excess (c − e) larger than 0.7 Poisson standard errors √e? Yes → 1, no → 0. Then apply a binary closing with a 3-cell structure (fills gaps of up to 2 cells), giving a set of contiguous runs of 1s — the candidate flare runs. The first and last cells are judged separately, so an edge cannot cut a flare in half.

第 3 步：逐段做泊松检验，决定切不切Step 3: test each run and decide whether to cut

对每个支撑段 [a, b)，把整段的观测计数 n 与期望计数 e 相加，算正的泊松偏差（Poisson deviance）D = 2·[n·ln(n/e) − n + e]。只有当三个条件同时成立才切：n > e（必须是正超出）、D > z² = 12.25（z = 3.5 时）、且该段活时间 ≥ 200 s。最后把切掉的段各向外膨胀 1 格（100 s 保护带）。

For each run [a, b), sum the observed counts n and expected counts e over the whole run and compute the positive Poisson deviance D = 2·[n·ln(n/e) − n + e]. A run is cut only when all three hold: n > e (it must be a positive excess), D > z² = 12.25 (for z = 3.5), and the run holds at least 200 s of live time. Finally each cut run is dilated by 1 cell (100 s of padding).

"core" 改了第 2 轮什么What "core" changed in round 2

第 1 轮的 MS 用一组很宽的滑动窗口（4/8/16/32/64 格）扫描，任何超标的窗口都把覆盖的格子标成坏。宽窗口的毛病是：一个大耀斑会污染很多与它重叠的长窗口，把旁边无关的安静时间一起删掉。第 2 轮改成上面的"支撑段 + 整段检验"。效果在 2801 上最清楚：第 1 轮 MS_z3.5 只保留 10.88% / 11.52%（M1 / M2），第 2 轮 MS_z3.5_core 保留 75.67% / 74.99%。

Round 1's MS scanned with a set of broad sliding windows (4/8/16/32/64 cells) and marked every covered cell bad whenever a window exceeded the threshold. The flaw: one big flare contaminates many long overlapping windows and deletes unrelated quiet time next to it. Round 2 replaced this with the "support run + whole-run test" above. The effect is clearest on 2801: round-1 MS_z3.5 kept only 10.88% / 11.52% (M1 / M2); round-2 MS_z3.5_core keeps 75.67% / 74.99%.

可以照抄的伪代码Pseudo-code a student can re-implement

def ms_z3p5_core(cells, z=3.5):
    # cells: list of (live_seconds t_i, high_band_counts c_i) on a 100 s grid
    # 1. reference rate mu on 500 s blocks
    blocks = group_by_5(cells)
    mu = lower_quartile([c_i / t_i for blocks])
    mu += 0.67449 * sqrt(mu / median(t))
    for _ in range(4):
        ok = [c <= mu*t + 3.5*sqrt(max(mu*t, 1)) for each block]
        mu = sum(c[ok]) / sum(t[ok])
    expected = [mu * t_i for each cell]
    # 2. support runs: 300 s smoothed excess above 0.7 sigma
    s = [ (smooth3(c)[i] - smooth3(expected)[i]) > 0.7*sqrt(max(smooth3(expected)[i],1)) ]
    s = binary_closing(s, 3 cells)          # fill gaps of up to 2 cells
    s[0] = (c[0]-expected[0]) > 0.7*sqrt(max(expected[0],1))   # edges judged alone
    s[-1] = (c[-1]-expected[-1]) > 0.7*sqrt(max(expected[-1],1))
    # 3. cut a run only if it is a positive, significant, long-enough excursion
    bad = zeros
    for (a, b) in runs_of_true(s):
        n = sum(c[a:b]); e = sum(expected[a:b])
        D = 2 * (n*log(n/e) - n + e)
        if n > e and D > z*z and sum(t[a:b]) >= 200:
            bad[a:b] = True
    keep = not dilate(bad, 1 cell)          # 100 s padding
    return keep

4. 五个观测的实做例子4. Five worked examples

每张图画的是同一批东西：蓝色细线是 20 s 格子的 2.5–8 keV 视场率；灰色虚线是 MS 自己估的安静参考率 μ；红色阴影是 MS 真正删掉的段；底部黑色粗条是归档 2.filter 保留的时间。所有数字都是从缓存光变曲线现算的（脚本 make_figures_policy_ms.py，与冻结模块同一份代码路径）。

Every figure draws the same things: the thin blue line is the 20 s-cell 2.5–8 keV FOV rate; the grey dashed line is MS's own quiet reference rate mu; the red shading is what MS actually cuts; the thick black bar at the bottom is the archived 2.filter kept time. All numbers are recomputed from the cached light curves (script make_figures_policy_ms.py, the same code path as the frozen module).

Lockman Hole 0101（0022740101）:缓存里也有它。MS 在 M1 上只保留 3.5%、M2 上 3.4%，2.filter 保留 23.4% / 17.4%。这是 RT-06 认定的 2.filter 失败（espfilt 峰锁在耀斑上，6.8 倍中位数），MS 在这里比 2.filter 切得更狠 —— 是更接近真相还是过度切割，R2（跨观测安静参考）要回答的正是这个问题。

Lockman Hole 0101 (0022740101): it is also in the cache. MS keeps 3.5% on M1 and 3.4% on M2, against 2.filter's 23.4% / 17.4%. This is the 2.filter failure identified by RT-06 (espfilt's peak locks onto a flare, 6.8× the median); MS cuts harder than 2.filter here — whether that is closer to the truth or over-cutting is exactly what R2 (a cross-observation quiet reference) is meant to answer.

5. 它在 Pareto 前沿上的位置5. Where it sits on the Pareto front

单次观测内：保留时间对纯泊松误差Within one observation: kept time against Poisson-only error

下面两张图画的是 157 个候选政策（基线 + E012 三个执行者的合并模块）在同一观测里的位置。横轴是保留活时间，纵轴是 σ_stat = 1/√N（N = 保留的 0.5–2 keV 计数，30″ 掩模）。注意这张图的纵轴只是计数噪声，不含系统项，所以"越靠右下越好"只对"曝光几乎免费"的深场成立；在 M31 上系统项主导（见第 6 节）。

The two figures below show 157 candidate policies (the baseline plus the merged E012 modules of three executors) in the same observation. The x-axis is kept live time and the y-axis is sigma_stat = 1/√N (N = kept 0.5–2 keV counts, 30" mask). Note that this y-axis is counting noise only, with no systematic term, so "lower right is better" holds only where exposure is nearly free (deep fields); on M31 the systematic term dominates (see section 6).

叠加后：σ_R 对保留时间，四个集 × 两个相机Stacked: sigma_R against kept time, four sets x two cameras

这是 C001 v3.1 判据（E019）下的位置：σ_R = √(Σ K_o τ²_o / Σ K_o + 1/N_P)，τ²_o 是每次观测的潜在一个非负超出（负二项后验），1/N_P 是拼接后的计数项。每张子图一个"集 × 相机"，横轴是相对 2.filter 的保留时间比。p16/p84 是观测重采样的散布，不是置信区间（RT-08 F02：覆盖率 52%）。

This is the position under the C001 v3.1 criterion (E019): sigma_R = √(Σ K_o τ²_o / Σ K_o + 1/N_P), where τ²_o is each observation's latent nonnegative excess (a negative-binomial posterior) and 1/N_P is the concatenated counting term. Each panel is one set × camera, and the x-axis is the retained time relative to 2.filter. p16/p84 are the spread from resampling observations, not confidence intervals (RT-08 F02: coverage 52%).

与 2.filter 的配对差，带它的散布Paired difference against 2.filter, with its spread

五个政策 × 四集 × 两相机的完整表The full table: five policies x four sets x two cameras

6. 一致性与剖面6. Consistency and profiles

E014:31 个 M31 观测叠成一个马赛克E014: 31 M31 observations as one mosaic

E014 把 31 个 M31 halo 观测当成一次拼接观测，用当时的 v_W 判据打分。在保留 ≥ 0.88 倍 2.filter 曝光的政策里，MS_z3.5_core 是最好的一档：1.3%（M1）与 1.6%（M2），对应 0.88 倍的 2.filter 时间。同一份产品里 2.filter 是 4.4% / 4.9%，绝对最小点（EH_B300_n-1，MOS2）只有 0.55% 但只保留 0.16 倍时间。P(best) 分散在 10–30 个政策上，最大 0.18 —— 最好的政策没有定论。

E014 treated the 31 M31 halo observations as one concatenated pointing and scored them with the then-current v_W criterion. Among policies keeping at least 0.88× of 2.filter's exposure, MS_z3.5_core is the best tier: 1.3% (M1) and 1.6% (M2), at 0.88× of 2.filter's time. In the same products 2.filter gives 4.4% / 4.9%, and the absolute minimum (EH_B300_n-1, MOS2) is 0.55% but keeps only 0.16× of the time. P(best) is spread over 10-30 policies with a maximum of 0.18 — the best policy is not resolved.

E020:垂直剖面一致性 —— "未检出"，而且功效很低E020: vertical-profile consistency — "not detected", and with little power

E020 问的是一个更接近科学产出的问题：换 GTI 政策能不能让 M31 halo 的垂直剖面在不同观测之间更一致？答案经过 RT-12 修正后是：未检出，且这个检验的功效很低。

E020 asks a question closer to the scientific output: can changing the GTI policy make the M31 halo vertical profile more consistent between observations? After the RT-12 correction the answer is: not detected, and the test has little power.

RT-12 的原话：如果 E019 的观测内项与一个结构性项正交相加，从 2.filter 换成领先政策只预测 Δτ_b ≈ −0.8…−1.1 pp（M1）与 −1.9 pp（M2），而配对 bootstrap 的 sd 是 0.7–1.9 pp —— 预测只有 1–2σ。所以"没有政策显著改善剖面"必须读成"未检出，功效低"，不是"剖面与 GTI 无关"。

RT-12's own words: if E019's within-observation term adds in quadrature to a structural term, switching from 2.filter to the leaders predicts only Δτ_b ≈ −0.8…−1.1 pp (M1) and −1.9 pp (M2), while the paired bootstrap sd is 0.7-1.9 pp — the prediction is only 1-2σ. So "no policy significantly improves the profile" must be read as "not detected; low power", not as "the profile is unrelated to the GTI".

RT-12 还撤回了一句话："剩下的散布不是 GTI 效应"是错的。去掉每次观测的 z 无关偏移后，中位 τ_b 从 8.8% → 6.5%（M1）与 11.5% → 8.7%（M2），即 τ_b² 的 43–46% 是一个可用 GTI 修的逐观测偏移。MS 在这个偏移统计上：M1 +2.20 pp [−1.22, +3.78]、M2 −1.28 pp [−3.14, +0.13] —— M2 上方向为负但区间含 0，记持平；M1 上为正。做这个统计的 Q_d0.05_n2 用角落，MS 不用。

RT-12 also withdrew one sentence: "the remaining scatter is not a GTI effect" is wrong. Removing each observation's z-independent offset takes the median τ_b from 8.8% to 6.5% (M1) and 11.5% to 8.7% (M2), i.e. 43-46% of τ_b² is a per-observation offset a GTI can fix. For MS on that offset statistic: M1 +2.20 pp [-1.22, +3.78] and M2 -1.28 pp [-3.14, +0.13] — on M2 the sign is negative but the interval contains 0, so no significant difference; on M1 it is positive. The policy that does well there, Q_d0.05_n2, uses the corner; MS does not.

E010:Al 线稳定性 —— 线检验目前排不出政策名次E010: Al-line stability — the line test cannot rank policies at present precision

E010 对 20 个政策叠合了 ln(L_K/L_B)。结论：Al-K 线在保留与拒绝时间之间稳定到约 1%，所有政策共有一个约 +1% 的偏移，政策之间的差别 ≤ 1% 且可由稀释解释。线检验不能用来给政策排名。MS_z3.5_core 不在 E010 的 20 个政策名单里（那份工作在 E012 之前），所以这里没有 MS 的专属数字 —— 这是空缺，不是通过。

E010 stacked ln(L_K/L_B) over 20 policies. The conclusion: the Al-K line is stable between kept and rejected time to about 1%, all policies share a common +1% offset, and differences between policies are ≤ 1% and explained by dilution. The line test cannot rank policies. MS_z3.5_core is not among E010's 20 policies (that work predates E012), so there is no MS-specific number here — that is a gap, not a pass.

7. 已知失败与限度7. Known failures and limits

8. 如何亲手跑它8. How to run it by hand

政策代码在数据树里，是冻结的只读模块；打分脚本在代码树里。下面每条命令都在 lang 上实跑过（作者 2026-10-16，退出码 0），输出写到 $TMPDIR，不改动任何收据。

The policy code is a frozen read-only module in the data tree; the scoring scripts are in the code tree. Every command below was run for real on lang (by the author on 2026-10-16, exit code 0), writing to $TMPDIR and touching no receipt.

8.1 模块在哪里，怎么读它8.1 Where the module is and how to read it

cd /u/Lang3/program/DiffuseXrayCalibration/ParetoGTICal
# The frozen module (SHA-256 fb2e0451...79b26, unchanged since the E012 search):
#   /u/Lang3/Data/DiffuseXrayCalibration/ParetoGTICal/results/c001_policy_search/sol/policies_sol.py
# MS_z3.5_core = scan_core(t, c, 3.5) inside policies(): out[f"MS_z{z}_core"]
# It is loaded by the scorer under the name "sol." through the merged module:
#   results/c001_policy_search/merged/policies_merged.py   -> "sol.MS_z3.5_core"
# and by c001_risk.py under the name "merged:" -> "merged:sol.MS_z3.5_core"

8.2 导出 GTI 区间8.2 Export GTI intervals

cd /u/Lang3/program/DiffuseXrayCalibration/ParetoGTICal
PYTHONDONTWRITEBYTECODE=1 python3 scripts/c001_export_gti.py --out "$TMPDIR/gti_ms.json"
# Prints per observation, in ks, for none / 2.filter / EH_B1000_n2 / sol:MS_z3.5_core /
# sonnet:CP_h_p6_d0.05.  The JSON is {obsid: {policy: [[t0, t1], ...]}} in TT seconds,
# MOS1 and MOS2 intersected (a time is kept only when both cameras keep it).

实跑输出（2026-10-16，作者）：0201 的 MS 保留 15.6 ks（1 个区间，与 none 相同）；3101 的 MS 保留 2.266 ks（2 个区间），2.filter 2.8 ks。每个 ObsID 只跑六个基本案例，这是 E013 的五仪器检验入口。

Real output (2026-10-16, the author): on 0201 MS keeps 15.6 ks (1 interval, the same as none); on 3101 MS keeps 2.266 ks (2 intervals) against 2.filter's 2.8 ks. It runs only the six basic cases; this is the entry point for E013's five-instrument test.

8.3 重新打分（v3.1 判据）8.3 Rescore with the v3.1 criterion

cd /u/Lang3/program/DiffuseXrayCalibration/ParetoGTICal
# M31 halo (the default --glob 080073*01):
PYTHONDONTWRITEBYTECODE=1 python3 scripts/c001_risk.py --out "$TMPDIR/risk_m31"
# One of the other three sets:
PYTHONDONTWRITEBYTECODE=1 python3 scripts/c001_risk.py --set cdfs     --out "$TMPDIR/risk_cdfs"
PYTHONDONTWRITEBYTECODE=1 python3 scripts/c001_risk.py --set lh_stage --out "$TMPDIR/risk_lh"
PYTHONDONTWRITEBYTECODE=1 python3 scripts/c001_risk.py --set virgo    --out "$TMPDIR/risk_virgo"
# Reads: per_obs.json, risk.json, summary.txt.  Look for the row "merged:sol.MS_z3.5_core".