如何手工续作（不用 AI）How to continue by hand (without AI)

本页写给一个刚接触项目的人：读完之后，你应该能在不借助 AI 的情况下，复现关键数值、理解数据在哪、知道下一步做什么。所有命令都已在 lang 机器上实际运行过（只读命令真实运行，拟合命令只验证用法）。

This page is written for a newcomer: after reading it you should be able to reproduce the key numbers, find the data, and know what to do next — all without AI. Every command was actually run on lang (read-only commands run for real; fitting commands were verified for usage only).

机器与分工The machines and their roles

来源：E013/E014 BRIEF（.agent/jobs/2026-10-14-e013-joint-profile/BRIEF.md 与 ...-e014-nw-crosscheck/BRIEF.md）的「Fit on lang only」与数据路径约定；data 符号链接指向 ../../Data/XMM/M31_data。

Source: the "Fit on lang only" rule and data-path convention in the E013/E014 BRIEFs (.agent/jobs/2026-10-14-e013-joint-profile/BRIEF.md and ...-e014-nw-crosscheck/BRIEF.md); the data symlink points to ../../Data/XMM/M31_data.

数据、代码、结果在哪Where the data, code and results live

可直接复制的命令（均已验证）Copy-paste commands (all verified)

命令 1：复现形状上限快速检验（只读，几秒）Command 1: reproduce the quick shape-limit check (read-only, seconds)

读冻结的 v19 逐场流量表，重新做「轮廓是平的吗、外 NW 四场亮吗、M31 β 分量能有多大」的快速检验。输出与 newangels3_discussion/quick_shape_limit_20261014.txt 一致（该脚本也会重写同目录下两个 CSV——它们是派生文件；若不想覆盖，先备份）。

It reads the frozen v19 per-field flux table and re-runs the quick check "is the profile flat, are the outer-NW four bright, how large can an M31 beta component be". The output matches newangels3_discussion/quick_shape_limit_20261014.txt (the script also rewrites two derived CSVs in the same directory; back them up first if you do not want that).

cd /u/Lang3/program/M31CGM
/home/huangrui/software/miniconda3/bin/python newangels3_discussion/quick_shape_limit_20261014.py

实际输出（2026-10-14 运行）：

Actual output (run 2026-10-14):

constant: chi2=11.99/13 p=0.529
linear in r: coef=[0.0126] sig=[1.92] chi2=8.31/12 dchi2=3.68
linear in b: coef=[0.151] sig=[2.49] chi2=5.77/12 dchi2=6.22
linear in r+b: coef=[0.0006 0.1471] sig=[0.06 1.59] chi2=5.77/11 dchi2=6.22
outer NW four (a posteriori): 1.114 +- 0.071
other ten: 0.870 +- 0.056

 beta  rc_kpc mw_model     A20  A20_err  A20_ul95_approx   chi2  dof
  0.4     4.0    const -0.1276   0.0751           0.1236 9.1031   12
  0.4     4.0 linear_b  0.0122   0.1072           0.1886 5.7564   11
  0.5     4.0    const -0.0737   0.0461           0.0758 9.4302   12
  0.5     4.0 linear_b  0.0109   0.0638           0.1158 5.7402   11
  0.6     4.0    const -0.0453   0.0304           0.0499 9.7635   12
  0.6     4.0 linear_b  0.0092   0.0407           0.0761 5.7187   11

来源：newangels3_discussion/quick_shape_limit_20261014.py；输出与 newangels3_discussion/quick_shape_limit_20261014.txt 逐行一致。

Source: newangels3_discussion/quick_shape_limit_20261014.py; the output matches newangels3_discussion/quick_shape_limit_20261014.txt line by line.

命令 2：列出 v19 主样本 14 场表（只读）Command 2: list the v19 primary 14-field table (read-only)

一行一行看 14 个主样本场的半径、kT、norm、N_H 与吸收流量。也可以直接打开 CSV：independent_cgm_results_v19_mos14to20keV_excluded_hi4pi_nh_per_pointing/independent_cgm_v19_primary_dual_mos_measurements.csv（14 行 + 表头）。

Inspect, field by field, the radius, kT, norm, N_H and absorbed flux of the 14 primary fields. You can also open the CSV directly: independent_cgm_results_v19_mos14to20keV_excluded_hi4pi_nh_per_pointing/independent_cgm_v19_primary_dual_mos_measurements.csv (14 rows plus header).

cd /u/Lang3/program/M31CGM
/home/huangrui/software/miniconda3/bin/python -c "import pandas as pd; \
d = pd.read_csv('independent_cgm_results_v19_mos14to20keV_excluded_hi4pi_nh_per_pointing/independent_cgm_v19_primary_dual_mos_measurements.csv', dtype={'obsid': str}); \
print(d[['obsid','rproj_kpc','mwhalo_kt_keV','mwhalo_norm']].to_string(index=False))"

实际输出（节选前 5 行；norm 默认显示为 6 位小数）：

Actual output (first 5 rows; the norm column prints 6 decimals by default):

obsid  rproj_kpc  mwhalo_kt_keV  mwhalo_norm
0800730201  28.793728       0.167030     0.000005
0800730301  28.687554       0.164971     0.000005
0800730501  23.396669       0.167798     0.000005
0800730601  25.028293       0.219833     0.000004
0800730701  19.544786       0.222737     0.000002

来源：independent_cgm_results_v19_.../independent_cgm_v19_primary_dual_mos_measurements.csv（本页运行时的实际打印）。

Source: independent_cgm_results_v19_.../independent_cgm_v19_primary_dual_mos_measurements.csv (printed live when this page was written).

命令 3：运行一个 v19 生产拟合（只展示用法，不要运行）Command 3: run one v19 production fit (usage shown only — do NOT run)

v19 产品已冻结（读只读目录），重新拟合会覆盖冻结会话。如果你确实要重跑某个 ObsID（例如在新建的实验目录里），用法如下——参数是一个 10 位 ObsID，脚本会校验哈希、在 logs_obsid_fitting_v19_... 写日志。运行约需数分钟到更久（单 ObsID 多阶段拟合）。

The v19 products are frozen (their directory is read-only); re-fitting would overwrite the frozen sessions. If you truly need to re-run one ObsID (e.g. inside a new experiment directory), the usage is — one 10-digit ObsID argument; the script checks hashes and writes to logs_obsid_fitting_v19_.... Expect several minutes or more (multi-stage fit).

cd /u/Lang3/program/M31CGM
bash run_m31_cgm_v19_single_obsid.sh 0800730201

不带参数时的实际输出（证明用法行来自脚本本身）：

Actual output without arguments (proving the usage line comes from the script):

Usage: run_m31_cgm_v19_single_obsid.sh 10_DIGIT_OBSID

脚本内部先做依赖检查（脚本/helper/N_H 表/HI4PI 图/验证器），再导出 CIAO 环境变量（run_m31_cgm_v19_single_obsid.sh 第 95–106 行），然后调用 background_spectrum_fitting_2T_basedon_obsid_v19.py。不要在 lang 上与正在运行的 E013/E014 任务并发重跑。

The script first checks dependencies (script/helper/N_H table/HI4PI map/validators), then exports the CIAO environment (lines 95–106 of run_m31_cgm_v19_single_obsid.sh), then calls background_spectrum_fitting_2T_basedon_obsid_v19.py. Do not re-run it concurrently with the E013/E014 jobs on lang.

来源：run_m31_cgm_v19_single_obsid.sh 第 1–30 行与 95–106 行（本页写作时读取）。

Source: run_m31_cgm_v19_single_obsid.sh lines 1–30 and 95–106 (read when this page was written).

命令 4：恢复一个 v19 Sherpa 会话（只展示，不要运行）Command 4: restore one v19 Sherpa session (shown only — do NOT run)

恢复（restore）会话是 E013 A1 与 E014 B3 的第一步。恢复本身只读，但脚本里任何 save() 或写文件都会破坏冻结状态——所以这里只展示。必须用 CIAO 的 Python（/home/huangrui/software/ciao-4.18/binexe/python3.12），并且要先导出 CIAO 环境变量，否则 from sherpa.astro import xspec 会报 ImportError: XSPEC support is not enabled（本页写作时实测确认）。

Restoring a session is the first step of E013 A1 and E014 B3. Restoring itself is read-only, but any save() or file write inside the script would break the frozen state — so this is shown only. You must use the CIAO Python (/home/huangrui/software/ciao-4.18/binexe/python3.12) and export the CIAO environment first; otherwise from sherpa.astro import xspec fails with ImportError: XSPEC support is not enabled (confirmed live while writing this page).

export HEADAS=/home/huangrui/software/ciao-4.18/bin
export ASCDS_INSTALL=/home/huangrui/software/ciao-4.18
export ASCDS_CALIB=/home/huangrui/software/ciao-4.18/data
export ASCDS_CONFIG=/home/huangrui/software/ciao-4.18/config
export ASCDS_BIN=/home/huangrui/software/ciao-4.18/bin
export PFILES="/u/Lang2/ai_tmp//pfiles;/home/huangrui/software/ciao-4.18/param"
export PYTHONHOME=/home/huangrui/software/ciao-4.18
export PYTHONPATH=/home/huangrui/software/ciao-4.18/lib/python3.12/site-packages
export LD_LIBRARY_PATH="/home/huangrui/software/ciao-4.18/lib:$LD_LIBRARY_PATH"
export MPLBACKEND=Agg

然后在 Python 里（恢复会话前先设置 XSPEC 物理配置，与 v19 脚本第 812–820 行一致）：

Then in Python (set the XSPEC physics before restoring, matching lines 812–820 of the v19 script):

from sherpa.astro import xspec
xspec.set_xsabund('angr')
xspec.set_xsxsect('vern')
xspec.set_xsxset('APECROOT', '/home/huangrui/software/ciao-4.18/spectral/modelData/apec_v3.0.9')

import sherpa.astro.ui as ui
ui.restore('background_spectrum_fitting_2T_basedon_obsid_v19_mos14to20keV_excluded_hi4pi_nh_per_pointing/0800730201_spectrum.fitting')
print(ui.get_stat_name())          # chi2gehrels
print(ui.calc_stat())              # compare with the CSV column fit_stat (30.571259873047417)

判读：恢复后 calc_stat() 应与 CSV 的 fit_stat 一致（E013 验收标准 |Δ| ≤ 1e-3 相对）。0800730201 的 fit_stat = 30.571259873047417（21 dof）来自测量 CSV。

How to read it: after restore, calc_stat() must match the CSV column fit_stat (E013 acceptance |Δ| ≤ 1e-3 relative). For 0800730201, fit_stat = 30.571259873047417 (21 dof) comes from the measurement CSV.

来源：环境变量与 XSPEC 设置照抄 run_m31_cgm_v19_single_obsid.sh 第 95–106 行与 background_spectrum_fitting_2T_basedon_obsid_v19.py 第 812–820 行；fit_stat 与 dof 来自 independent_cgm_v19_all22_measurements.csv 的 0800730201 行；ImportError 文案为实测。

Source: the environment variables and XSPEC settings copy run_m31_cgm_v19_single_obsid.sh lines 95–106 and background_spectrum_fitting_2T_basedon_obsid_v19.py lines 812–820; fit_stat and dof from the 0800730201 row of independent_cgm_v19_all22_measurements.csv; the ImportError text was reproduced live.

新手必须遵守的规则Rules a newcomer must keep

1. 临时文件只写到 /u/Lang2/ai_tmp/<你>/，绝不写 /tmp、/var/tmp 或家目录以外的磁盘根。Temp files go to /u/Lang2/ai_tmp// only; never /tmp, /var/tmp, or the root disk.
2. v19 产品已冻结：会话目录与测量 CSV 只读，不得覆盖、移动或重命名。The v19 products are frozen: the session directory and measurement CSVs are read-only; never overwrite, move, or rename them.
3. MOS 1.4–2.0 keV 一律剔除（Al/Si 仪器线），拟合带为 0.4–1.4 + 2.0–8.0 keV；能量上限必须与代码和日志一致。MOS 1.4–2.0 keV is always excluded (the Al/Si instrumental lines); the fitted band is 0.4–1.4 + 2.0–8.0 keV, and the upper bound must match code and logs.
4. N_H 用逐指向 HI4PI 值（15 arcmin 孔径平均，原子 N_HI），不用单一全样本值。N_H is per pointing from HI4PI (15-arcmin aperture mean, atomic N_HI), never one sample-wide value.
5. 软质子（SP）与仪器线 norm 逐谱自由；任何天体物理联动假设必须显式声明。Soft-proton (SP) and instrumental-line norms stay free per spectrum; any astrophysical linking assumption must be stated explicitly.
6. 共享的 0.175 keV 分量是 MW+M31 之和（CGMsum），绝不称它「纯 MW」或「纯 M31」。The shared 0.175 keV component is the MW+M31 sum (CGMsum); never call it "pure MW" or "pure M31".
7. 非负分量的上限需要边界感知区间（profile likelihood 或模拟校准），不能用普通 F 检验。Limits on non-negative components need boundary-aware intervals (profile likelihood or simulation calibration), not a naive F-test.
8. lang 共享：新任务 OMP/OPENBLAS/MKL_NUM_THREADS=1；不碰别人的进程，不用默认 tmux server。lang is shared: any new job sets OMP/OPENBLAS/MKL_NUM_THREADS=1; do not touch other processes and never use the default tmux server.

来源：规则 1/8 来自 TEMP_RULE 与 E013/E014 BRIEF 的 Rules 节；规则 2–7 来自 cognition-os/00_AI_BRIEFING.md「Methodology invariants」1–8 条；规则 7 另见 03_TRUST_TABLE.md「delta_stat F-test classification (D22)」。

Source: rules 1/8 from TEMP_RULE and the Rules sections of the E013/E014 BRIEFs; rules 2–7 from the "Methodology invariants" (items 1–8) of cognition-os/00_AI_BRIEFING.md; rule 7 also 03_TRUST_TABLE.md "delta_stat F-test classification (D22)".

编号计划：下一步怎么做The numbered plan: what to do next

步骤 1：等待并核对 E013 A1 曲面Step 1: wait for and check the E013 A1 surfaces

• 目标：20 个可恢复会话的逐场 Δstat(kT, norm) 曲面（E013 A1）。Goal: per-field Δstat(kT, norm) surfaces for the 20 restorable sessions (E013 A1).
• 输入：冻结 v19 会话 + independent_cgm_v19_all22_measurements.csv。Inputs: the frozen v19 sessions and independent_cgm_v19_all22_measurements.csv.
• 方法：按 E013 BRIEF 的网格（kT 0.155–0.200 × norm 11 点）逐点冻结再拟合其余参数。Method: per the E013 BRIEF grid (kT 0.155–0.200 × 11 norm points), freeze kT/norm and re-fit the rest.
• 成功判据：恢复回归 |Δstat| ≤ 1e-3 相对、通量比 0.99–1.01、无缺失格点。Success/failure: restore regression |Δstat| ≤ 1e-3 relative, flux ratio 0.99–1.01, no missing grid points.
• 时间：≤12 h 墙钟（≤6 workers）。Time: ≤12 h wall (≤6 workers).
• 先读：.agent/jobs/2026-10-14-e013-joint-profile/BRIEF.md。Read first: .agent/jobs/2026-10-14-e013-joint-profile/BRIEF.md.

步骤 2：E013 A2 空间模型联合似然Step 2: the E013 A2 joint spatial-model likelihood

• 目标：MW（常数或线性 b）+ M31 β 项的谱级上限，替代 I033 的快速流量级上限。Goal: a spectral-level limit on MW (constant or linear in b) + an M31 beta term, superseding the quick flux-level limit of I033.
• 输入：A1 曲面 CSV。Inputs: the A1 surface CSVs.
• 方法：Σ_fields Δstat_i(kT, norm_i = N_MW(b_i) + A·g(r_i))，profile 掉 N_MW/梯度/kT，A=0 边界用模拟校准单侧上限。Method: Σ_fields Δstat_i(kT, norm_i = N_MW(b_i) + A·g(r_i)); profile over N_MW/the gradient/kT; calibrate one-sided limits at the A=0 boundary by simulation.
• 成功判据：给出 (r_c, β) 表上的 A 上限；与 I033 快速版同数量级为预期，差异需解释。Success/failure: a table of A limits over (r_c, β); the same order of magnitude as the quick I033 version is expected, and any difference must be explained.
• 时间：1–2 天（插值与模拟，CPU 轻量）。Time: 1–2 days (interpolation and simulations; light CPU).
• 先读：cognition-os/experiments/E013_joint_spatial_profile_likelihood.md。Read first: cognition-os/experiments/E013_joint_spatial_profile_likelihood.md.

步骤 3：读 E014 结果并跟进 D49 排队检验Step 3: read the E014 results and follow up the D49 queued tests

• 目标：外 NW 四场偏亮是天空结构、SWCX 还是仪器？——E014 已回答：都不是。剩余候选是残余软质子/仪器（SP 比值 3.1 但相关性弱，Spearman +0.35，p=0.22）、小于 15′ 的结构（RASS 孔径以下）、或统计涨落（a posteriori 2.7σ）。Goal: are the bright outer-NW fields sky structure, SWCX, or instrumental? — E014 answered: none of these. The remaining candidates are residual soft-proton/instrumental (SP ratio 3.1 but weak correlation, Spearman +0.35, p=0.22), structure below 15′ (under the RASS aperture), or a statistical fluctuation (a posteriori 2.7σ).
• 输入：E014 的 rass_per_field.csv、omni_per_field.csv、nuisance_per_field.csv（newangels3_discussion/E014_/）。Inputs: E014's rass_per_field.csv, omni_per_field.csv, nuisance_per_field.csv (newangels3_discussion/E014_/).
• 方法：核对三个数——RASS R45 外四/其余十比值 0.91 ± 0.12 vs XMM 的 1.28（排除大尺度天空）；R45 纬度斜率无 +15.6%/deg 梯度；OMNI 无外 NW 窗口超 4e8 cm⁻² s⁻¹（排除 SWCX）。然后做 D49 排队的检验 (a)：对 0201/0301/0501/0601 加同期对照（0701/0801/1501/1601），在 lang 上用现有谱、把软质子折断幂律指数自由化重拟（不重处理数据）。Method: check three numbers — the RASS R45 outer-four/other-ten ratio 0.91 ± 0.12 vs the XMM 1.28 (large-scale sky excluded); no +15.6%/deg latitude slope in R45; no outer-NW OMNI window above 4e8 cm⁻² s⁻¹ (SWCX excluded). Then run D49's queued test (a): refit 0201/0301/0501/0601 plus the same-epoch controls (0701/0801/1501/1601) from the existing spectra on lang with the soft-proton broken power-law indices freed (no reprocessing).
• 成功判据：重拟后若四场的 CGMsum 流量显著回落（与对照场持平），软质子解释成立；若流量不变，则检验 (b)（Virgo ODF 级更严格耀斑过滤后重处理）成为下一步。Success/failure: if after the refit the four fields' CGMsum fluxes drop to the controls' level, the soft-proton explanation holds; if they do not move, queued test (b) (reprocess on Virgo with a stricter ODF-level flare filter) is next.
• 时间：读结果半天；重拟约 1–2 天（8 个 ObsID，每个数分钟到更久）。Time: half a day to read the results; about 1–2 days for the refit (8 ObsIDs, minutes or more each).
• 先读：cognition-os/experiments/E014_nw_gradient_rass_omni_crosscheck.md（「Result after luna correction 1」节）；cognition-os/04_DECISION_LOG.md D49。Read first: cognition-os/experiments/E014_nw_gradient_rass_omni_crosscheck.md (the "Result after luna correction 1" section); cognition-os/04_DECISION_LOG.md D49.

步骤 4：设计 300 kpc RASS 图（O4）Step 4: design the 300-kpc RASS map (O4)

• 目标：本地（M31 方向）MW 前景模型 + Local Hot Bridge 检验的设计文档。Goal: a design document for a local (M31-direction) MW foreground model and a Local Hot Bridge test.
• 输入：/home/huangrui/Data/M31_RASS/data/hp_maps/（nside-256 RASS 带；用原生 counts+exposure，修正图只能作次级列——B026）。Inputs: /home/huangrui/Data/M31_RASS/data/hp_maps/ (nside-256 RASS bands; use native counts+exposure — the corrected maps are a secondary column only, per B026).
• 方法：Poisson 前向模型 rate = c0(1+s(b−b_M31))，加扫描/瓦片噪声层；掩掉 M32/NGC 205 与亮源。Method: a Poisson forward model rate = c0(1+s(b−b_M31)) with scan/tile nuisance layers; mask M32/NGC 205 and bright sources.
• 成功判据：设计能回答「MW 在 M31 方向的宽带亮度是多少、Local Hot Bridge 需要多少」。Success/failure: the design can answer "what is the MW broadband brightness toward M31, and how much would a Local Hot Bridge need".
• 时间：设计 1 天，执行视数据而定。Time: 1 day for the design; execution depends on data.
• 先读：cognition-os/insights/I030 与 I031（RASS 前向模型要点）。Read first: cognition-os/insights/I030 and I031 (the RASS forward-model points).

步骤 5：把三条科学线写进论文讨论Step 5: write the three science lines into the paper discussion

• 目标：New-ANGELS III 草稿的讨论部分覆盖 (i) 总约束与 Locatelli 紧张、(ii) 轮廓形状限制、(iii) NW 起源。Goal: the New-ANGELS III draft discussion covers (i) the total constraint and the Locatelli tension, (ii) the profile-shape limit, (iii) the NW origin.
• 输入：E013/E014 结果 + I017/I033 + 冻结测量表。Inputs: the E013/E014 results, I017/I033, and the frozen measurement tables.
• 方法：直接总和优先（D30/D34）；条件场景全部带先验标签；提交门槛按 I013。Method: direct-sum first (D30/D34); every conditional scenario carries its prior label; submission gates stay those of I013.
• 成功判据：每句话都能映射到 03_TRUST_TABLE 的一行；无未标注的建议。Success/failure: every sentence maps to a row of 03_TRUST_TABLE; no unlabelled suggestion.
• 时间：1–2 周。Time: 1–2 weeks.
• 先读：paper_apj_v19_cgmsum_draft/README_draft_status_and_submission_gaps.md。Read first: paper_apj_v19_cgmsum_draft/README_draft_status_and_submission_gaps.md.

来源：步骤 1–2 依据 E013 BRIEF 与 cognition-os/experiments/E013；步骤 3 依据 cognition-os/experiments/E014（「Result after luna correction 1」节）、cognition-os/04_DECISION_LOG.md D49 与 I033 第 5–6 条；步骤 4 依据 02_CURRENT_QUESTION（2026-10-14 块）与 I030/I031；步骤 5 依据 D48 与 I013。所有「时间」为建议（未经评审）。

Source: steps 1–2 from the E013 BRIEF and cognition-os/experiments/E013; step 3 from cognition-os/experiments/E014 (the "Result after luna correction 1" section), cognition-os/04_DECISION_LOG.md D49, and I033 items 5–6; step 4 from 02_CURRENT_QUESTION (2026-10-14 block) and I030/I031; step 5 from D48 and I013. All "time" entries are a Suggestion (not reviewed).