动手指南：不用 AI 继续Handbook: Continue Without AI

机器与角色Machines & Roles

Lang3 资源门Lang3 Resource Gate：任何计算启动前读Before any compute, read .hermes/research-orchestration/20260922_parallel_theory/runtime/resource_state.json，只有only if "status": "READY" 且and sampled_at_epoch 距今within < 60 秒才启动；否则等一会重读，不许绕过。seconds then start; else wait and re-read, no bypass.

临时文件规则（用户 2026-10-13 定）Temporary File Rule (user 2026-10-13)：任何主机都不用 /tmp、/var/tmp 或 $HOME。Lang3 用No host uses /tmp, /var/tmp or $HOME. Lang3 uses $TMPDIR (/u/Lang2/ai_tmp/codex/v2_<T>/)，virgo 用 AI_tmp。用完就删（删前 lsof 确认没有进程在用、结果已另存）。代码里不许硬编码 /tmp。virgo uses AI_tmp. Delete after use (lsof first, results saved elsewhere). Code must not hard-code /tmp.

文件都在哪（仓库根 /u/Lang3/program/DET_ML_Uncertainty）Where Files Are (repo root /u/Lang3/program/DET_ML_Uncertainty)

常用命令（除第 5 节的 phase-1 测试外，每条都对照过源码）Common Commands (Each Verified Against Source, except the phase-1 test in §5)

0. PythonPython

用项目约定的 conda Python（路径见 WORKER_RULES.md：“Python: /home/<用户>/software/miniconda3/bin/python”，带 numpy/scipy/pandas/pyarrow）。计算类命令都加环境变量和 nice：Use the project's conda Python (see WORKER_RULES.md: "Python: /home/<user>/software/miniconda3/bin/python", with numpy/scipy/pandas/pyarrow). All compute commands get env vars and nice:

export OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1
nice -n 10 /home/<用户>/software/miniconda3/bin/python <脚本>

1. 快速自检（几秒）Quick Self-Check (seconds)

来源：quick_test_phase0.py（10 张图的小模拟，末尾打印 "Quick test PASSED"）：Source: quick_test_phase0.py (10-image toy sim, prints "Quick test PASSED" at end):

cd /u/Lang3/program/DET_ML_Uncertainty
python quick_test_phase0.py

2. 跑一个模拟Run a Simulation

来源：scripts/run_phase0_sims.py（遍历 config.py 的 S_GRID × B_GRID，每 (S,b) 做 N_SIM=10000 次，写 results/phase0_sims.npz）：Source: scripts/run_phase0_sims.py (iterates config.py's S_GRID × B_GRID, N_SIM=10000 per (S,b), writes results/phase0_sims.npz):

cd /u/Lang3/program/DET_ML_Uncertainty/scripts
python run_phase0_sims.py

带位置拟合的版本（3 参数，写 results/phase1_sims.npz）：With position fit (3 params, writes results/phase1_sims.npz):

cd /u/Lang3/program/DET_ML_Uncertainty
python scripts/run_phase1_sims.py

两个脚本的启动目录不同，别搞混（2026-10-17 核对源码）：run_phase0_sims.py 第 9 行是 sys.path.insert(0, "..")，".." 按进程当前目录解析，所以必须从 scripts/ 里跑；从仓库根跑会 ModuleNotFoundError: No module named 'config'（已复现）。run_phase1_sims.py 第 12 行用 Path(__file__).parent.parent（绝对路径），从任何目录跑都可以。两者都用相对路径写 results/，所以在各自正确的目录下运行即可。The two runners need different start directories — do not mix them up (source-checked 2026-10-17): run_phase0_sims.py line 9 is sys.path.insert(0, ".."), and ".." resolves against the process cwd, so it must be run from inside scripts/; run from the repo root it raises ModuleNotFoundError: No module named 'config' (reproduced). run_phase1_sims.py line 12 uses Path(__file__).parent.parent (absolute), so it runs from any directory. Both write results/ by a relative path, so run each from its own correct directory.

3. 跑基准计分器（v1，冻结版）Run the Benchmark Scorer (v1, frozen)

来源：scripts/run_benchmark_deltac.py 的 argparse（第 190–214 行确认）：Source: scripts/run_benchmark_deltac.py argparse (confirmed lines 190–214):

python scripts/run_benchmark_deltac.py --dataset catalog --config M2_band4 \
    --width-mode new --mu-min 8 --layout exact --label my_check

可选数据集：python（自研拟合器模拟）、emldetect（SAS 模拟）、catalog（4XMM DR14）。输出到 results/ 和 figures/。v2 计分器在 scratchpad/v2/T5/scorer_v2*.py（Claude 审查后才会移入 scripts/）。Optional datasets: python (custom fitter sim), emldetect (SAS sim), catalog (4XMM DR14). Output to results/ and figures/. v2 scorer at scratchpad/v2/T5/scorer_v2*.py (moves to scripts/ after Claude review).

二维损失模型（如 v2_T3_joint_2D）需要 L_2d 列Two-dimensional loss models (e.g. v2_T3_joint_2D) need an L_2d column，标准脚本没有该列。用 T11 的副本（带 --l2d-file）：the standard script lacks. Use the T11 copy (with --l2d-file):

python scratchpad/v2/T11/run_benchmark_deltac_T11_L2d.py --dataset catalog \
    --config M2_band4 --width-mode new --mu-min 8 --layout exact --iso 60 \
    --model benchmark/models/v2_T3_joint_2D.py \
    --l2d-file results/v2/T11/t11_dr14_L2d_M2_band4.parquet --label my_check

或 v2 计分器（带 --extra-column）：Or the v2 scorer (with --extra-column):

python scratchpad/v2/T5/pareto_submit_v2.py \
    --id MyEntry_v1 --layer C-iso \
    --model benchmark/models/v2_T3_joint_2D.py \
    --title "一句话" --description "一段话" \
    --hypothesis "它检验什么物理想法" \
    --note cognition-os/experiments/Exxx_mine.md \
    --fit-data "none" --params 0 \
    --extra-column L2d=results/v2/T11/t11_dr14_L2d_M2_band4.parquet:L2d \
    --salt-file <PI 给你的盐文件路径>

来源：scratchpad/v2/T11/run_benchmark_deltac_T11_L2d.py 的 argparse（--l2d-file 选项）；scratchpad/v2/T5/pareto_submit_v2.py 的 argparse（--extra-column）。每条命令都已用 --help 核对；二维损失命令已干跑通过（C-iso M2 b4，A1 0.156，与登记值一致）。Source: scratchpad/v2/T11/run_benchmark_deltac_T11_L2d.py argparse (--l2d-file); scratchpad/v2/T5/pareto_submit_v2.py argparse (--extra-column). Every command checked with --help; two-dimensional loss command dry-ran (C-iso M2 b4, A1 0.156, matches registered value).

4. 画一张图Plot a Figure

本站的 14 张图全部产自：All 14 figures on this site come from:

python scratchpad/v2/T12/make_figures.py

（先查资源门 READY；单线程，几分钟。）想画自己的图，仿照它读 results/ 里的 CSV/parquet。(Check resource gate READY first; single-thread, a few minutes.) To plot your own, mimic it reading CSV/parquet from results/.

5. 跑测试Run Tests

cd /u/Lang3/program/DET_ML_Uncertainty
python tests/test_phase0.py
python tests/test_phase1.py

两个测试都从仓库根跑（它们用绝对路径找 lib/，但用相对路径读 results/）。test_phase0.py 读 results/phase0_sims.npz——就是上面第 2 节那个脚本写的文件，链条是通的。Run both tests from the repo root (they find lib/ by absolute path but read results/ by a relative path). test_phase0.py reads results/phase0_sims.npz — the file the runner in §2 above writes, so that chain is intact.

登记一个条目Register an Entry

1. 先写实验笔记Write experiment note first：cognition-os/experiments/E<编号>_<slug>.md，记录问题、协议（跑之前冻结并记哈希）、方法。cognition-os/experiments/E<number>_<slug>.md, record question, protocol (freeze and hash before running), method.
2. 写模型插件Write model plugin：拷贝 benchmark/models/ref_V4_delta_newwidth.py 改。插件必须定义 N_PARAMS、N_PARAMS_FULL_GRID、THEORY 和 predict(obs, kern, ctx) -> (mean, sd)。obs 只含可观测量（S=Ŝ、b、position_id、MASKFRAC、mu0）。Copy benchmark/models/ref_V4_delta_newwidth.py and modify. Plugin must define N_PARAMS, N_PARAMS_FULL_GRID, THEORY, and predict(obs, kern, ctx) -> (mean, sd). obs contains only observables (S=Ŝ, b, position_id, MASKFRAC, mu0).
3. 提交Submit（来源：scratchpad/v2/T5/pareto_submit_v2.py 的 argparse）： (Source: scratchpad/v2/T5/pareto_submit_v2.py argparse):

   python scratchpad/v2/T5/pareto_submit_v2.py \
       --id MyEntry_v1 --layer C-iso \
       --model benchmark/models/MyEntry.py \
       --title "一句话" --description "一段话" \
       --hypothesis "它检验什么物理想法" \
       --note cognition-os/experiments/Exxx_mine.md \
       --fit-data "none" --params 0 \
       --salt-file <PI 给你的盐文件路径>

   参数计数必须诚实：迁移来的拟合数也算；在 catalog 行上拟合过的加 --in-sample。若模型需要额外可观测量（如 L_2d），用 --extra-column file:column。私有分数由 PI 在里程碑时通过计分器驱动开启；学生永远不做。Parameter count must be honest: transferred fit counts count too; if fitted on catalogue rows add --in-sample. If model needs extra observables (e.g. L_2d), use --extra-column file:column. Private scores are opened by PI via scorer at milestones; students never do this.
4. 条目不可变Entry is immutable：已存在的 id 永不覆盖；改了模型就换新 id。学生不执行登记——把插件和笔记准备好，交给 PI（或 PI 指定的人）执行；salt 文件由 PI 保管。Existing id is never overwritten; change model = new id. Students do not execute registration — prepare plugin and note, hand to PI (or PI's designee) to run; salt file held by PI.

必须遵守的规则Rules You Must Follow

1. 铁律（IRON RULE）Iron Rule：条件于观测量 + 声明的先验 π；绝不按 ML/DET_ML 切样本；统计在 ΔC 空间做；S_true 只能通过 π 积分进入。（全文：CLAUDE.md 和 PARETO_BENCHMARK_V2.md §1。）Condition on observables + declared prior π; never slice by ML/DET_ML; statistics in ΔC space; S_true enters only via π integral. (Full text: CLAUDE.md and PARETO_BENCHMARK_V2.md §1.)
2. 目录约定陷阱Catalogue Convention Traps（CLAUDE.md “Catalog Convention Traps I041”）： (CLAUDE.md "Catalog Convention Traps I041"):(a) DET_ML 用自然对数；(b) Q 的第二参数是 ΔC/2；(c) 目录 delta_c_nu3 列存 ΔC/2；(d) 因子化代理参数不能当物理量用；(e) κ≈3 只是 1 参数代理诊断；(f) ecut 的默认值分模式（standalone 0.68；DR14 的 CUTRAD=60″ 相当于 15）。任何目录分析先过“验证清单”。(a) DET_ML uses natural log; (b) Q's second argument is ΔC/2; (c) catalogue delta_c_nu3 column stores ΔC/2; (d) factorised proxy params are not physical quantities; (e) κ≈3 is only a 1-param proxy diagnostic; (f) ecut default varies by mode (standalone 0.68; DR14 CUTRAD=60″ ≡ 15). Any catalogue analysis must pass the verification checklist first.
3. 不许用 /tmpNo /tmp（任何主机）；临时文件用上面列的目录并随用随清。(any host); temp files use the dirs above and clean as you go.
4. CPU 上限CPU Cap：Lang3 整机 ≤ 60%；virgo 整机 ≤ 60% 且启动批次前跑守卫。Lang3 whole-machine ≤ 60%; virgo whole-machine ≤ 60% and run guard before each batch.
5. 记录每个来源Record Every Source：每个数字标来源文件；每个实验先写笔记再跑；协议先冻结（记 sha256）再看数据。Every number cites its source file; every experiment writes note before running; protocol frozen (sha256 recorded) before looking at data.
6. 不改冻结的东西Don't Modify Frozen Things：scripts/、lib/、benchmark/entries/、source_catalog/、已有 results/ 文件。要改就拷进自己的 scratchpad。scripts/, lib/, benchmark/entries/, source_catalog/, existing results/ files. If you must change, copy into your own scratchpad.
7. 负结果也登记Register Negative Results Too：被支配的条目、失败的假设都是信息。Dominated entries, failed hypotheses are information.
8. 不 commit、不部署No Commits, No Deploys：git 提交和网站部署由 PI 做。git commits and site deploys are done by PI.

下一步计划（编号）Next Steps (Numbered)

依据：cognition-os/02_CURRENT_QUESTION.md（Q1–Q7，2026-10-17T15:41Z；PI 2026-10-17 07:10Z 把模拟宽度 Q7 定为主攻，目录问题降为 bonus）、08_RUNNING.md rev 33（§3 在跑、§5 可派发）、07_REVISION_LOG.md R001–R039、03_TRUST_TABLE.md T209–T229、reviews/T13_001–008、RT-22 到 RT-42 的分诊、04_DECISION_LOG.md 2026-10-17 12:57 前的条目。每步给：目标、输入、方法、成败判据、时间估计、先读什么。Based on: cognition-os/02_CURRENT_QUESTION.md (Q1–Q7, 2026-10-17T15:41Z; the PI made the simulation width Q7 the main line on 2026-10-17 07:10Z and demoted the catalogue questions to bonus), 08_RUNNING.md revision 33 (§3 in flight, §5 ready to dispatch), 07_REVISION_LOG.md R001–R039, 03_TRUST_TABLE.md T209–T229, reviews T13_001–008, the RT-22 through RT-42 triages, and 04_DECISION_LOG.md entries to 2026-10-17 12:57. Each step has: goal, inputs, method, success/failure test, time estimate, what to read first.

第 1 步 · T11 第 21 阶段的判定（已完成，结论：ℓ_impl ≈ 0.8·(1−w)/μ）Step 1 · T11 Stage 21 Adjudication (Done, Conclusion: ℓ_impl ≈ 0.8·(1−w)/μ)

• 目标：Goal: 按 T11_020 评审的冻结判读，判定 Ŝ 条件 PN 总体 MC 的结果。Per T11_020 review's frozen adjudication, adjudicate the Ŝ-conditioned PN population MC result.
• 结果：Result: ℓ_impl ≈ 0.8·(1−w)/μ（接近而非吻合，比值 0.78–0.87）；L_2d/(1−w) 对所有相机恒等于 2，相机分化来自数据；Ŝ 条件下 PN 的 r_MC ≈ 0，第 19 阶段矛盾解决。候选假说 ΔC ≈ (1−m_cam)(μ+δ) − (1−w)（未确立）。（来源：reviews/T11_021_claude.md。）ℓ_impl ≈ 0.8·(1−w)/μ (closer to, not matches; ratio 0.78–0.87); L_2d/(1−w) = 2 for every camera, so the camera split lies in the data; Ŝ-conditioned r_MC ≈ 0 on PN, closing the stage-19 tension. Candidate hypothesis ΔC ≈ (1−m_cam)(μ+δ) − (1−w) (not established). (Source: reviews/T11_021_claude.md.)

第 2 步 · 守护 L0 生产到完成并正式计分（进行中，最后核实 163/251，2026-10-16T17:03Z）Step 2 · Guard L0 Production to Completion and Formal Scoring (In Progress, Last Verified 163/251, 2026-10-16T17:03Z)

• 目标：Goal: L0 的 251 块全部完成，取回验证，用 v2 计分器在全统计量下正式登记 L0 条目。All 251 L0 blocks complete, retrieved and verified, formally register L0 entries with v2 scorer on full statistics.
• 输入：Inputs: virgo 上的生产/局部补算监督进程（按 PID 文件管理）；results/v2/T1/production/（已有 14 块 + prefix_manifest.json）；scratchpad/v2/T5/scorer_v2_s6.py 的摄取模式。virgo production/local-patch supervision processes (managed by PID file); results/v2/T1/production/ (14 blocks + prefix_manifest.json); scratchpad/v2/T5/scorer_v2_s6.py ingestion mode.
• 方法：Method: 每天看一次日志与块计数；完成后逐块 sha256 验证、拉回 Lang3、写清单；然后用 T5 的计分流程在局部约定下跑全部插件（先演练到 scratch，交给 PI 正式登记）。Check logs and block counts daily; after completion verify each block sha256, pull back to Lang3, write manifest; then run T5 scoring flow on all plugins under local convention (dry-run to scratch first, hand to PI for formal registration).
• 成败判据：Success/Failure Test: 验收来自 T1 卡：位元一致重跑 1% 子集；固定位置 1 参数 z-sd 在 μ ≥ 20 时与 1 差 < 1%；失败率 ≤ 0.1%。From T1 card: bitwise-identical re-run of 1% subset; fixed-position 1-param z-sd within 1% of 1 at μ ≥ 20; failure rate ≤ 0.1%.
• 时间：Time: 守护 1–2 周（生产预计 2026-10-20 完）；计分与演练 2–3 天。Guard 1–2 weeks (production ETA 2026-10-20); scoring and dry-run 2–3 days.
• 先读：Read First: cognition-os/experiments/E152_certified_fitter_L0.md（设计与证书）→ 检查点 T1_004.md → prefix_manifest.json。cognition-os/experiments/E152_certified_fitter_L0.md (design & certificate) → checkpoint T1_004.md → prefix_manifest.json.

第 3 步 · 同样守护 L1 生产（进行中，最后核实 13,414 个 parquet、近 2 小时 388 个，2026-10-16T17:05Z）Step 3 · Likewise Guard L1 Production (In Progress, Last Verified 13,414 Parquets, 388 in the Last 2 h, 2026-10-16T17:05Z)

• 目标：Goal: 16,200 幅图像的 emldetect–理想拟合器逐行配对表完成，量化工程差异 ΔC_eml − ΔC_ideal。16,200 images emldetect–ideal fitter row-by-row pairing table complete, quantify engineering gap ΔC_eml − ΔC_ideal.
• 输入：Inputs: virgo 的 4 个 L1 worker；results/v2/T2/ 已有试点。virgo's 4 L1 workers; results/v2/T2/ has pilot.
• 方法：Method: 守护 + 完成后按 T2 卡验收：每行 emldetect 都被匹配或解释；认证拟合器在这些图像上与 L0 一致（同 PSF、同背景处）。Guard + on completion verify per T2 card: every emldetect row matched or explained; certified fitter on these images agrees with L0 (same PSF, same background handling).
• 成败判据：Success/Failure Test: 见 T2 卡；试点已确立关键事实（emldetect 跟随局部最优：S=2.5 时 eml−局部位置差 p95 0.39 px vs 全局 15 px）。See T2 card; pilot established key fact (emldetect follows local optimum: at S=2.5 eml–local position diff p95 0.39 px vs global 15 px).
• 时间：Time: 守护约 6 天 + 分析 2 天。Guard ~6 days + analysis 2 days.
• 先读：Read First: cognition-os/experiments/E153_L1_paired_staging.md → 检查点 T2_005.md。cognition-os/experiments/E153_L1_paired_staging.md → checkpoint T2_005.md.

第 4 步 · T11 第 20 阶段：二维模型的窗口平坦性检验（已完成，结论：不是闭合项）Step 4 · T11 Stage 20: Two-Dimensional Model Window Flatness Test (Done, Conclusion: Not a Closing Term)

• 目标：Goal: 二维损失模型必须在 6–10、10–20、>20 三个窗口里都平坦，才算真正闭合（catalog 的亏缺在三个窗口里都是 −0.046 到 −0.048 的平坦比例）。The two-dimensional loss model must be flat across all three windows (6–10, 10–20, >20) to be a true closing term (catalogue deficit is flat at −0.046 to −0.048 across all three).
• 输入：Inputs: results/v2/T11/ 的分箱表；T4_007 的三向判读规则作对照。results/v2/T11/ binned tables; T4_007 three-way adjudication rules for comparison.
• 方法：Method: 对每个 MOS 配置算三个窗口的 frac 和 A1；对照 catalog 的平坦值与 M2 b4 的 r 斜率（−0.033）。For each MOS config compute frac and A1 in three windows; compare against catalogue flat value and M2 b4 r slope (−0.033).
• 成败判据：Success/Failure Test: 预先声明（写进实验笔记再算）：三窗口 frac 与 catalog 各差 ≤ 0.01 且斜率同号。Pre-declare (write in experiment note before computing): three-window frac each within 0.01 of catalogue and slope same sign.
• 时间：Time: 1 天。1 day.
• 先读：Read First: 04_DECISION_LOG.md 2026-10-12 06:40 条目（T4_007 的 NOT MATCHED 结论和排队中的第 20 阶段）。04_DECISION_LOG.md 2026-10-12 06:40 entry (T4_007 NOT MATCHED conclusion and queued stage 20).

第 5 步 · 攻宽度缺口（C-iso A2 0.08–0.11 vs 0.05）Step 5 · Attack Width Gap (C-iso A2 0.08–0.11 vs 0.05)

• 目标：Goal: 把零参数条目的 A2 往下压。Push zero-parameter entries' A2 down.
• 输入：Inputs: E147 的壳层×相位混合网格（宽度已证明有效，A2 0.080）；E141 的高背景捕获核；E107 系列的局部 Edgeworth 展开。E147 shell×phase mixture grid (width proven effective, A2 0.080); E141 high-background capture kernel; E107 series local Edgeworth expansion.
• 方法：Method: 以 E147mix 宽度为起点，在 L0（第 2 步的产物）上检验哪个零参数宽度律在所有 α 下都过噪声底；不要在 catalog 上调——在 catalog 上拟合的数都算参数并使条目变样本内。Start from E147mix width, test on L0 (Step 2 product) which zero-parameter width law passes noise floor for all α; do NOT tune on catalogue — numbers fitted on catalogue count as parameters and make the entry in-sample.
• 成败判据：Success/Failure Test: 配对两级区间相对当前前沿条目排除 0。Paired two-level interval relative to current front entry excludes 0.
• 时间：Time: 1–2 周。1–2 weeks.
• 先读：Read First: cognition-os/experiments/E147_catalog_shell_phase_mixture.md → E141_capture_kernel_high_background.md → PARETO_BENCHMARK_V2.md §9.4。cognition-os/experiments/E147_catalog_shell_phase_mixture.md → E141_capture_kernel_high_background.md → PARETO_BENCHMARK_V2.md §9.4.

第 6 步 · PN 的损失形式（视第 1 步结果）Step 6 · PN Loss Form (Depends on Step 1 Result)

• 目标：Goal: 根据第 1 步的判定，确定 PN 用的损失形式（(1−w) 还是 L_2d 还是都不是），构造零参数修正。Based on Step 1 adjudication, determine PN's loss form ((1−w) or L_2d or neither), construct zero-parameter correction.
• 输入：Inputs: 第 1 步的 MC 结果（100000 实现/波段）；E162 第 15–20 阶段的排除清单。Step 1 MC results (100000 realizations/band); E162 stages 15–20 exclusion list.
• 方法：Method: 把定案的损失做成 PN 的零参数核，重新计分 C-iso PN。注意评审 T11_014 的决定：不许把 PN 项“打补丁”进已登记的相机统一模型——PN 的失败要如实公布。Make the adjudicated loss a zero-parameter PN kernel, re-score C-iso PN. Note T11_014 review decision: do not "patch" the PN term into the already-registered camera-unified model — PN's failure must be reported honestly.
• 成败判据：Success/Failure Test: C-iso PN 的 A1 降到与 MOS 相当（< 0.2），且不引入任何拟合数。C-iso PN A1 drops to MOS-like (< 0.2) with zero fitted parameters.
• 时间：Time: 1–2 周。1–2 weeks.
• 先读：Read First: cognition-os/experiments/E162_joint_fit_catalog_transfer.md 第 15–20 阶段 → reviews/T11_014_claude.md。cognition-os/experiments/E162_joint_fit_catalog_transfer.md stages 15–20 → reviews/T11_014_claude.md.

第 7 步 · T5 阶段 12：高计数模型失效与真实数据乘性亏损（已完成，T5_009 评审通过）Step 7 · T5 Stage 12: High-Count Model Failure & Real-Data Multiplicative Deficit (Done, T5_009 Review PASS)

• 结果：Result: L0 对照 m_L0 = 0.0038（不截断主读数，RT-13；0.0077 是 ML<690 截断值，只作敏感性），目录 m 减 L0 m = 0.058 [0.056, 0.061]——乘性亏损属于真实数据，已接受；核高 S 失效确诊为 E114 网格范围（只覆盖 S = 1–400，之上线性外推；S_kernel ≈ 673 时 frac = +0.21，1122 时 +0.65）；C-all 上 m 无环/背景依赖，共同行上 C-all 与 C-iso 相同（更正 T5_008 评审第 4 条）；T3_joint_L07 的 a 随背景从 +0.33 变到 −0.31。m、a 已显示在高计数报告页（用户决定，决策日志 20:45）。（来源：reviews/T5_009_claude.md。）L0 control m_L0 = 0.0038 (the uncut primary reading, RT-13; 0.0077 is the ML < 690 cut value, a sensitivity only), catalogue m minus L0 m = 0.058 [0.056, 0.061] — the multiplicative deficit belongs to the real data, accepted; the kernel high-S failure is diagnosed as the E114 grid range (covers only S = 1–400, extrapolated linearly above; frac = +0.21 at S_kernel ≈ 673, +0.65 at 1122); on C-all m has no ring/background dependence, and C-all equals C-iso on common rows (correcting T5_008 review item 4); T3_joint_L07's a drifts from +0.33 to −0.31 with background. m and a are shown on the high-count report page (user decision, decision log 20:45). (Source: reviews/T5_009_claude.md.)

第 7a 步 · T5 阶段 13–14：扩展核网格（已完成，T5_010/T5_011 评审通过）Step 7a · T5 Stages 13–14: Extend the Kernel Grid (Done, T5_010/T5_011 Reviews PASS)

• 目标：Goal: 找到 E114 网格的生成方式，向上扩展 S（新版本化文件），用新模拟验证，6–10 分数逐位复现。Find how the E114 grid is generated, extend it upward in S (new versioned file), verify against fresh simulations, reproduce 6–10 scores exactly.
• 结果：Result: 全部 15 个网格（M1/M2/PN × b1–b5）已扩展到 S = 5000（E116 生成器，只改 S 轴；旧节点逐一复现，差值 0.0）；新模拟在 S_true = 300、600、1000、2000、4000 时 |frac| < 0.01；已登记的 6–10 成绩复现，差值 0.0；扩展网格设为高计数报告默认核，前沿不变。剩约 0.6% 乘性底线（L0 m = 0.004–0.008）。T11 的星表总体从未超过 S = 300（加载器截断），T11 的 m_cam 结论不受影响（T5_011 评审更正 T5_010 评审的担忧）。All 15 grids (M1/M2/PN × b1–b5) extended to S = 5000 (E116 generator, only the S axis changed; every old node reproduces to 0.0); fresh simulations give |frac| < 0.01 at S_true = 300, 600, 1000, 2000, 4000; registered 6–10 scores reproduce to 0.0; the extended grids are the default kernel for high-count reports, the front is unchanged. A ~0.6% multiplicative floor remains (L0 m = 0.004–0.008). T11's catalogue populations never exceed S = 300 (loader cap), so T11's m_cam conclusions are unaffected (T5_011 review corrects T5_010 review's concern).

第 7b 步 · T11 阶段 22–23：m_cam 假说检验（已完成，T11_023 评审通过）Step 7b · T11 Stages 22–23: The m_cam Hypothesis Test (Done, T11_023 Review PASS)

• 目标：Goal: 检验候选假说 ΔC ≈ (1−m_cam)(μ+δ) − (1−w)：用高计数冻结的 m_cam 预测低窗口；再按星表列拆分 m。Test the candidate hypothesis ΔC ≈ (1−m_cam)(μ+δ) − (1−w): predict the low windows with a high-count-frozen m_cam, then split m by catalogue columns.
• 阶段 22 结果：Stage 22 Result: 未闭合：12 个配置只闭合 4 个；b3/b4 低窗口残差为正（亏损随 S 增大）；m 随波段增大（b4−b1 ≈ +0.034），PN b1 为负——"每相机一个常数"假说被否定，原因开放。M2 b4 的 m_cam = 0.0659 与 T5 的 C-all m = 0.066 独立一致。（来源：reviews/T11_022_claude.md。）Not closed: only 4 of 12 configs closed; b3/b4 low-window residuals positive (deficit grows with S); m grows with band (b4−b1 ≈ +0.034), PN b1 negative — the "one constant per camera" hypothesis is refuted, cause open. M2 b4 m_cam = 0.0659 agrees with T5's C-all m = 0.066 independently. (Source: reviews/T11_022_claude.md.)
• 阶段 23 结果：Stage 23 Result: M2 b4 上 ℓ_impl = 0.1124 [0.1111, 0.1136]；ℓ_impl / ((1−w)/μ + m) = 1.02 [1.01, 1.04]（m 取 MOS b4 均值 0.0584），换用 M2 b4 自己的 m_cam = 0.0659 则为 0.96；ℓ_impl / (L_2d/μ) = 1.10；ℓ_impl / ((1−w)/μ) = 2.19——MOS 也是"一份 (1−w) 联合损失加真实数据的乘性 m"，假说在原本看起来最像二维形式的相机上得到支持；拟合器项 r_MC = +0.017 已记录。拆分：全部 12 个配置的 m 都随观测年代下降；MASKFRAC 与年代拆分按固定规则在 M1 和 PN 上"解释" m，M2 用哪种拆分都解释不了。更正：MASKFRAC 是未遮挡比例（中位数 0.981），两组只差约 0.02，不足以直接造成 0.05 的 m 差异，更可能是年代或探测器位置的代理变量；EP_EXTENT 和 SUM_FLAG 因非零组太小未能检验；m(μ) 有 3 个配置的点估计落在自己的 bootstrap 区间之外，待第 24 阶段查明。（来源：reviews/T11_023_claude.md。）On M2 b4, ℓ_impl = 0.1124 [0.1111, 0.1136]; ℓ_impl / ((1−w)/μ + m) = 1.02 [1.01, 1.04] with the MOS b4-mean m 0.0584, or 0.96 with M2 b4's own m_cam = 0.0659; ℓ_impl / (L_2d/μ) = 1.10; ℓ_impl / ((1−w)/μ) = 2.19 — MOS too is "one joint loss (1−w) plus the real-data multiplicative m", the hypothesis is supported on the camera where the two-dimensional form looked right; the fitter-level term r_MC = +0.017 is recorded. Splits: m falls with epoch on all 12 configs; the MASKFRAC and epoch splits "explain" m on M1 and PN by the fixed rule, M2 resists every split. Correction: MASKFRAC is the unmasked fraction (median 0.981), the two groups differ by only ~0.02, too small to produce a 0.05 difference in m directly, probably a proxy for epoch or detector position; EP_EXTENT and SUM_FLAG could not be tested (non-zero groups too small); for m(μ), 3 configs have point estimates outside their own bootstrap intervals, to be explained by stage 24. (Source: reviews/T11_023_claude.md.)

第 7c 步 · T11 阶段 24–26：分离遮挡与年代（已完成，T11_024–026 评审通过；T11 暂停）Step 7c · T11 Stages 24–26: Separate Mask from Epoch (Done, T11_024–026 Reviews PASS; T11 Paused)

• 目标：Goal: 把 MASKFRAC 信号与年代信号分开；构造零参数的遮挡修正项；修正 m(μ) 的区间。Separate the MASKFRAC signal from the epoch signal; build a zero-parameter mask-aware arm; fix the m(μ) intervals.
• 结果：Result: 年代效应在 MASKFRAC 高低两半里都显著（12 个配置全部如此）；corr(MASKFRAC, OBS_ID) ≈ 0，两种拆分互相独立（更正 T11_023 评审的代理猜测）。遮挡修正项是上界（MOS 降约 0.03、PN 降约 0.047，修正后 PN 出现超出）；符号问题未解释（完全覆盖的源亏损反而更大）。m(μ) 的权重不一致 bug 已修好。T11_025：M1 在 2012-12（MOS1 CCD3 损失）处有阶跃 −0.016，M2/PN 是漂移（约 −0.001/年）；Ŝ 构成仅凭星表列无法检验。T11_026：阶跃不是覆盖效应（加入 MASKFRAC 后仍为 −0.0132 ± 0.0024）；星表协变量都解释不了漂移；星表列路线已走完，T11 暂停。（来源：reviews/T11_024/025/026_claude.md。）The epoch effect is significant in both mask halves (all 12 configs); corr(MASKFRAC, OBS_ID) ≈ 0, the two splits are independent (correcting the T11_023 review's proxy guess). The mask-aware arm is an upper bound (lowers m by ~0.03 on MOS, ~0.047 on PN; PN shows an excess after correction); the sign problem is unexplained (fully covered sources have the larger deficit). The m(μ) weight-inconsistency bug is fixed. T11_025: M1 has a step of −0.016 at 2012-12 (the MOS1 CCD3 loss), M2/PN are drifts (~−0.001/yr); Ŝ formation is not testable from the catalogue columns. T11_026: the step is not a coverage effect (−0.0132 ± 0.0024 with MASKFRAC included); no catalogued covariate explains the drift; the catalogue-column route is exhausted, T11 paused. (Sources: reviews/T11_024/025/026_claude.md.)

第 7d 步 · T13 真实图像本地重拟合（第 2–6 阶段已评审，第 7 阶段第 2 部分在跑）Step 7d · T13 Real-Image Local Refit (Stages 2–6 Reviewed, Stage-7 Part 2 Running)

• 目标：Goal: 用冻结的本地拟合器重拟合真实星表图像，直接检验流量归一化与 PSF 通道——这是星表列路线走完后剩下的决定性实验。Refit the real catalogue images with the frozen local fitter to test the flux-normalisation and PSF channel directly — the decisive experiment remaining after the catalogue-column route was exhausted.
• 输入：Inputs: 每个观测的计数图像、曝光图、背景图、探测掩膜（4XMM PPS 产品，或在 virgo 上用 SAS 从 ODF/PPS 事件清单重新生成）；M2 b4（加 M1 b4）的星表行；试点 3 个观测：早（2004 前）、中（2008–2011）、晚（2016 后），每个至少 5 个 M2 b4 的 ML > 50 的 C-iso 公开选择源。Per observation: the counts image, the exposure map, the background map, the detection mask (4XMM PPS products, or regenerated with SAS on virgo from ODF/PPS event lists); the M2 b4 (plus M1 b4) catalogue rows; pilot of 3 observations: early (before 2004), middle (2008–2011), late (after 2016), each with at least 5 M2 b4 ML > 50 sources in the C-iso public selection.
• 方法：Method: 用 T4/T11 的冻结本地拟合器（FineFamily Nelder–Mead 3 参数拟合，见 scratchpad/v2/T11/s22_popmc_worker.sh），在星表位置处用真实图像、真实背景图和探测掩膜做拟合；PSF 模板按源的探测器位置和波段能量生成（psfgen）。记录 Ŝ_local、ΔC_local 和拟合位置。健全性检验：ΔC_local 与 ΔC_cat 的秩应一致，Ŝ_local / Ŝ_cat 应为 O(1)。不许改拟合器。Use the T4/T11 frozen local fitter (FineFamily Nelder–Mead 3-parameter fit, see scratchpad/v2/T11/s22_popmc_worker.sh) at the catalogue positions, with the real image, the real background map and the detection mask; PSF template at the source's detector position and band energy (psfgen). Record Ŝ_local, ΔC_local and the fitted position. Sanity test: ΔC_local and ΔC_cat should agree in rank, and Ŝ_local / Ŝ_cat should be O(1). Do not change the fitter.
• 成败判据：Success/Failure Test: 三条事先固定的判读（prompts/T13.md）：(1) 流量归一化——median(Ŝ_local / Ŝ_cat) 按年，若随年代漂移且其大小通过核斜率解释 m 的大部，则原因是 (a)；(2) 亏损所在位置——若 m_local < 0.02 而同一批源的 m_cat ≈ 0.05–0.07，则亏损在星表的处理（a 或 c）；若 m_local ≈ m_cat，则亏损在真实 PSF（b）；(3) 时间——M1 的阶跃与 M2 的漂移能否在本地重拟合中复现。Three readings fixed in advance (prompts/T13.md): (1) flux normalisation — median(Ŝ_local / Ŝ_cat) per year; if it drifts with year and its size accounts for most of m through the kernel slope, the cause is (a); (2) where the deficit lives — if m_local < 0.02 while m_cat ≈ 0.05–0.07 on the same sources, the deficit is in the catalogue's processing (a or c); if m_local ≈ m_cat, the deficit is in the real PSF (b); (3) time — whether the local refit reproduces the catalogue's M1 step and M2 drift.
• 时间：Time: 试点约 1 天（3 个观测，virgo 上最多 4 个进程，≤ 5 GB）；第 2 阶段（按年分层约 8 个观测/年，2000–2023）的规模由试点估计，先写方案再跑。Pilot ~1 day (3 observations, at most 4 processes on virgo, ≤ 5 GB); stage 2 (~8 observations per year, 2000–2023, stratified by year) is sized from the pilot — plan first, run later.
• 先读：Read First: prompts/T13.md（三条固定判读与 virgo 限额）→ reviews/T13_001_claude.md → reviews/T11_026_claude.md。prompts/T13.md (the three fixed readings and the virgo limits) → reviews/T13_001_claude.md → reviews/T11_026_claude.md.
• 第 1 阶段结果（T13_001 评审通过）：Stage-1 Result (T13_001 Review PASS): 产品路线端到端打通：NXSA 的 PPS 产品（4 个产品 + OBLSLI）、EXPID 匹配、共享 WCS 全部验证（box-list 的 X_IMA = WCS 像素 + 1，误差 < 0.002 px）；冻结拟合器原样可用（只重定向 PSF31 路径）。3 个观测的健全性检验通过：S_local/S_cat 为 1.05–1.11；两个孤立场的 ΔC 排序相关 ρ 为 0.81–0.98；拥挤的 2001 场把 ρ 拉低到约 0.35（拟合移到邻近源，r_uv 长尾到 10 px）。M1 与 M2 的流量比配对差 +0.0006 [−0.007, +0.015]，区间含 0，持平。版本陷阱：档案现在提供的是 2024 年重处理产品（PPSVERS 21.51，SAS 21），而 DR14 用的是管线 18/19——0512180101 的图像计数约为星表值的 3.5 倍，即版本错配。（来源：reviews/T13_001_claude.md。）The product route verified end to end: NXSA's PPS products (4 products plus OBLSLI), EXPID matching and the shared WCS all checked (box-list X_IMA = WCS pixel + 1, to < 0.002 px); the frozen fitter unchanged (only the PSF31 path redirected). Sanity passes on 3 observations: S_local/S_cat is 1.05–1.11; ρ(ΔC) is 0.81–0.98 on the two isolated fields; the crowded 2001 field drags ρ to about 0.35 (fits move to a neighbour, r_uv tail to 10 px). The M1–M2 flux-ratio paired difference is +0.0006 [−0.007, +0.015], interval contains 0, no significant difference. Pipeline-version trap: the archive now serves the 2024 reprocessing (PPSVERS 21.51, SAS 21) while DR14 used pipeline 18/19 — 0512180101's image counts are about 3.5× the catalogue values, a version mismatch. (Source: reviews/T13_001_claude.md.)
• 第 2 阶段方案（如派发时所述）：Stage-2 Plan (as dispatched): 约 190 个观测（每年约 8 个）；下载已完成，psfgen 与拟合运行中。三个测量量分开判读：m_cat（DR14 数据、DR14 管线）、m_newcat（2024 图像、2024 管线自己的 ΔC，用稳定反解）、m_local（2024 图像、我们的冻结拟合器）。m_newcat 对 m_local（同一批图像）把星表的拟合与处理 (c) 从数据 (b) 分开；m_cat 对 m_newcat 量出重处理改变了什么——不必用 SAS 18 重新生成。筛选不能建立在读数对象上（评审更正）：用星表对星表的管线间筛选——2024 PPS 源表与 DR14 按位置匹配，比较 M2/M1 b4 的 RATE；所有读数都报全集和 |median(RATE_2024/RATE_DR14) − 1| < 0.10 子集两套。拥挤读数报 |r_uv| ≤ 2 px 前后两套。中间文件放 NVMe（试点误用了 HDD）。（来源：reviews/T13_001_claude.md。）About 190 observations (~8 per year); downloads done, psfgen and the fits running. Three measurements separate the readings: m_cat (DR14 data, DR14 pipeline), m_newcat (2024 images, the 2024 pipeline's own ΔC, via the stable inversion), m_local (2024 images, our frozen fitter). m_newcat against m_local (same images) isolates the catalogue's fitting and processing (c) from the data (b); m_cat against m_newcat measures what the reprocessing changed — no SAS 18 regeneration needed. The screen must not select on the reading's object (review correction): a catalogue-to-catalogue pipeline screen — match the 2024 PPS source list to DR14 by position, compare the M2/M1 b4 RATE; every reading is reported for the full set and for the |median(RATE_2024/RATE_DR14) − 1| < 0.10 subset. Crowding readings reported with and without a |r_uv| ≤ 2 px cut. Intermediates on the NVMe (the pilot mislabelled the HDD). (Source: reviews/T13_001_claude.md.)
• 第 2 阶段结果（T13_002 评审通过；后经 RT-05 分诊发现位置文件 bug，下列逐行数字与"共享位置"读数作废，保留作历史）：Stage-2 Result (T13_002 Review PASS; later voided in its row-level numbers and the "shared-position" reading by the RT-05 triage's position-file bug — kept as history): 185 个观测（种子在下载前固定；并发下载的共享 tar bug 已修复，只用移动恢复；星表对星表筛选：RATE_2024/RATE_DR14 中位数 1.001，0/185 超出 ±10%）。|r_uv| ≤ 2 px 筛选下：M2 m_cat = 0.086 [0.075, 0.097]、m_newcat = 0.083 [0.073, 0.093]、m_local = 0.048 [0.034, 0.061]；M1 0.067 [0.053, 0.082]、0.065 [0.050, 0.080]、0.033 [0.012, 0.053]（无筛选时 m_local 更低：0.027 和 0.014——移到邻近源的拟合会抬高 ΔC_local，带筛选的行才是可靠的）。读数是"混合"（评审更正 checkpoint 的"主要在处理链"）：大致一半（配对区间很宽：M2 0.56 [0.40, 0.70]，M1 0.50 [0.21, 0.77]）由 SAS 链产生——RT-01 红队分诊把这一半定位为共享星表位置的代价：把本地拟合固定在星表联合位置上，m_fixed ≈ m_cat（按中位数两台相机都成立），放开位置的 ΔC 中位数增益 7.5–8.5，远高于正确位置下 χ²₂ 的期望（约 1.4）；约一半（m_local，自由位置，与 3 参数核同口径）在我们自己的重拟合中存留，m_local = 0.033 / 0.048 仍然成立。流量归一化 (a) 排除为主因：RATE 比每年 +0.0003；S_local/S_cat 1.01–1.03，经核斜率 1.4 最多约 0.035。M1 阶跃在 m_cat、m_newcat 中确立，在 m_local 中未分辨（点估计 0.022 [−0.025, +0.062]）；M2 漂移在 m_local 中确立（−0.0032/年 [−0.0056, −0.0010]）。适用范围：DR14–2024 匹配源。注意事项：这里的 M2 漂移（均值亏损，ML > 50，双相机都有探测的观测）与 T11 的 m_cam 漂移（−0.001/年）估计量和选择不同，只在表内比较。（来源：reviews/T13_002_claude.md。）185 observations (seed fixed before download; the shared-tar concurrency bug found, fixed, repaired by moves only; catalogue-to-catalogue screen: RATE_2024/RATE_DR14 median 1.001, 0/185 outside ±10%). With the |r_uv| ≤ 2 px cut: M2 m_cat = 0.086 [0.075, 0.097], m_newcat = 0.083 [0.073, 0.093], m_local = 0.048 [0.034, 0.061]; M1 0.067 [0.053, 0.082], 0.065 [0.050, 0.080], 0.033 [0.012, 0.053] (without the cut m_local is lower, 0.027 and 0.014 — fits that move to a neighbour inflate ΔC_local, so the cut rows are the reliable ones). The reading is "mixed" (the review corrects the checkpoint's "mostly the processing chain"): roughly half (wide paired interval: M2 0.56 [0.40, 0.70], M1 0.50 [0.21, 0.77]) is produced by the SAS chain — the RT-01 red-team triage locates that half in the cost of the shared catalogue position: fixing the local fit at the catalogue joint position gives m_fixed ≈ m_cat (by medians, on both cameras), and freeing the position gains a median ΔC of 7.5–8.5, far above the χ²₂ expectation (~1.4) at a correct position; about half (m_local, free position, apples-to-apples with the 3-parameter kernel) survives our own refit, and m_local = 0.033 / 0.048 still stands. Flux normalisation (a) is ruled out as the main cause: the RATE ratio is +0.0003 per year; S_local/S_cat 1.01–1.03, at most ~0.035 through the kernel slope of 1.4. The M1 step is resolved in m_cat and m_newcat, unresolved in m_local (point estimate 0.022 [−0.025, +0.062]); the M2 drift in m_local is resolved (−0.0032/yr [−0.0056, −0.0010]). Scope: matched DR14–2024 sources. Caveat: the M2 drift here (mean deficit, ML > 50, observations with detections in both cameras) and T11's m_cam drift (−0.001/yr) differ in estimator and selection — compare only within a table. (Source: reviews/T13_002_claude.md.)
• 第 3 阶段（已评审，T13_003，通过附条件）：Stage 3 (Reviewed, T13_003, PASS with conditions): SAS emldetect 逐项拆分，约 30 个观测、4 种配置：复现管线、单相机单波段、似然支撑改为 15px 圆、关闭展源拟合。哪个开关把 m 从约 m_newcat 拉到约 m_local，哪个就是机制；若原因是支撑截断（ecut/scut），可得到零参数修正：在管线自己的支撑上计算核函数。评审结果：复现 A 与 2024 星表 DET_ML 之比中位数 0.9998（M2）/ 0.9995（M1）——首个 SAS 对 SAS 验证过的复现；B（联合拟合）与 C（截断支撑）不是机制（Δm_B ≈ 0，Δm_C 区间含 0）；关闭展源拟合（D）使 m 降低 0.022–0.032，区间不含 0，但"展源拟合是机制"尚未确立（红队 RT-12 更正：科学样本里几乎没有展源拟合（M2 0/242），D 的 Δm 在 EXT=0 点源上依然存在，所以不是少数展源行造成的；emldetect 文档写明点源始终按 ν=2+n，自由度约定也被排除。剩下的解释是拟合流程效应，待 T13 确立。另：T13 第 3 阶段的 run 表跨观测匹配，约 8% 行受影响，第 5 阶段修正）。条件（第 4 阶段）：先修 t13_s2_make_pos.sh 的位置 bug 并重做第 2 阶段；在星表 RA/DEC 与校正前管线位置两处做固定位置对照；按 EXT 拆分重读第 3 阶段（预先登记判读），并加报中位数与 5% 截尾均值。Item-by-item SAS emldetect decomposition, ~30 observations, 4 configurations: reproduce the pipeline, single camera single band, likelihood support as a 15px circle, extent fitting off. Whichever switch moves m from about m_newcat to about m_local is the mechanism; a support-truncation cause (ecut/scut) would admit a zero-parameter fix: the kernel computed on the pipeline's own support. Review result: the replica A matches the 2024 catalogue DET_ML with per-observation median ratio 0.9998 (M2) / 0.9995 (M1) — the first SAS-vs-SAS validated replica; B (joint fit) and C (truncated support) are not the mechanism (Δm_B ≈ 0, Δm_C intervals across 0); switching extent fitting off (D) lowers m by 0.022–0.032, interval excluding 0, but "extent fitting is the mechanism" is not established (red team RT-12 correction: extent fits are almost absent in the science rows (M2 0/242) and D's Δm persists among EXT = 0 point sources, so it is not carried by a few extended rows; the emldetect documentation states ν = 2 + n for point sources, so a DOF convention is ruled out. What remains is a fitting-procedure effect, to be established by T13. Also: the stage-3 run table matched across observations, affecting ~8 % of rows; fixed in stage 5). Conditions (stage 4): fix the t13_s2_make_pos.sh position bug and redo stage 2 first; the fixed-position control at both the catalogue RA/DEC and the pre-correction pipeline position; re-read stage 3 split by EXT (pre-registered reading), with median and 5% trimmed mean reported beside the mean.
• 第 4 阶段（已评审，T13_004，通过附条件）：Stage 4 (Reviewed, T13_004, PASS with conditions): 位置 bug 修复已证明（t13_s2_make_pos_v2.sh，错配 0，最大距离 0 px），第 2 阶段重做与 RT-05 的临时数字完全一致（m_local 0.061 / 0.038；差距 0.020 [−0.006, +0.042] / 0.023 [+0.007, +0.037]）。固定位置对照：每观测、两台相机共有的偏移真实但很小（中位 0.22–0.31 px）；固定在管线校正前位置 (b) 时 m ≈ 0.069 / 0.062，接近 m_newcat（0.078 / 0.058）——catcorr 假设不是主因；自由度约定被文档和数据排除（DC_D/DC_B = 1.000）。Δm_D 按 D−B 的 ML 位移拆分：ML_D − ML_B ≥ 10 的一组（M2 41/243、M1 53/223，全部是星表点源）承载 Δm_D 的 65–71%。方向更正（评审核对 checkpoint 后）：开启展源拟合时（A、B 和管线本身），这组点源的拟合计数更低（中位 S_B/S_D = 0.77 / 0.73，S_A/S_D = 0.93 / 0.84，星表 S/S_D = 0.88 / 0.80，即 12–20%），DET_ML 低 27–43——checkpoint 里"B 用展源参数抬高它们的 ΔC"方向反了。若属实，这是比 ΔC 亏缺更强的结论（在这 30 个重拟合观测里约五分之一的亮点点源流量偏低——M2 40/242、M1 54/223；4XMM 全目录的份额未测，该组只占目录左尾的约 25 %，07 R006），但尚未定论：是管线欠拟合还是纯点源拟合过拟合，必须用已知真值的注入检验决定（第 5 阶段）。条件：(1) run 匹配器仍有 RT-12 的跨观测 bug（s4_reread.py:90–103 不限制 OBS_ID，8% 的行取了别的观测的值），须限制 OBS_ID 后重导第 4–6 节全部表格；(2) 报中位数与 5% 截尾均值；(3) 按 EXT 拆分 Δm_D；(4) 排查 nmulsou = 2 的源拆分。The position-bug fix is proven (t13_s2_make_pos_v2.sh, 0 mismatches, max distance 0 px) and the stage-2 redo reproduces RT-05's provisional numbers exactly (m_local 0.061 / 0.038; gap 0.020 [−0.006, +0.042] / 0.023 [+0.007, +0.037]). Fixed-position controls: the per-observation, camera-common offset is real but small (median 0.22–0.31 px); fixing at the pipeline's pre-catcorr position (b) brings m to ≈ 0.069 / 0.062, close to m_newcat (0.078 / 0.058) — the catcorr hypothesis is not the main story; the DOF convention is refuted by the documentation and the data (DC_D/DC_B = 1.000). The decomposition of Δm_D by the D−B ML shift: the ML_D − ML_B ≥ 10 group (M2 41/243, M1 53/223; all catalogue point sources) carries 65–71 % of Δm_D. Direction corrected (the review checked the checkpoint): with fitextent = yes (runs A, B and the pipeline itself) these point sources get lower fitted counts (median S_B/S_D = 0.77 / 0.73, S_A/S_D = 0.93 / 0.84, catalogue S/S_D = 0.88 / 0.80, i.e. 12–20 %) and DET_ML lower by 27–43 — the checkpoint's "B inflated their DC via the extent parameter" has the direction backwards. If real, this is a stronger conclusion than a ΔC deficit (about one bright point source in five has low flux in these 30 refitted observations — M2 40/242, M1 54/223; the share across all of 4XMM is not measured, and the group is only ≈ 25 % of the catalogue's left tail, 07 R006), but not yet decided: whether the pipeline under-fits or the pure point fit over-fits must be decided by a known-truth injection test (stage 5). Conditions: (1) the run matcher still has the RT-12 cross-observation bug (s4_reread.py:90–103 does not restrict OBS_ID, so 8 % of rows take another observation's value) — restrict to the row's OBS_ID and re-derive every table of §4–6; (2) report median and 5 % trimmed mean; (3) split Δm_D by EXT; (4) rule out nmulsou = 2 source splitting.
• 第 2 阶段 bug 与下一步（RT-05 分诊）：The stage-2 bug and the next step (RT-05 triage): t13_s2_make_pos.sh 按原始行序计算像素位置、却与按 SRC_NUM 排序的行配对（带排序一侧的 zip bug；试点脚本 t13_make_pos.py 是对的），约 90% 的行用了别的源的位置——第 2 阶段所有逐行量和 RT-01 的固定位置对照作废。下一步（第 5 阶段，运行中）：第 4 阶段已修复 bug、重做第 2 阶段（与 RT-05 临时数字完全一致）并完成两处固定位置对照（catcorr 假设排除）。第 5 阶段做已知真值的注入检验：注入已知 S 的点源（E094 设置或 PI 的 M31 注入图像，virgo，SAS），跑管线配置（A 类：fitextent = yes、withtwostage、nmulsou = 2、dmlextmin = 4）和 D（fitextent = no），报每个 run 的 S_fit/S_true 与 ΔC、以及落入"大位移"类的注入点源比例随 S 与 b 的变化；同时排查 nmulsou = 2 的源拆分（大位移组里 A、B 是否在约 20″ 内多出一个 D 没有的探测，S_split1 + S_split2 ≈ S_D 吗），并修正 run 匹配器的跨观测 bug（RT-12）。修正标签后的临时数字：m_local ≈ 0.061 / 0.038（M2/M1）；位置差 25%（M2，未分辨）到 38%（M1），不是"约一半"。新假设（待检验，不是结论）：4XMM 的 RA/DEC 带 catcorr 天体测量校正而 PPS 图像 WCS 没有。t13_s2_make_pos.sh computed pixel positions in raw row order but paired them with rows sorted by SRC_NUM (a zip bug with one sorted side; the pilot script t13_make_pos.py was correct), so about 90% of rows used another source's position — all stage-2 row-level quantities and RT-01's fixed-position control are void. Next step (stage 5, running): stage 4 fixed the bug, redid stage 2 (matching RT-05's provisional numbers exactly) and completed both fixed-position controls (the catcorr hypothesis ruled out). Stage 5 runs the known-truth injection test: inject point sources of known S (the E094 setup or the PI's M31 injection images, virgo, SAS), run the pipeline configuration (A-like: fitextent = yes, withtwostage, nmulsou = 2, dmlextmin = 4) and D (fitextent = no), report S_fit/S_true and ΔC per run and the fraction of injected point sources falling into the "big-shift" class versus S and b; also rule out nmulsou = 2 source splitting (in the big-shift group, do A and B have a second detection within about 20″ that D lacks, and does S_split1 + S_split2 ≈ S_D?) and fix the run matcher's cross-observation bug (RT-12). Corrected provisional numbers: m_local ≈ 0.061 / 0.038 (M2/M1); the position gap is 25% (M2, unresolved) to 38% (M1), not "about half". New hypothesis (to be tested, not a result): 4XMM's RA/DEC carry the catcorr astrometric correction while the PPS image WCS does not.

第 7e 步 · T5 阶段 15–18：高 S 报告总体（已完成，T5_012–015 评审通过；T5 暂停）Step 7e · T5 Stages 15–18: The High-S Report Populations (Done, T5_012–015 Reviews PASS; T5 Paused)

• 目标：Goal: 把高计数报告总体扩到 S ≤ 4000（C-all-hi、C-iso-hi，只作报告，新增窗口 100–300 和 >300），检验 m 是否延续到亮源。Extend the high-count report populations to S ≤ 4000 (C-all-hi, C-iso-hi, report-only, new windows 100–300 and >300), to test whether m persists to bright sources.
• 阶段 15 结果：Stage 15 Result: 已建好，6–10 窗口复现（差值 0.0），但 >300 窗口作废：ML→ΔC 换算 ΔC = 2·gammainccinv(1.5, exp(−ML)) 在 ML ≳ 745 处 exp(−ML) 下溢为 inf；S > 300 只剩 467 行（原始 300 < S ≤ 4000 的源有 17,958 个，其中 25.7% 的 DET_ML ≥ 690）。>300 窗口的 A1 = 16、m = 0.83 是数值伪迹，不是物理。（来源：reviews/T5_012_claude.md。）（评审更正，T5_013：467 行的流失来自筛选条件，EP_EXTENT==0 去掉 12,514 行，SUM_FLAG==0 去掉 4,347 行，公开切分再去一半；溢出没有删行，而是让这 467 行中 402 行的 ΔC 失真。）Built; the 6–10 window reproduces (difference 0.0), but the >300 window is void: the ML→ΔC inversion ΔC = 2·gammainccinv(1.5, exp(−ML)) underflows to inf at ML ≳ 745; only 467 rows survive at S > 300 (17,958 raw sources with 300 < S ≤ 4000, of which 25.7% have DET_ML ≥ 690). The >300 window's A1 = 16 and m = 0.83 are numerical artefacts, not physics. (Source: reviews/T5_012_claude.md.) (Review correction, T5_013: the attrition to 467 rows comes from the filters, EP_EXTENT==0 removing 12,514 and SUM_FLAG==0 removing 4,347, then the public split; the overflow removed no rows but distorted ΔC for 402 of the 467.)
• 阶段 16 结果（T5_013 评审通过）：Stage 16 Result (T5_013 Review PASS): 稳定反解上线（erfcx 恒等式 + brentq，往返精度 4.4e-13，ML < 600 处与旧代码一致到 2e-13）；已登记 6–10 成绩复现差值 0.0。m 延续到亮源：未修正参照 V4c_no_eta 在 >300 的 frac = −0.0647，与 100–300 的 −0.0663 相同；Δm = −0.006 [−0.0077, −0.0004]（C-all）、+0.0007 [−0.0029, +0.0034]（C-iso）——约 6% 的乘性亏损从 S ≈ 100 延续到 S ≈ 4000。ML-700 截断曾让 m 偏高约 0.002：参照的 m(S ≤ 300) 从 0.0660 修正为 0.0643。Stable inversion live (erfcx identity + brentq, round-trip accuracy 4.4e-13, agrees with the old code below ML 600 to 2e-13); registered 6–10 scores reproduce to 0.0. m extends to bright sources: the uncorrected reference V4c_no_eta has frac = −0.0647 in >300 against −0.0663 in 100–300; Δm = −0.006 [−0.0077, −0.0004] (C-all), +0.0007 [−0.0029, +0.0034] (C-iso) — the ~6% multiplicative deficit holds from S ≈ 100 to S ≈ 4000. The ML-700 clip had biased m up by ~0.002: the reference's m(S ≤ 300) moves from 0.0660 to 0.0643.
• 阶段 17 结果（T5_014 评审通过）：Stage 17 Result (T5_014 Review PASS): 15 个 Φ 网格扩展到 S = 5000（旧节点差值 0.0）；S > 300 的 w 和 L2d 补齐，覆盖率 100%，200 行对照与 T11 差值 0.0；已登记成绩逐位复现（最大差 5.6e-16，既有浮点归约）。六个零参数 T3 条目全部延续：|Δm| < 0.01（C-all 上 −0.006 到 +0.0002，C-iso 上 +0.001 到 +0.004），乘性亏损在亮源上不变；在 >300 背景窗口，它们的配对 A1 差全部为负且区间不含 0（2-D 条目增益最大，A1 −0.43）。V5c_eta（样本内 6 参数）在亮源上比不修正更差：A1 差 +0.58 [+0.43, +0.75]（C-all）、+0.89 [+0.86, +1.01]（C-iso）——"为了修正而修正"的实例。现行标记只覆盖 ≤300 的窗口，所以它未被标记。15 Φ grids extended to S = 5000 (old nodes 0.0); w and L2d filled in above S = 300, coverage 100%, 200-row cross-check against T11 to 0.0; registered scores reproduce bit for bit (max diff 5.6e-16, pre-existing floating-point reduction). All six T3 entries (four zero-parameter) extend: |Δm| < 0.01 (−0.006 to +0.0002 on C-all, +0.001 to +0.004 on C-iso), the multiplicative deficit is unchanged on bright sources; in the >300 context window their paired A1 differences are all negative with intervals excluding 0 (the 2-D entries gain the most, A1 −0.43). V5c_eta (in-sample 6-parameter) is worse than no correction on bright sources: A1 difference +0.58 [+0.43, +0.75] (C-all), +0.89 [+0.86, +1.01] (C-iso) — the "correction for its own sake" case. The current flag covers windows ≤ 300 only, so it is not flagged.
• 阶段 18 结果（T5_015 评审通过）：Stage 18 Result (T5_015 Review PASS): 区间缺陷的根因不是小窗口偏差，而是两个统计量：bootstrap 重抽样的是原始 A1，点估计却是噪声校正后的 A1_nc。修复后 bootstrap 直接计算所报告的统计量（固定 N_eff、逐模型簇 N_eff、底线扣除、每个重抽样丢弃空 bin），点估计从充分统计量精确复现。440/440 主区间包含点估计（规则：BCa 稳定时用 BCa，否则 basic 并扩展到含点；只有 C-iso V5c_eta >300 一行需要扩展，已注明）；70/70 个标记判定无一改变；已登记 6–10 成绩复现。V5c_eta 结论分层：|frac| 差在两个层上所有方法都不含 0（C-all [+0.033, +0.034]）；A1 差在 C-iso 上所有方法都不含 0（BCa [+0.90, +0.93]），在 C-all 上依赖方法（百分位 [−0.18, +0.72]，basic [+0.44, +1.34]，BCa [+0.31, +0.78]，只有 68 个 bin）。高计数报告页已发布：../v2/highcount.html（含稳定反解、-hi 总体、m/a、标记、>300 背景表、L0 对照与方法说明）。（来源：reviews/T5_015_claude.md。）The interval defect's root cause was not small-window bias but two different statistics: the bootstrap resampled raw A1 while the point estimate was the noise-corrected A1_nc. After the fix the bootstrap evaluates the reported statistic (fixed N_eff, per-model cluster N_eff, the floor subtraction, empty bins dropped per replicate), and the point is reproduced exactly from the sufficient statistics. 440/440 headline intervals contain their point (rule: BCa when stable, otherwise basic expanded to contain the point; only C-iso V5c_eta >300 needed the expansion, documented); no flag verdict changed (70/70); registered 6–10 scores reproduce. The V5c_eta finding, layered: its |frac| difference excludes 0 under every method on both layers (C-all [+0.033, +0.034]); its A1 difference excludes 0 on C-iso under every method (BCa [+0.90, +0.93]), and on C-all depends on the method (percentile [−0.18, +0.72], basic [+0.44, +1.34], BCa [+0.31, +0.78], only 68 bins). The high-count report page is published: ../v2/highcount.html (stable inversion, -hi populations, m/a, flags, the >300 context table, the L0 control, and method notes). (Source: reviews/T5_015_claude.md.)
• 用户决定（待定）：User Decision (Pending): 高计数退化标记是否扩展到 >300 窗口（若扩展，只有 V5c_eta 被标记；前沿不受影响，因为 V5c_eta 是样本内条目）；高 S 总体是否进层定义。Whether the high-count degradation flag extends to the >300 window (if yes, only V5c_eta is flagged; the front is unaffected since V5c_eta is in-sample); whether a high-S population enters the layer definitions.

第 8 步 · J 层产能（3 相机 × 5 波段）Step 8 · J Layer Capacity (3 Cameras × 5 Bands)

• 目标：Goal: 多相机联合模拟扩到 15 幅图像，检验二维损失和背景误差律的最终形式。Scale multi-camera joint sim to 15 images, test final form of two-dimensional loss and background-error law.
• 输入：Inputs: E155 的试点（约定往返闭合到 1e-6；产能估算约 4,200 CPU-h）；sirius 上的原始图像。E155 pilot (round-trip closure to 1e-6; capacity estimate ~4,200 CPU-h); raw images on sirius.
• 方法：Method: 按 E155 第 7 节的生产计划；先波段 4，审查后再 3×5。每场记注入表（含未探测源）。Per E155 §7 production plan; band 4 first, review then 3×5. Every run logs injection table (including undetected sources).
• 成败判据：Success/Failure Test: 约定往返 ≤ 1e-6；入选可由存储列复现。Round-trip ≤ 1e-6; selection reproducible from stored columns.
• 时间：Time: 数周（SAS 产能主导）。Several weeks (SAS capacity driven).
• 先读：Read First: cognition-os/experiments/E155_joint_emldetect_virgo.md → BENCHMARK_V2_TASKS.md 的 T4 卡。cognition-os/experiments/E155_joint_emldetect_virgo.md → BENCHMARK_V2_TASKS.md T4 card.

第 9 步 · 里程碑：请 PI 开私有位置Step 9 · Milestone: Ask PI to Open Private Positions

• 目标：Goal: 用从未见过的 1/3 位置检验前沿是否过拟合。这是 PI 的动作；学生只准备好材料。Test front for over-fitting on the never-seen 1/3 positions. This is PI's action; student only prepares materials.
• 输入：Inputs: 第 2、3、5、6、7 步的完成；一份“哪些条目在前沿”的清单。Completion of Steps 2, 3, 5, 6, 7; a list of "which entries are on the front".
• 方法：Method: 把清单和理由写成实验笔记交给 PI。Write list and rationale into experiment note, hand to PI.
• 成败判据：Success/Failure Test: 私有分数与公开分数在噪声内一致。Private scores consistent with public scores within noise.
• 时间：Time: 半天（PI 执行）。Half day (PI executes).
• 先读：Read First: PARETO_BENCHMARK_V2.md §6（Holdout and access）。PARETO_BENCHMARK_V2.md §6 (Holdout and access).

第 10 步 · T13 第 7 阶段第 2 部分：近邻占目录 m 的份额（在跑，待定 / Pending）Step 10 · T13 Stage-7 Part 2: The Neighbour's Share of the Catalogue's m (Running, Pending)

• 目标：Goal: 测出近邻（DIST_NN < 40″）能解释目录高计数亏损 m 的多大份额。第 1、3 部分已写入 checkpoint，第 2 部分等 virgo 驱动。Measure how much of the catalogue's high-count deficit m near neighbours (DIST_NN < 40″) explain. Parts 1 and 3 are in the checkpoint; part 2 waits for the virgo driver.
• 输入：Inputs: virgo 上的隔离运行驱动（重注入行，逐源比较近邻与源自身的残差）；全目录 r_cat 按 Ŝ × DIST_NN 分箱。The virgo isolation-run driver (re-injected rows, per-source neighbour-vs-source residual); catalogue-wide r_cat binned by Ŝ × DIST_NN.
• 方法：Method: 预注册的隔离运行；份额用亏损（−r）而不是恢复比例（1+r）——s7_part3.py:183-185 的 bug 已定位，须由第 2 部分 worker 修正后重写该段。隔离臂（isoA）至今没有测到任何东西（07 R024）：s7_iso_build.py 静默跳过，0/282 个近邻模板存在，isoA 与 base 在 559/564 行上完全相同（virgo parsed.csv），所以"近邻模型移除只解释 ≲ 1 % 的大−对照差距"这句已撤回——近邻效应尚未被检验。第 7b 阶段必须用真实的近邻模板重跑，并加"两臂图像必须不同"的断言。Pre-registered isolation run; the share uses the deficit (−r), not the recovered fraction (1+r) — the bug at s7_part3.py:183-185 is located and the part-2 worker must correct it and rewrite that paragraph. The isolation arm (isoA) has measured nothing so far (07 R024): s7_iso_build.py skipped silently, 0/282 neighbour templates exist, and isoA equals base on 559/564 rows (virgo parsed.csv), so "removing the neighbour model accounts for ≲ 1 % of the big − control gap" is withdrawn — the neighbour effect is untested. Stage 7b must re-run with real neighbour templates and add an "the two arms' images must differ" assertion.
• 成败判据：Success/Failure Test: 份额带一个区间；重注入与目录的 big − control 差在模板因子范围内解释得了（重注入÷目录的 big−control 比值，中位数、配对 bootstrap：M2 A 类 1.54 [1.18, 1.67]、M1 1.09 [0.88, 1.15]；D run 0.77 / 0.60；不要再引"1.2–1.5×"这个笼统区间，07 R025）。在份额测出前不得写成结论。The share carries an interval; the re-injection-vs-catalogue big − control difference is explained within the template factor (the re-injection ÷ catalogue big − control ratio, median with paired bootstrap: M2 A-like 1.54 [1.18, 1.67], M1 1.09 [0.88, 1.15]; the D run 0.77 / 0.60; stop quoting the blanket "1.2–1.5×" range, 07 R025). Do not write it as a result before the share is measured.
• 时间：Time: virgo 驱动 21:30Z 时 152/234 幅图像，ETA ≈ 23:40Z；之后是复核。The virgo driver was at 152 of 234 images at 21:30Z, ETA ≈ 23:40Z; then the review.
• 先读：Read First: checkpoints/T13_007.md（第 1、3 部分）→ reviews/RT-27_triage_claude.md（O1/O2 必须用）→ 07 R021。checkpoints/T13_007.md (parts 1 and 3) → reviews/RT-27_triage_claude.md (O1/O2 must be used) → 07 R021.

第 11 步 · Q1 的决定性一步：按目录给定的 Ŝ 采 b 的联合模拟（待派发，待定 / Pending）Step 11 · Q1's Decisive Step: A Joint Simulation at the Catalogue's b Given Ŝ (To Dispatch, Pending)

• 目标：Goal: 给 Q1 一个受控比较：联合 emldetect 模拟必须采目录的 b（给定 Ŝ），用估计背景按 Ŝ 分箱在所有窗口计分，对用同一代码重算的目录。Give Q1 a controlled comparison: the joint emldetect simulation must sample the catalogue's b given Ŝ, scored per Ŝ bin over all windows with the estimated background, against the catalogue recomputed with the same code.
• 输入：Inputs: 现有联合模拟行（可按 b 重加权）或新的一批；cat_sens.py 的口径；目录的 b(Ŝ) 关系（0.02→0.14）。The existing joint-simulation rows (reweightable by b) or a new batch; the cat_sens.py convention; the catalogue's b(Ŝ) relation (0.02→0.14).
• 方法：Method: 试点曾用恒定的 b \(\approx\) 0.028，两个 6–10 窗口因此落在不同 Ŝ（7.3–8.5 对 9.5–13）。新运行必须匹配 Ŝ，并检验亏损是否延续到 Ŝ > 30（试点在那里消失，目录停在 −7 %）。The pilot used a flat b \(\approx\) 0.028, so its two 6–10 windows sat at different Ŝ (7.3–8.5 against 9.5–13). The new run must match Ŝ and test whether the deficit persists to Ŝ > 30 (where the pilot's vanishes while the catalogue stays at −7 %).
• 成败判据：Success/Failure Test: 在匹配 Ŝ 后，模拟的亏损与目录的 −9.282 % [−9.350, −9.219] 之差落在噪声底内；否则 2-D 位置损失不能算闭合项。At matched Ŝ the gap between the simulation's deficit and the catalogue's −9.282 % [−9.350, −9.219] lies within the noise floor; otherwise the 2-D position loss cannot count as the closing term.
• 时间：Time: brief 待 Lang3/virgo 负载允许后写；产能与第 8 步同量级。The brief is written when Lang3/virgo load allows; capacity of the same order as Step 8.
• 先读：Read First: reviews/RT-26_triage_claude.md（决定一节）→ 07 R019 → 02 Q1。reviews/RT-26_triage_claude.md (the Decision section) → 07 R019 → 02 Q1.

第 12 步 · 双成分混合拟合（E165 行对 P7，样本外）（待 hybrid 结果，待定 / Pending）Step 12 · The Two-Component Mixture Fit (E165 rows against P7, held out) (Awaiting the hybrid result, Pending)

• 目标：Goal: 按 08 §7.1 拟合 (m₀, s₀, f, Δ, off-axis) 五个参数，样本外检验，回答 Q3 的宽度形状问题。Fit (m₀, s₀, f, Δ, off-axis) per 08 §7.1, held out, to answer Q3's width-shape question.
• 输入：Inputs: E165 的行；P7 参照；按相机与离轴角分层。E165's rows; the P7 reference; stratified by camera and off-axis angle.
• 方法：Method: 混合模型的最大似然拟合；成败判据同时看 sd 与尾部覆盖（PIT 直方图，不只是 sd）。已知的坑：m 随离轴角最多动 0.09，s 随背景上升（PN 0.090 → 0.139），MOS 里 ≈ 11–12 % 的低效率子总体拟合更好（07 R007）。Maximum-likelihood fit of the mixture; the test looks at the sd and the tail coverage together (a PIT histogram, not just the sd). Known traps: m moves with off-axis by up to 0.09, s rises with background (PN 0.090 → 0.139), and in MOS an ≈ 11–12 % low-efficiency sub-population fits better (07 R007).
• 成败判据：Success/Failure Test: 每个窗口的样本外 A2 ≤ 0.10 且参数 ≤ 8 个，尾部覆盖正确。Held-out A2 ≤ 0.10 in every window with ≤ 8 parameters, plus correct tail coverage.
• 时间：Time: 等 E165 的 hybrid 结果存在后约 1–2 天。About 1–2 days once E165's hybrid result exists.
• 先读：Read First: cognition-os/08_RUNNING.md §7.1 → 02 Q3 → 07 R007。cognition-os/08_RUNNING.md §7.1 → 02 Q3 → 07 R007.

第 13 步 · 把附录基线计分推到其余配置（待派发，待定 / Pending）Step 13 · Score the Appendix Baseline on the Other Configurations (To Dispatch, Pending)

• 目标：Goal: 回答 Q4：已登记的 appendix_eff（+ appendix_psfgen 作参考）目前只有 M2 band 4。把计分推到波段 1–4 的其余配置并重算前沿。Answer Q4: the registered appendix_eff (plus appendix_psfgen as a reference) currently covers M2 band 4 only. Push the scoring to the other configurations in bands 1–4 and recompute the fronts.
• 输入：Inputs: 已登记的四条目；波段 1–4 的核网格；基准计分器。The four registered entries; the kernel grids for bands 1–4; the benchmark scorer.
• 方法：Method: 逐配置计分，重算样本外前沿；报告时区分原始分数（前沿用）与噪声校正值（头条用）。Score configuration by configuration and recompute the out-of-sample fronts; when reporting, distinguish raw scores (used by the fronts) from noise-corrected values (used in headlines).
• 成败判据：Success/Failure Test: 零参数条目在其余配置上仍在样本外前沿上，或其落选的原因能写清楚。已知：0.44 的优势只对被压制的参照成立；对前沿上最好的均值分数条目 appendix_eff 更差（C-iso 0.1527 对 0.1321）；它靠宽度上前沿。The zero-parameter entry stays on the out-of-sample front in the other configurations, or the reason it drops off is written down. Known: the 0.44 edge is only against the dominated reference; against the front's best mean-score entry appendix_eff is worse (C-iso 0.1527 vs 0.1321) and earns its seat on width.
• 时间：Time: 约 1 天（计分器已存在）。About 1 day (the scorer already exists).
• 先读：Read First: reviews/RT-25_triage_claude.md → 07 R018 → 02 Q4。reviews/RT-25_triage_claude.md → 07 R018 → 02 Q4.

第 14 步 · 高 S 层（C-all-hi / C-iso-hi，S 上限 4000）（已批准，未开跑，待定 / Pending）Step 14 · The High-S Layers (C-all-hi / C-iso-hi, S cap 4000) (Approved, Not Yet Run, Pending)

• 目标：Goal: PI 2026-10-16 批准：在现有层之外新增两个高 S 报告层，检验亏损是否延续到亮源。PI-approved on 2026-10-16: add two high-S report layers beside the existing ones and test whether the deficit persists to bright sources.
• 输入：Inputs: 已扩展到 S = 5000 的核网格（T5_014）；稳定反解（T5_013）。The kernel grids extended to S = 5000 (T5_014); the stable inversion (T5_013).
• 方法：Method: 建层并计分；新层只作报告，不替代已登记层，已登记条目与 6–10 前沿不变。Build the layers and score them; report-only, not replacing the registered layers, with the registered entries and the 6–10 front unchanged.
• 成败判据：Success/Failure Test: 已登记成绩在新层上复现差值 0.0；亏损的高 S 行为带区间。Registered scores reproduce to 0.0 on the new layers; the high-S behaviour of the deficit carries an interval.
• 时间：Time: 等 Lang3 负载允许；约 1–2 天。Waits for Lang3 load; about 1–2 days.
• 先读：Read First: reviews/T5_015_claude.md → 08_RUNNING.md §5 → 02 的层定义。reviews/T5_015_claude.md → 08_RUNNING.md §5 → 02's layer definitions.

第 15 步 · 宽度线四任务：玩具理论、L0/L1 梯子、SAS 版本重放、Q1 目标（已完成并评审，T229/T228/T227/R035）Step 15 · The Four Width-Line Tasks: Toy Theory, the L0/L1 Ladder, the SAS Version Replay and the Q1 Target (Done and Reviewed, T229/T228/T227/R035)

• 目标：Goal: 回答 Q7——零参数插入式 \(\sqrt{V}\) 能不能描述我们自己的模拟里给定 \(\hat{S}\) 的 \(\Delta C\) 宽度，在哪里失效。PI 2026-10-17 07:10Z 把这个问题定为主攻方向。Answer Q7 — does the zero-parameter plug-in \(\sqrt{V}\) describe the \(\Delta C\) width given \(\hat{S}\) in our own simulations, and where does it fail. The PI made this the main line on 2026-10-17 07:10Z.
• 输入：Inputs: W1：玩具模拟（15\(\times\)15 高斯 PSF，\(\sigma\)=1.8 px，局部 3 参数拟合，每格 40,000 次重复），每个公式附与精确模拟对照的测试脚本。W2：L0 生产表（181 块、2,425,396 行）与匹配的 L1 行，注册 e114 核在标称位置的插入式，逐行 z。SAS 重放：SAS 20.0.0 与 22.1 同输入、显式同参数。Q1：目录 M2 b4 的 cat_sens.py 口径与试点的 L2d 表。W1: the toy simulation (15\(\times\)15 Gaussian PSF, \(\sigma\) = 1.8 px, local 3-parameter fit, 40,000 reps per cell), every formula shipped with a test script against exact simulation. W2: the L0 production tables (181 chunks, 2,425,396 rows) and the matched L1 rows, the registered e114 kernel plug-in at the nominal position, per-row z. The SAS replay: SAS 20.0.0 and 22.1 on identical inputs with explicit parameters. Q1: the catalogue M2 b4 cat_sens.py convention and the pilot's L2d table.
• 方法：Method: 理论派发规则（PI 2026-10-17）：Claude 先写"已知结论 + 可执行判据"，两个独立臂并行，红队用数值反例攻击推导，Claude 复核判据脚本后才接受任何公式。可执行判据：|mean z| < 0.02 且 |std z − 1| < 0.02（逐行条件化，模拟重复数 \(\geq\) 10,000 处）。宽度一律按典型 DET_ML 分箱，不用 \(\hat{S}\) 分位数分箱（R030）。Theory-dispatch rule (PI 2026-10-17): Claude writes the known results and the executable key first, two independent arms run in parallel, red teams attack the derivations with numerical counterexamples, and Claude replays each formula's test script before accepting it. Executable key: |mean z| < 0.02 and |std z − 1| < 0.02 (per-row conditioning, wherever the simulation has \(\geq\) 10,000 reps). Widths are always binned by typical DET_ML, never by \(\hat{S}\) quantiles (R030).
• 成败判据：Success/Failure Test: 逐格通过/失败表，达不到就带数字报失败；SAS 重放按预注册的 identical / equivalent / different / ambiguous 规则字面执行。A per-cell pass/fail table, anything else reported as a failure with numbers; the SAS replay applies the pre-registered identical / equivalent / different / ambiguous rule literally.
• 时间：Time: W1 两臂 07:09Z \(\to\) 13:22Z（codexnim）与 \(\approx\) 14:05Z（Hermes）；W2 09:49Z \(\to\) 12:00Z；SAS 重放 06:58Z \(\to\) \(\approx\) 10:19Z；Q1 步骤 1–2 当天完成。W1's arms 07:09Z \(\to\) 13:22Z (codexnim) and \(\approx\) 14:05Z (Hermes); W2 09:49Z \(\to\) 12:00Z; the SAS replay 06:58Z \(\to\) \(\approx\) 10:19Z; Q1's steps 1–2 finished the same day.
• 先读：Read First: cognition-os/experiments/E170_w1_width_theory_toy.md → E169_sas22_vs_sas20_source_detection.md → reviews/W1_codexnim_claude.md → reviews/W2_001_claude.md → reviews/SASVER_001_claude.md → reviews/Q1_s12_claude.md → 07 R035–R038。cognition-os/experiments/E170_w1_width_theory_toy.md → E169_sas22_vs_sas20_source_detection.md → reviews/W1_codexnim_claude.md → reviews/W2_001_claude.md → reviews/SASVER_001_claude.md → reviews/Q1_s12_claude.md → 07 R035–R038.
• 结果（四条都带来源与信任标签）：Result (all four with their source and trust label): W1（T229/E170，Provisional，07 R037；宽度失效一半 07 R042 起为 Questioned）：位置固定的 Fisher 宽度 \(V_{3p}\) 在 DET_ML \(\gtrsim\) 20 与 b = 0.005 通过（std z 0.98–1.00），在 DET_ML 6–20 且 b \(\geq\) 0.02 失效（1.17–2.49，上尾重）——但 RT-43 定位到这些失效格来自 codexnim 臂拟合器的 \(\pm 8\) px 位置截断（0.3–2 % 的重复；一个被追踪的重复 \(\Delta C\) 7.29 对 31.78），项目 Nelder–Mead 在三个争议格给 1.008 / 0.987 / 0.989（N = 2000）、尾偏度 +0.51（原 +9.09），所以"宽度失效"待仲裁复核——本阶段（T12）用项目拟合器复跑三个争议格（N = 2000，种子 777）得 std z 1.008 / 1.010 / 0.963，对记录的 1.171 / 2.486 / 1.331；第三格差 −0.037 未满足 R042 预注册判据，调度方不采用字面 revert，而是废弃记录的失效并把 (15.32, 0.08) 的宽度标为 open（偏窄不偏宽，N = 10,000 复跑入队，07 R044）；该复跑给出 std z 0.975（约 3.6\(\sigma\) 低于 1），按预注册规则记为窄侧偏离约 2.5 %（Provisional，07 R045，方向与 L1 在 6–10 的 0.95 一致）；条件均值 \(\mu_{3p}\) 每格偏低 +0.8…+1.7 \(\Delta C\)（三种拟合器一致；R047 撤回"无均值公式通过"：E088 \(\delta_0\) 在 6 格中 4 格把 mean z 关到 −0.002…+0.018，DET_ML \(\approx\) 8 的两格差 +0.06；T231）；截断（支撑截断）矩 27/27 格全通过——位置固定（R048：位置自由时只有拟合位置模板接近，+0.10…+0.16 / 1.03–1.05，开放）；背景盒项只在 N_bg = 10,000 通过（R048：均值只在 N_bg \(\geq\) 1000 修正）。Hermes 臂的相反结论被仲裁运行拒绝（项目拟合器给出 1.024 / 0.993）。W2（T228，Provisional，07 R036、R039；L1 数字 07 R041 起换数）：L1 emldetect std z 0.95 / 0.975 / 1.00 / 1.02（DET_ML 6–10 / 10–20 / 20–50 / 50–100），mean z −0.05…−0.07——用 emldetect 实际得到的背景 BG_MAP_EML 打分；用设计水平 b_true 打分得到的 std z 1.02–1.12、mean z +0.28…+0.47 与"PSF/通量标度"解释已撤回（RT-45 分诊）；L0 合并 std 1.45 / 1.52 是两种总体的混合（远行 \(\Delta\)C 低于插入式，mean z −1.70 / −4.03；近处行 0.89 / 0.96 按拟合结果条件化，不是给定 \((\hat{S}, b)\) 的宽度）。SAS 重放（T227/E169，Provisional，07 R034、R038）：emldetect 与 eboxdetect 逐位一致，只有 esplinemap 变；真背景的 L1/J 生产无需版本注记，esplinemap 臂带一条。Q1（07 R035，bonus）：目标冻结 −9.282 % [−9.350, −9.219]；试点的"L2d/\(\mu\) 4.9 %"是代码伪迹（10.4 %）。W1 (T229/E170, Provisional, 07 R037): the position-fixed Fisher width \(V_{3p}\) passes at DET_ML \(\gtrsim\) 20 and at b = 0.005 (std z 0.98–1.00) and fails at DET_ML 6–20 with b \(\geq\) 0.02 (1.17–2.49, heavy upper tail); the conditional mean \(\mu_{3p}\) is low by +0.8…+1.7 \(\Delta C\) in every cell (R047 withdraws "no passing formula": E088 \(\delta_0\) closes mean z to −0.002…+0.018 at 4 of 6 cells, +0.06 at the two DET_ML \(\approx\) 8 cells; T231); the masked (support-truncated) moments pass in 27 of 27 cells — with the position fixed (R048: with it free only the fitted-template plug-in is close, +0.10…+0.16 / 1.03–1.05, open); the background-box term passes only at N_bg = 10,000 (R048: its mean is fixed only for N_bg \(\geq\) 1000). The Hermes arm's contrary conclusion was rejected by an arbiter run (the project fitter gives 1.024 / 0.993). 07 R042 (RT-43's triage): the width-failure half is now Questioned — the failing cells come from the codexnim arm's fitter \(widthlib.fit_3p\) walking to its \(\pm 8\) px position clip (0.3–2 % of reps; on one traced rep \(\Delta C\) 7.29 against 31.78), and the project Nelder–Mead fitter gives std z 1.008 / 0.987 / 0.989 at the three disputed cells (N = 2000) with tail skew +0.51 (was +9.09); this stage (T12) re-ran them (N = 2000, seed 777) for 1.008 / 1.010 / 0.963 against the recorded 1.171 / 2.486 / 1.331; the third cell misses R042's pre-registered criterion by −0.037, so the dispatcher declines the literal revert, deprecates the recorded failure and marks the width at (15.32, 0.08) open (narrower, not wider; a N = 10,000 re-run is queued, 07 R044); that re-run gives std z 0.975 (about 3.6\(\sigma\) below 1), recorded by the pre-registered rule as a narrow-side departure of about 2.5 % (Provisional, 07 R045, the same direction as L1's 0.95 at 6–10). W2 (T228, Provisional, 07 R036, R039; the L1 numbers were replaced by 07 R041): L1 emldetect std z 0.95 / 0.975 / 1.00 / 1.02 (DET_ML 6–10 / 10–20 / 20–50 / 50–100) with mean z −0.05…−0.07 — scored with the background emldetect was actually given (BG_MAP_EML); the design-level b_true scoring's std z 1.02–1.12, mean z +0.28…+0.47 and the "PSF/flux-scale" explanation are withdrawn (RT-45's triage); L0's pooled 1.45 / 1.52 is a mixture of two populations (the far rows' \(\Delta\)C sits below the plug-in, mean z −1.70 / −4.03; the near rows' 0.89 / 0.96 is conditioned on a fit outcome, not the width given \((\hat{S}, b)\)). The SAS replay (T227/E169, Provisional, 07 R034, R038): emldetect and eboxdetect are bit-identical, only esplinemap changes; the truth-background L1/J production needs no version caveat, while the esplinemap arms carry one. Q1 (07 R035, bonus): the target is frozen at −9.282 % [−9.350, −9.219]; the pilot's "L2d/\(\mu\) 4.9 %" was a code artefact (10.4 %).

第 16 步 · W3：全局搜索拟合器的条件律（已关闭，双臂，07 R052）Step 16 · W3: The Conditional Law for the Global-Search Fitter (Closed, Two Arms, 07 R052)

• 目标：Goal: 给出 L0（我们的全局-over-支撑拟合器）的条件律：拟合位置可以跑到 31\(\times\)31 盒子里 ecut = 15 px 圆的任何地方，所以给定 \((\hat{S}, b)\) 的 \(\Delta C\) 不是 \(\hat{S}\) 的单函数——W2 量到它是一个两种总体的混合。这是 Q7 剩下的最后一块。Give the conditional law for L0 (our global-over-support fitter): the fitted position can land anywhere inside the centred ecut = 15 px circle of a 31\(\times\)31 box, so \(\Delta C\) given \((\hat{S}, b)\) is not a function of \(\hat{S}\) alone — W2 measured it to be a mixture of two populations. This is the last piece of Q7.
• 输入：Inputs: L0 生产行（virgo，181 块，2,425,396 收敛行，列含 b、position_id、S_true、S_hat、DeltaC、DET_ML、r_hat = 拟合偏移、converged）；recert13 试点设计（17,420 行，Lang3，S_true ~ logU[0.5, 150]）；注册 e114 核在标称位置的插入式（w2_kernel.py 端口）与 S 换算表 s4_s_conversion_by_position.csv（E156）。The L0 production rows (virgo, 181 chunks, 2,425,396 converged rows, columns include b, position_id, S_true, S_hat, DeltaC, DET_ML, r_hat = fitted offset, converged); the recert13 pilot design (17,420 rows, Lang3, S_true ~ logU[0.5, 150]); the registered e114 kernel plug-in at the nominal position (the w2_kernel.py port) and the S-conversion table s4_s_conversion_by_position.csv (E156).
• 方法：Method: 四项：(1) 推导给定 \((\hat{S}, b, \hat{r})\) 的条件律——支撑截断的模板（把 p 换成在 \((\hat{x}, \hat{y})\) 处渲染并限制在中心圆内的模板，说明那里 \(\hat{S}\) 怎么定义），按 \(\hat{r}\) 分箱（0–2、2–5、5–10、10–15 px）\(\times\) 典型 DET_ML 测试；(2) 给定 \((\hat{S}, b, \hat{r})\) 的律在全部行上测，再给出只看 \((\hat{S}, b)\) 的混合律，权重 \(P(\hat{r} | \hat{S}, b)\) 由源总体 \(\pi(S_true)\) 经贝叶斯导出；(3) 解释近处行 std 0.89；(4) 说明用户只知道 \((\hat{S}, b, \hat{x}, \hat{y})\) 与只知道 \((\hat{S}, b)\) 时各用哪条律。Four items: (1) derive the law given \((\hat{S}, b, \hat{r})\) — the support-truncated template (replace p by the template rendered at \((\hat{x}, \hat{y})\) and restricted to the centred circle, stating how \(\hat{S}\) is defined there), tested per \(\hat{r}\) bin (0–2, 2–5, 5–10, 10–15 px) \(\times\) typical DET_ML; (2) test the \((\hat{S}, b, \hat{r})\) law on all rows, then give the \((\hat{S}, b)\) law as a mixture with weight \(P(\hat{r} | \hat{S}, b)\) derived from the source population \(\pi(S_true)\) by Bayes; (3) explain the near rows' std 0.89; (4) say which law a user should apply when only \((\hat{S}, b, \hat{x}, \hat{y})\) are known and when only \((\hat{S}, b)\) are.
• 成败判据：Success/Failure Test: 每个公式附 Python 函数与测试脚本，打印逐格的模拟与预测 mean z、std z 带 position-cluster bootstrap 区间。成功：每个 \(\geq\) 10,000 行的格 |mean z| < 0.02 且 |std z − 1| < 0.02，外加 recert13 的 6–20 窗口；否则带数字报失败。Every formula ships as a Python function with a test script printing, per cell, the simulated and predicted mean z and std z with a position-cluster bootstrap interval. Success: |mean z| < 0.02 and |std z − 1| < 0.02 in every cell with \(\geq\) 10,000 rows, plus the recert13 6–20 window; anything else is a reported failure with numbers.
• 时间：Time: 12:54Z 起在跑（codexnim + Hermes 双臂，HERMES_HOME 隔离）；生产分析在 virgo（先过 guard），Lang3 只做小检查。Running since 12:54Z (codexnim + Hermes, two arms, isolated HERMES_HOME); production analysis on virgo (guard first), Lang3 for small checks only.
• 先读：Read First: .agent/jobs/2026-10-17-detml-width-globalmle/BRIEF.md（含 15:41Z 的 ERRATA：recert13 的 1.015 / 0.997 是按观测 DET_ML 分箱的，按典型 DET_ML 是 std 1.34 / 1.29、远行 28 % / 4.7 %；"支撑截断"只是 r \(\gtrsim\) 12 px 那部分）→ reviews/RT-41_triage_claude.md → 07 R039。.agent/jobs/2026-10-17-detml-width-globalmle/BRIEF.md (including its 15:41Z ERRATA: recert13's 1.015 / 0.997 came from binning by observed DET_ML, and by typical DET_ML it is std 1.34 / 1.29 with far fractions 28 % / 4.7 %; "support truncation" is only the r \(\gtrsim\) 12 px part) → reviews/RT-41_triage_claude.md → 07 R039.

第 17 步 · W5：给定 \((\hat{S}, \hat{r})\) 的条件宽度——用精确导数重推 \(\sigma_{ratio}\)（已关闭，双臂，判据未达成）Step 17 · W5: The Conditional Width Given \((\hat{S}, \hat{r})\) — Re-deriving \(\sigma_{ratio}\) with Exact Derivatives (Closed, Two Arms, Key Not Met)

• 目标：Goal: 解释为什么随包的 \(\sigma_{ratio}\) 在 L0 近行通过而在玩具核过校正。三件事：(1) 用拟合器模板族的精确导数重推 \(\sigma_{ratio}\)（振幅条件化的位置曲率方差修正）并放到玩具核上；(2) 推导玩具的 Var[\(\Delta C\) | \(\hat{S}\), b, r̂]（W5 已关闭、判据未达成；"核上 2–4 % 偏窄，T233"这个固定 \(S_{true}\) 格子的读数已被 R059 取代：按可观测 \(\hat{S}\) 分箱后 DET_ML 1–3 的 std z 是 0.82–0.86），说明每一项属于哪个条件化；(3) 在 E171 的行上用 np.gradient 与精确导数各算一次 \(\sigma_{ratio}\)，判定哪个对，并检查修正项是否跟上 6–10 内部的 r̂ 趋势（0.88 \(\to\) 0.83）。最后给出用户该用哪个宽度公式及其适用域。Explain why the shipped \(\sigma_{ratio}\) passes L0 near rows but over-corrects the toy core. Three things: (1) re-derive \(\sigma_{ratio}\) (the amplitude-conditioned position-curvature variance correction) with the fitter family's exact derivatives and apply it to the toy core; (2) derive the toy's Var[\(\Delta C\) | \(\hat{S}\), b, r̂] (W5 is closed and the key was not met; the "2–4 % core narrowing, T233" fixed-\(S_{true}\) reading is superseded by R059: binned by the observable \(\hat{S}\), the std z at DET_ML 1–3 is 0.82–0.86), stating which conditioning each term belongs to; (3) compute \(\sigma_{ratio}\) with np.gradient and with exact derivatives of the L0 template on E171's rows, say which is right, and check whether the corrected term follows the r̂ trend inside 6–10 (0.88 \(\to\) 0.83). State which width formula a user should apply and its domain.
• 输入：Inputs: 两臂已有的玩具行（只读）：/u/Lang2/ai_tmp/hermes/detml-width-toy2/out/cell_*.npz 与 /u/Lang2/ai_tmp/codexnim/detml-width-toy2/cell_*.npz；E171 的 L0 行（recert13_mockconv.npz、prod_item1.npz）；lib/beyond_fisher.py、lib/conditional_mean_correction.py。The existing toy rows of both arms (read-only): /u/Lang2/ai_tmp/hermes/detml-width-toy2/out/cell_*.npz and /u/Lang2/ai_tmp/codexnim/detml-width-toy2/cell_*.npz; E171's L0 rows (recert13_mockconv.npz, prod_item1.npz); lib/beyond_fisher.py, lib/conditional_mean_correction.py.
• 成败判据：Success/Failure Test: 玩具核 r̂ \(\leq\) 4 px 五个格 |std z − 1| < 0.02 且不破坏 DET_ML \(\geq\) 24 两格；L0 每个 N \(\geq\) 10,000 生产格近行 std z 落在 1 的 0.02 内；同一函数；常数不许在判据格上调。达不到就带数字报失败。|std z − 1| < 0.02 on the toy core r̂ \(\leq\) 4 px in all five cells without breaking the two DET_ML \(\geq\) 24 cells; L0 near-row std z within 0.02 of 1 in every production cell with N \(\geq\) 10,000; the same function; no constant tuned on the key cells. Anything else is a reported failure with numbers.
• 时间：Time: 2026-10-18T06:58:41Z 发出（codexnim + Hermes 两臂，红队 RT-53 同时攻击 T231 + T233）。Launched 2026-10-18T06:58:41Z (codexnim + Hermes, two arms; red team RT-53 attacks T231 + T233 in parallel).
• 结果（07 R054、R055、R057，E173）：Result (07 R054, R055, R057, E173): 判据未达成。用拟合模板族的精确位置导数后 \(\sigma_{ratio}\) 在 L0 近行 DET_ML 6–10 过度修正——std z 1.67（codexnim 臂）/ 1.83（Hermes 臂），生产子样本给 6.95；在玩具核上也过度修正（1.05–2.18）。RT-54 给出决定性证据（R057）：只有 np.gradient 的 3 点 h = 1 模板通过，5 点模板给 1.31、h = 2 给 0.81、精确导数给 1.68——原先"18/19 格通过"是离散化误差的偶然抵消。所以 \(\sigma_{ratio}\) 是一个经验校准，不是推导出来的宽度，T230 的 L0 近行通过因此标为 Questioned。玩具核的方差律同样未达成。后续：给低计数偏窄一个尺寸预测（R060 之后仍缺）。The key is not met. With the fitter template family's exact position derivatives \(\sigma_{ratio}\) over-corrects L0 near rows at DET_ML 6–10 — std z 1.67 (codexnim arm) / 1.83 (Hermes arm), a production subsample giving 6.95 — and over-corrects the toy core too (1.05–2.18). RT-54 gave the decisive evidence (R057): only np.gradient's 3-point h = 1 stencil passes, while the 5-point stencil gives 1.31, h = 2 gives 0.81 and the exact derivatives give 1.68 — the old "18 of 19 cells pass" was an accidental cancellation of a discretisation error. So \(\sigma_{ratio}\) is an empirical calibration, not a derived width, and T230's L0 near-row pass is therefore marked Questioned. The toy core's variance law is likewise not met. Follow-up: a size prediction for the low-count narrowing (still missing after R060).
• 先读：Read First: .agent/jobs/2026-10-18-detml-width-w5-sigma-ratio/BRIEF.md → cognition-os/experiments/E172_w4_toy_second_order.md（R052 修正）→ E171_w3_globalmle_conditional_law.md（R050、R052 修正）→ reviews/W4_hermes_claude.md → 07 R052。.agent/jobs/2026-10-18-detml-width-w5-sigma-ratio/BRIEF.md → cognition-os/experiments/E172_w4_toy_second_order.md (the R052 amendment) → E171_w3_globalmle_conditional_law.md (the R050 and R052 amendments) → reviews/W4_hermes_claude.md → 07 R052.
• 开放检查（R048，无执行者）：Open check (R048, no executor yet): 掩膜位置自由玩具，插入式在拟合模板处取值，N = 10,000，带与不带 \(\delta_0\)，含脆弱的 DET_ML \(\approx\) 7–15 格。The masked position-free toy with the plug-in at the fitted template, N = 10,000, with and without \(\delta_0\), including the fragile DET_ML \(\approx\) 7–15 cells.

第 18 步 · 真实 PSF 宽度配料研究（08 §7 item 0，仍开放）Step 18 · The real-PSF width ingredient study (08 §7 item 0, still open)

• 目标：Goal: 玩具的可观测偏窄到 DET_ML \(\approx\) 12 就消失，真实 PSF 的到 \(\approx\) 20 才消失（R062）。找出是哪个配料把交叉点从 \(\approx\) 12 推到 \(\approx\) 20，再给 DET_ML 6–20 一个可观测修正。The toy's observable narrowing is gone by DET_ML \(\approx\) 12 while the real PSF's lasts to \(\approx\) 20 (R062). Find which ingredient moves the crossover from \(\approx\) 12 to \(\approx\) 20, then give an observable correction for DET_ML 6–20.
• 输入：Inputs: 同一套构造（M31 真实 PSF 图、cfg_base_script 的 ecut 15 px 掩膜与曝光图、样条模板族、全局网格 + NM、位置自由 \(\nu\) = 3、总体 S_true ~ dN/dS \(\propto\) S^−2.5 于 [5.58, 51.20]）。The same construction (the M31 real-PSF stamp, cfg_base_script's ecut 15 px mask and exposure map, the spline template family, a global grid + NM, position free with \(\nu\) = 3, a population S_true ~ dN/dS \(\propto\) S^−2.5 on [5.58, 51.20]).
• 方法：Method: 一次只换一个配料——同 HPD 的高斯 PSF（同一掩膜/曝光图内）、平曝光、无掩膜、位置固定——每个变体约 25,000 行，记录 2-SE 交叉点落在哪儿；再在留出的源位置上测一个可观测修正因子。Swap one ingredient at a time — a Gaussian PSF of the same HPD (inside the same mask/exposure map), flat exposure, no mask, position fixed — about 25,000 rows per variant, recording where the 2-SE crossover lands; then test an observable correction factor on a held-out source position.
• 成败判据：Success/Failure Test: 至少一个变体把交叉点从 \(\approx\) 20 显著挪开（超出该变体自己的 SE），从而定位成因；若四个变体都不动，则成因不在这些配料里，要另找。At least one variant moves the crossover significantly away from \(\approx\) 20 (beyond that variant's own SE), locating the cause; if none of the four moves it, the cause lies outside these ingredients and must be sought elsewhere.
• 时间：Time: virgo 上约 170 ms/行（RT-58 实测 170.8 / 171.1 ms/行），每个变体约 25,000 行。About 170 ms/row on virgo (RT-58 measured 170.8 / 171.1 ms/row), about 25,000 rows per variant.
• 先读：Read First: R062 的分诊与复算，及其报告 §2 的分箱表（见来源）。R062's triage and replays, and §2 of its report (the bin table, see the sources)。
• 2026-10-19 更新（RT-60 已分诊，R064、R068）：Update of 2026-10-19 (RT-60 is triaged, R064, R068): 机制已定位但不是一条可用的修正：公式在真实参数处是对的（固定 S_true 与真实位置时残差的 std 除以 \(\sqrt{V_{3p}(S_{true})}\) 是 1.12 / 1.07 / 1.03 / 1.01 / 1.03，DET_ML 3 / 5 / 9 / 15 / 30），偏窄来自把凸的 \(V_{3p}\) 放在偏高的 \(\hat{S}\) 上计算（DET_ML 3 时 Ŝ 高 +14.5 %）。但 R064 提示的修正被 RT-63 推翻（R068）：在可观测的 Cox–Snell 偏差校正振幅 \(\hat{S} - b_S(\hat{S})\) 上求值 \(V_{3p}\)，DET_ML 6.9 / 9 / 12 / 18.5 的 std z 只从 0.88 / 0.92 / 0.94 / 0.96 升到 0.975 / 0.959 / 0.970 / 0.979（消掉 45–79 %），回不到 1 \(\pm\) 0.02，所以 08 §7 第 0a 项以否定结论关闭。本步仍然开放：换配料的研究要回答的是"为什么真实 PSF 的偏窄持续到 DET_ML \(\approx\) 20 而玩具的到 \(\approx\) 12 就消失"，这与"偏差校正在哪里失效"是同一个问题的两面。RT-65（逐行 \(\delta\) 约定）与 RT-66（残差的先验依赖）已分诊（R070、R069），RT-68 也已分诊（R071），RT-67 也已分诊（R072：更好的偏差修正路线关闭，残差与先验无关（\(\leq\) 0.6 SE），机制已改写为 \(V_{3p}\) 的近乎全抵消（R077/R078）），RT-69 也已分诊（R073：目录区间的净过覆盖只有 +2–5 pp，目标改为区间覆盖率误差 \(\leq\) 2 pp）；一支也已分诊（R074：0.5\(\times\) 闭合已废弃，不存在随背景变化的闭合）；z_off 是 L0 打分（R076），机制是 Fisher penalty 错配被 \(V_{3p}\) 抵消放大（R077/R078）；L0 的低计数中心项超额是 L0 流水线自己的性质（R082/R084/R085；72/28 归因经 R087 撤回，成因开放；\(\hat{S}\) 分箱 +0.389 / +0.14–0.18 / +0.058 \(\pm\) 0.007）；候选的二阶 penalty 修正未达成（R080），但该机制也解释 L1 宽度（R086 推翻 R083：合并估计量下 L1 错配 +1.19 % [+1.03, +1.35]，解释箱内宽度缺口的 0.89 [0.84, 0.93]）；主线是 emldetect 条件律（PI 2026-10-20 04:51 EDT）：任务 0b 已完成（R089，已复算）——emldetect 与 L0 高计数一致、DET_ML ≲ 7 处中心低 0.16–0.53 z、宽 0.05–0.14，所以 L0 低计数超额不是 emldetect 的性质；RT-84/RT-85/RT-86 已分诊（R091、R088/R090），T-LAW 在跑。The mechanism is located but it is not a usable correction: the formula is right at the true parameters (at fixed S_true and the true position the residual's std divided by \(\sqrt{V_{3p}(S_{true})}\) is 1.12 / 1.07 / 1.03 / 1.01 / 1.03 at DET_ML 3 / 5 / 9 / 15 / 30) and the narrowing comes from evaluating the convex \(V_{3p}\) at a biased-high \(\hat{S}\) (+14.5 % at DET_ML 3). But the correction R064 suggested is refuted by RT-63 (R068): evaluating \(V_{3p}\) at the observable Cox–Snell bias-corrected amplitude \(\hat{S} - b_S(\hat{S})\) lifts std z at DET_ML 6.9 / 9 / 12 / 18.5 only from 0.88 / 0.92 / 0.94 / 0.96 to 0.975 / 0.959 / 0.970 / 0.979 (45–79 % of the deficit removed) and not to 1 \(\pm\) 0.02, so 08 §7 item 0a is closed negative. This step stays open: the ingredient study asks why the real PSF's narrowing lasts to DET_ML \(\approx\) 20 while the toy's is gone by \(\approx\) 12, which is the other face of "where does the bias correction fail". RT-65 (the per-row \(\delta\) convention) and RT-66 (the residual's prior dependence) are triaged (R070, R069); RT-68 is triaged too (R071); RT-67 is triaged too (R072: the better-bias-correction route is closed, the residual is prior-independent (\(\leq\) 0.6 SE) and the mechanism is rewritten as \(V_{3p}\)'s near-total cancellation (R077/R078)); RT-69 is triaged too (R073: the catalogue interval's net over-coverage is only +2–5 pp, and the goal is re-stated as interval-coverage error \(\leq\) 2 pp); one is triaged too (R074: the 0.5\(\times\) closure is deprecated and no background-dependent closure exists); z_off is the L0 scoring (R076) and the mechanism is the Fisher-penalty mismatch amplified by the \(V_{3p}\) cancellation (R077/R078); the L0 low-count centre excess is a property of the L0 pipeline itself (R082/R084/R085; the 72/28 attribution withdrawn by R087, cause open; \(\hat{S}\)-bin gaps +0.389 / +0.14–0.18 / +0.058 \(\pm\) 0.007); the candidate second-order-penalty fix is not achieved (R080), but the mechanism does explain the L1 width too (R086 reverses R083: with pooled estimators the L1 mismatch is +1.19 % [+1.03, +1.35], explaining 0.89 [0.84, 0.93] of the within-cell width deficit); the main line is the emldetect conditional law (PI 2026-10-20 04:51 EDT): task 0b is completed (R089, replayed) — emldetect and L0 agree at high counts and at DET_ML ≲ 7 emldetect's centre is 0.16–0.53 z lower and its width 0.05–0.14 larger, so the L0 low-count excess is not an emldetect property; RT-84/RT-85/RT-86 are triaged (R091, R088/R090) and T-LAW is running.

第 19 步 · L0 对 L1 的决定性实验：在真实 PSF 图像上跑 emldetect（08 §7 item 0b；已完成，R089：L0 超额不是 emldetect 的性质）Step 19 · The decisive L0-vs-L1 run: emldetect on the real-stamp images (08 §7 item 0b; completed, R089: the L0 excess is not an emldetect property))

• 目标：Goal: 把 L0 与 L1 的宽度差里的混淆因素一个一个去掉。R067 发现：同 67 位置、同 S^−2.5 密度、同 (S, b) 格子权重下，L1 比 L0 宽 +0.24 / +0.19 / +0.12 / +0.08 / +0.07 于 DET_ML \(\approx\) 2 / 3 / 4 / 5 / 7（约 50 处持平）；但两边用的是不同流水线加不同打分（真实 PSF 模板族在拟合位置 对 e114 替代核在名义位置）与不同振幅估计量，PSF 与拟合器的贡献用现有产品分不开。本步要留下拟合器这唯一的差别。Remove the confounders in the L0-vs-L1 width gap one at a time. R067 found: at the same 67 positions, the same S^−2.5 density and the same (S, b) cell masses, L1 is wider than L0 by +0.24 / +0.19 / +0.12 / +0.08 / +0.07 at DET_ML \(\approx\) 2 / 3 / 4 / 5 / 7 (level by about 50); but the two sides use different pipeline + scoring pairs (the stamp family at the fitted offset against the e114 surrogate at the nominal position) and different amplitude estimands, so the PSF and fitter contributions cannot be separated with the current products. This step leaves the fitter as the only difference.
• 输入：Inputs: L0 生产的真实 PSF 图像（同 67 位置、同 S/b 网格、带背景图）——先确认这些图像产物还在（sirius / virgo）再写任务说明；一个带位置轴的真实 PSF 核网格。The L0 production real-stamp images (the same 67 positions, the same S/b grid, with background maps) — check that these image products still exist (sirius / virgo) before briefing; a real-stamp kernel grid that has an offset axis.
• 方法：Method: 在这批图像上跑 emldetect（SAS，virgo），然后用同一个真实 PSF 核分别按名义位置与 emldetect 拟合位置打分。同一批像素、同一个核，所以 PSF 与计算点的混淆被去掉，只剩拟合器一个差别。Run emldetect on those images (SAS, virgo), then score with the same real-stamp kernel at both the nominal position and the emldetect-fitted position. Same pixels, same kernel, so the PSF and evaluation-point confounds are removed and only the fitter differs.
• 成败判据：Success/Failure Test: 在匹配的 (S, b) 格子与先验下，两个位置的 std z 差带区间报出；若名义位置打分显著宽于拟合位置打分，则 R067 观察到的差主要来自计算点而不是拟合器；若两者都宽，则拟合器或核本身有差。Report the std z difference between the two positions with an interval, at matched (S, b) cells and prior; if the nominal-position scoring is significantly wider than the fitted-position scoring, R067's gap comes mainly from the evaluation point rather than the fitter; if both are wide, the fitter or the kernel itself differs.
• 时间：Time: 对现有图像集跑一遍 emldetect + 一个核网格的成本（SAS 在 virgo，成本由图像数主导；写说明前先确认图像还在）。One emldetect pass over the existing stamp set plus a kernel grid (SAS on virgo, cost driven by the image count; confirm the images still exist before briefing).
• 先读：Read First: reviews/RT-64_triage_claude.md（O1/O2/O4 的匹配比较与分解）→ 07 R067 → 08_RUNNING.md §7 第 0b 项。reviews/RT-64_triage_claude.md (the matched comparison and decomposition in O1/O2/O4) → 07 R067 → 08_RUNNING.md §7 item 0b.
• 第 20 步 · L1 宽度的背景图与估计量（已完成，R086 推翻 R083：合并估计量下 L1 错配 +1.19 % [+1.03, +1.35]）Step 20 · The L1 width: background map and estimator (done, R086 reverses R083: with pooled estimators the L1 mismatch is +1.19 % [+1.03, +1.35])
  • 目标：Goal: 判定 L1（emldetect）宽度亏损的背景图约定（BG_MAP_EML 与 b_true）与汇总（非逐格）错配估计量是否改变 R083 的结论。Decide whether the L1 (emldetect) width deficit depends on the background-map convention (BG_MAP_EML vs b_true) and whether the pooled (not cell-averaged) mismatch estimator changes R083's conclusion.
  • 输入：Inputs: L1 生产图像与其背景图；R083 的复现脚本（rt79_m5b.py）。The L1 production images with their background maps; R083's replay script (rt79_m5b.py).
  • 方法：Method: 同一批图像在两种背景下打分；错配用汇总估计量重算（不逐格平均）。Score the same images under both backgrounds; recompute the mismatch with the pooled estimator (no cell averaging).
  • 成败判据：Success/Failure Test: 两种约定下 L1 宽度差的配对区间报出；若含 0 则写 持平。Report the paired interval of the L1 width difference under both conventions; if it contains 0, write “no difference”.
  • 时间：Time: 一个红队会话（约 2 小时）。One red-team session (about 2 h).
  • 先读：Read First: reviews/RT-79_triage_claude.md（O4 的 BG_MAP_EML 延期项）→ 07 R083。reviews/RT-79_triage_claude.md (O4, the deferred BG_MAP_EML item) → 07 R083.
• 第 21 步 · 0b 试点后的验收清单与 emldetect 条件律表（专家 P1，已接受；已完成：P1 清单与六支规律任务 R093–R103）Step 21 · After the 0b pilot: the acceptance manifest and the emldetect conditional-law table (expert P1, accepted; done: the P1 manifest and the six law tasks R093–R103)
  • 目标：Goal: （甲）一份只读的验收清单（T1/T2/T4 的哈希、行数、缺失/重复、版本、支撑、先验、分箱、估计量，逐门 PASS/FAIL）；（乙）一张声明参数的 emldetect 中心+宽度/CDF 候选表，留出集检验，无修正基线。(a) A read-only acceptance manifest (hashes, row counts, absent/duplicate rows, versions, support, prior, binning, estimator for T1/T2/T4, PASS/FAIL per gate); (b) a declared-parameter emldetect centre+width/CDF candidate table, held-out, with a no-correction baseline.
  • 输入：Inputs: 0b 的报告与图像清单；T1/T2/T4 的现有产物。The 0b report and image inventory; the existing T1/T2/T4 products.
  • 方法：Method: 清单只读、免费臂可做；条件律表按 GOAL 允许的声明参数形式给出。The manifest is read-only (a free arm can do it); the conditional-law table takes the declared-parameter form the GOAL allows.
  • 成败判据：Success/Failure Test: 清单逐门给出 PASS/FAIL/证据不足；条件律表在留出集上给出带区间的数字。The manifest gives PASS/FAIL/insufficient per gate; the conditional-law table gives held-out numbers with intervals.
  • 时间：Time: 清单约半天；条件律表视 0b 规模。The manifest about half a day; the table depends on the 0b scale.
  • 先读：Read First: reviews/EXPERT_20261020_astra_triage_claude.md（P1 两条）→ 08_RUNNING.md §5。reviews/EXPERT_20261020_astra_triage_claude.md (the two P1 items) → 08_RUNNING.md §5.
• 第 22 步 · 规律阶段：四条候选与一支验证过的评分器（已完成，R093–R104）Step 22 · The law stage: four candidates and one validated scorer (done, R093–R104)
  • 目标：Goal: 在 L1 emldetect 输出（M2 能段 4）上检验 PI 主线规律 \(p(\mathrm{DET\_ML} \mid \hat{S}, b, \mathrm{index})\) 的候选形式，并先有一支正确的评分器。Test candidate forms of the PI main-line law \(p(\mathrm{DET\_ML} \mid \hat{S}, b, \mathrm{index})\) on L1 emldetect output (M2 band 4), with a correct scorer first.
  • 输入：Inputs: L1 行表（1,066,923 行，67 位置）、拟合/留出半分（种子 20261020）、common.py/eval_lib.py。The L1 row table (1,066,923 rows, 67 positions), the fit/held half split (seed 20261020), common.py/eval_lib.py.
  • 方法：Method: 四个候选（C 插入式、D 制表 288 参数、DQ 二次 index、MIX3d 生成式混合）在 72 个留出格上打分；零假设先在"恰好为真"的数据上验证。Four candidates (C plug-in, D tabulated 288 parameters, DQ quadratic index, MIX3d generative mixture) scored on 72 held-out cells, with the null validated on exactly-true data first.
  • 成败判据：Success/Failure Test: 通过数高于完美规律的零假设第 5 百分位，且全局 Q 的 p 值高于 0.01。A pass count above the perfect law's null 5th percentile and a global-Q p-value above 0.01.
  • 时间：Time: 六个任务 + 六支红队，一天内完成并被复核。Six tasks plus six red teams, all finished and re-checked within one day.
  • 先读：Read First: 07 R093–R104 → .agent/jobs/2026-10-20-task-null-1/REPORT.codexnim.md（评分器）→ reviews/RT-84_triage_claude.md、reviews/RT-87_triage_claude.md。07 R093–R104 → .agent/jobs/2026-10-20-task-null-1/REPORT.codexnim.md (the scorer) → reviews/RT-84_triage_claude.md, reviews/RT-87_triage_claude.md.
  • 结果：Result: 四条候选全部被拒——能段 4 上 C 10/72、D 48/72、DQ 42/72、MIX3d 35/72，能段 1 上 D 重拟合 37/72（零假设均值约 62–63/72；D 的 p_pass = 0.0015）。这些通过数自此是历史诊断：2026-10-21 的专家复核（R114，已采纳）指出这个打分手混用三个估计对象——宽度是位置内加权 std，均值与覆盖率跨位置合并，零假设还跟着候选自适应。RT-98（R117，已复算）证明宽度指标对逐位置偏移恒等不变（1e-15）：一条带 0.30 z 逐位置偏移的规律仍通过宽度 70/72，而 cov68 失败从 4 涨到 29。所以"宽度一半可推广"只能读作"在同一个位置内可推广"。新打分手（SCORER-CONTRACT v2，R118，已复算）分开两个目标：未见位置总体上 D / H2 / MIX3d / C 通过 44 / 50 / 32 / 12 格（共 72，能段 4），能段 1 的 D / H2 为 36 / 43，全部被 Q 拒（p \(\leq\) 0.009）；R124 限定：这是一次不稳定的抽样——14 个位置分半上 D 22–46、H2 26–52（能段 4），能段 1 的 D 14–53、H2 30–52；H2 优于 D 在固定分半上不显著（McNemar p 0.18 / 0.25），所以不能作"H2 最好"的排名，任何规律排名之前先要 scorer v3；逐位置上每条规律都大幅失败（能段 4 为 1 / 0 / 1 / 0，能段 1 为 0 / 1），冻结位置效应 \(\tau_{mean}\) 0.14 z（能段 4）/ 0.18 z（能段 1），重尾 df \(\approx\) 3。中心项的主导成分是 D 表自身的样本内表征误差（R116：同一位置内失败 24 \(\to\) 3、未见位置 21 \(\to\) 8，72 参数诊断修正），不是位置泛化缺口。宽度读数也被限定（R115）：0.976 的留出宽度是一次偏低的分半，40 个分半平均 0.998；6 参数修正只通过 25/72 个宽度格。Every candidate is rejected — on band 4 C 10/72, D 48/72, DQ 42/72 and MIX3d 35/72, and on band 1 the D refit 37/72 (null mean about 62–63/72; D's p_pass = 0.0015). Those pass counts are historical diagnostics from here on: the 2026-10-21 expert review (R114, accepted) showed the scorer mixed three estimands — the width is a within-position weighted std, the mean and coverage pool rows across positions, and the null adapts to the candidate. RT-98 (R117, replayed) proved the width metric is identically invariant (1e-15) to per-position offsets: a law carrying a 0.30 z per-position offset still passes width 70/72, while cov68 failures rise from 4 to 29. So "the width half generalises" can only be read as "the width half generalises within one position". The new scorer (SCORER-CONTRACT v2, R118, replayed) separates two targets: on the population of unseen positions D / H2 / MIX3d / C pass 44 / 50 / 32 / 12 of 72 cells (band 4) and band-1 D / H2 36 / 43, all rejected on Q (p \(\leq\) 0.009); R124 qualifies: these are one draw of an unstable statistic — over 14 position splits D 22–46 and H2 26–52 (band 4), band-1 D 14–53 and H2 30–52; H2 above D is not significant on the fixed split (McNemar p 0.18 / 0.25), so no "H2 is best" ranking can be made, and scorer v3 is needed before any law ranking; per position every law fails by a wide margin (band 4: 1 / 0 / 1 / 0; band 1: 0 / 1), with the frozen position effect \(\tau_{mean}\) 0.14 z (band 4) / 0.18 z (band 1), heavy-tailed with df \(\approx\) 3. The centre term is dominated by an in-sample representation error of the D table itself (R116: within position 24 \(\to\) 3 failures, at unseen positions 21 \(\to\) 8 with a 72-parameter diagnostic correction), not a position-generalisation gap. The width reading is qualified too (R115): the 0.976 held-out width was one low split, 0.998 averaged over 40 splits, and the 6-parameter correction passes only 25/72 width cells.
• 第 23 步 · 符号回归试点：中心残差面的短解析式（已完成，R109：没有公式配得上 D）Step 23 · The symbolic-regression trial: a short analytic form for the centre residual surface (done, R109: no formula matches D)
  • 目标：Goal: 一个短解析式 \(f(\ln \hat{S}, \ln b, \mathrm{index})\) 能否替代 D 的 288 参数中心表，并迁到 M2 能段 1。Can a short analytic form \(f(\ln \hat{S}, \ln b, \mathrm{index})\) replace D's 288-parameter centre table and transfer to M2 band 1?
  • 输入：Inputs: L1 行表、D 的节点场、NULL-1 评分器。The L1 row table, D's node fields, the NULL-1 scorer.
  • 方法：Method: 两条臂：PySR（Julia，无 LLM）与 SRHarness（LLM 后端 = NIM 模型）；拟合半训练、留出打分、再测迁移。Two arms: PySR (Julia, no LLM) and SRHarness (LLM backend = a NIM model); train on the fit half, score held-out, then test transfer.
  • 成败判据：Success/Failure Test: 留出通过数高于 D 的 48/72，且迁移后不低于 D 的 17/72。A held-out pass count above D's 48/72 and a transfer no worse than D's 17/72.
  • 时间：Time: 两个红队会话。Two red-team sessions.
  • 先读：Read First: 04 2026-10-20 18:47 EDT（PI 批准）→ 07 R103 → .agent/jobs/2026-10-20-task-sr-pysr/BRIEF.md。04, 2026-10-20 18:47 EDT (PI approval) → 07 R103 → .agent/jobs/2026-10-20-task-sr-pysr/BRIEF.md.
  • 结果：Result: 没有公式配得上 D（R109，RT-97 已复算）：PySR（3 个种子 + r 特征）与 SRHarness（NIM 后端）都找不到；最好的 SRHarness 三常数式在 b/Ŝ 上给 28/72（能段 4，D 48），原样迁到能段 1 是 22 而 D 自己的能段 1 重拟合是 37（约 60 %）；只重拟合常数会让每个公式都更差。目标、评分器、只用拟合半三点都已确认；拟合半的增益不迁移；"index 无关所以 \(\beta\) 不必要"被驳回（\(\beta\) = 0 把 D 从 48 打到 24/72）。No formula matches D (R109, RT-97 replayed): neither PySR (3 seeds plus the r feature) nor SRHarness (NIM backend) finds one; the best SRHarness three-constant form in b/Ŝ gives 28/72 on band 4 (D 48) and transfers unchanged to band 1 at 22 against D's own band-1 refit of 37 (about 60 %), and refitting constants only makes every formula worse. Target, scorer and fit-half-only are all confirmed; the fit-half gain does not transfer; and "index-free, so \(\beta\) unnecessary" is refuted (\(\beta\) = 0 takes D from 48 to 24/72).
• 第 24 步 · 新打分手：未见位置总体与逐位置两个目标，以及 PI 的进度指标（评分器已完成，R118；基线点已到：R120 的已解决位置比例 4/33）Step 24 · The new scorer: the unseen-position population and the per-position targets, and the PI's progress metric (the scorer is done, R118; the baseline point is in: R120's solved-position fraction 4/33)
  • 目标：Goal: 把"这个规律好不好"拆成两个可辩护的目标，并给出 PI 要的读数：已解决位置的比例（与面积），可以从 0 开始增长。Split "is this law good" into two defensible targets and produce the reading the PI wants: the solved fraction (and area) of positions, which may grow from 0.
  • 输入：Inputs: L1 行表（1,066,923 行，67 位置）、SCORER_CONTRACT.md 与 score_v2.py（.agent/jobs/2026-10-21-task-scorer-contract/）、RT-98 的五条异议。The L1 row table (1,066,923 rows, 67 positions), SCORER_CONTRACT.md and score_v2.py (.agent/jobs/2026-10-21-task-scorer-contract/), and RT-98's five objections.
  • 方法：Method: 两个目标分开打分：POP-UNOBS（未见位置总体，边际宽度目标 \(\sqrt{1+\tau^2}\)，冻结的非自适应零假设，跨格相关）与 POS-CAL（逐位置，精确高斯零假设，支撑下限 neff_pc \(\geq\) 100、n_c \(\geq\) 8）。Monte-Carlo p 一律写成 k/N 加一个界。Score the two targets separately: POP-UNOBS (the unseen-position population, marginal width target \(\sqrt{1+\tau^2}\), a frozen non-adaptive null, cross-cell correlation) and POS-CAL (per position, an exact Gaussian null, support floors neff_pc \(\geq\) 100 and n_c \(\geq\) 8). Every Monte-Carlo p is written as k/N with a bound.
  • 成败判据：Success/Failure Test: PI 2026-10-21 的目标（05:39 EDT 定，11:42 EDT 修订）：域 = 离轴 \(\theta \leq \theta_{max}\) 且 MASKFRAC \(\geq m_{min}\)；一个位置通过 = 其中心 / 宽度 / 覆盖率逐位置都通过；修订后位置上没有通过/不通过阈值——进度指标 = 已解决位置的比例（与面积），可以从 0 开始，任何增长都算进展，90 % 默认值撤回；每条规律按已解决位置比例（按离轴半径 / MASKFRAC 分层的面积加权比例）加区间报告并随时间跟踪。The PI's target of 2026-10-21 (set 05:39 EDT, amended 11:42 EDT): domain = off-axis \(\theta \leq \theta_{max}\) and MASKFRAC \(\geq m_{min}\); a position passes if its centre / width / coverage all pass per position; after the amendment there is no pass/fail threshold on positions — the progress metric is the solved fraction (and solved area) of positions, which may start at 0, where any increase is progress and the 90 % default is withdrawn; every law is reported as a solved-position fraction (and an area-weighted fraction, stratified by off-axis radius / MASKFRAC) with its interval, tracked over time.
  • 时间：Time: 评分器一个任务日；域扫描一支红队会话。The scorer took one task day; the domain sweep is one red-team session.
  • 先读：Read First: 07 R114（专家 astra 的报告）→ .agent/jobs/2026-10-21-task-scorer-contract/REPORT.codexnim.md → .agent/jobs/2026-10-21-redteam-rt98-scorer-contract/REPORT.hermes.md → 04_DECISION_LOG.md 2026-10-21 05:39 EDT。07 R114 (the astra expert's report) → .agent/jobs/2026-10-21-task-scorer-contract/REPORT.codexnim.md → .agent/jobs/2026-10-21-redteam-rt98-scorer-contract/REPORT.hermes.md → 04_DECISION_LOG.md 2026-10-21 05:39 EDT.
  • 结果：Result: 两个目标上都没有候选通过。POP-UNOBS：能段 4 的 D / H2 / MIX3d / C 通过 44 / 50 / 32 / 12 格（共 72），能段 1 的 D / H2 为 36 / 43，全部被 Q 拒（p \(\leq\) 0.009）。POS-CAL：能段 4 为 1 / 0 / 1 / 0，能段 1 为 0 / 1，所有 Q_B 加一后 0.0005。冻结位置效应 \(\tau_{mean}\) 0.14 z（能段 4）/ 0.18 z（能段 1），r 0.472 / 0.492，df 3.16 / 3.00（重尾）。RT-98 的 tau 地板塌陷异议已被 RT-100 在最终版上否定（0/18 格在地板，冻结输出逐位复现），但"约 33 个位置无法辨认逐格位置尺度"的结构性结论保留。POS-CAL 的通过数零假设反保守约 2.4–3.3 倍，用 Q_B 更稳。R120（RT-100 已复算，进度曲线的第一个点）：能段 4 上 D / H2 各只有 4/33 个位置的条目通过比例 \(\geq\) 0.90（均值 0.643 / 0.651），能段 1 为 1/33（0.496 / 0.513）；离轴三分位 0.767 / 0.605 / 0.542——失败位置散布（13.75′、9′、10.8′），最好位置里也有 10.8′（1.00），不是视场外缘效应；\(\theta \leq\) 8′ 把能段 4 抬到约 0.75；离轴梯度属于可观测流水线，不属于 D/H2。No candidate passes under either target. POP-UNOBS: band-4 D / H2 / MIX3d / C pass 44 / 50 / 32 / 12 of 72 cells and band-1 D / H2 36 / 43, all rejected on Q (p \(\leq\) 0.009). POS-CAL: band 4 gives 1 / 0 / 1 / 0 and band 1 0 / 1, with every Q_B at 0.0005 plus-one. The frozen position effect \(\tau_{mean}\) is 0.14 z (band 4) / 0.18 z (band 1), r 0.472 / 0.492, df 3.16 / 3.00 (heavy-tailed). RT-98's tau-floor-collapse objection is refuted on the final version by RT-100 (0/18 cells at the floor, bit-exact freeze reproduction), but the structural conclusion stands that roughly 33 positions cannot identify a per-cell between-position scale. The POS-CAL pass-count null is anti-conservative by about 2.4–3.3\(\times\), so Q_B is the calibrated aggregate. R120 (RT-100 replayed, the first point of the progress curve): on band 4 only 4/33 positions have an entry pass fraction \(\geq\) 0.90 for D and for H2 (means 0.643 / 0.651), and 1/33 on band 1 (0.496 / 0.513); off-axis terciles 0.767 / 0.605 / 0.542 — the failing positions are scattered (13.75′, 9′, 10.8′), and the best positions include 10.8′ (1.00), so it is not an outer-field-of-view effect; \(\theta \leq\) 8′ lifts band 4 to about 0.75; the off-axis gradient belongs to the observable pipeline, not to D/H2.
• 第 25 步 · WIDTH-2ND：分解可观测 \(\hat{S}\) 箱的宽度亏缺并推导 \(\mathrm{Var}[\Delta C \mid \hat{\theta}]\) 的二阶修正（已完成，判据 FAIL，R121）Step 25 · WIDTH-2ND: decompose the observable-\(\hat{S}\)-bin width deficit and derive the second-order correction to \(\mathrm{Var}[\Delta C \mid \hat{\theta}]\) (done, verdict FAIL, R121)
  • 目标：Goal: 宽度这一半已经知道为什么错（相消放大机制，R077/R078/R086），但还没有一个闭合的公式。本步给出二阶方差修正，并把亏缺分解成三块。The width half already knows why it is wrong (the cancellation-amplification mechanism, R077/R078/R086) but has no closing formula. This step supplies the second-order variance correction and decomposes the deficit into three parts.
  • 输入：Inputs: results/e088_*_rows.csv 与 results/e089_*（理论匹配的 King 拟合器行）、E088 的 \(\delta_0\)、E090 的 \(\delta_1\)、Li et al. 2025 arXiv:2510.03466 定理 6。results/e088_*_rows.csv and results/e089_* (theory-matched King fitter rows), E088's \(\delta_0\), E090's \(\delta_1\), and Li et al. 2025 arXiv:2510.03466's Theorem 6.
  • 方法：Method: (a) 每个可观测 \(\hat{S}\) 箱（DET_ML 6–40）把宽度亏缺分解为：在偏高的 \(\hat{S}\) 上插入式取值 / 箱内 S_true 混合（用模拟设计权重，说明权重）/ 残差；(b) 推导二阶项（与 E090 同一套展开机器，sympy 允许），说明每个假设，在精确单像素例上验符号；(c) 在 E088/E089 行上检验修正后宽度。(a) Per observable \(\hat{S}\) bin (DET_ML 6–40), decompose the width deficit into plug-in at the biased-high \(\hat{S}\), within-bin S_true mixing (using the simulation design weights, stating the weights) and residual; (b) derive the second-order term (the same expansion machinery as E090, sympy allowed), state every assumption, and check the sign on the exact single-pixel case; (c) test the corrected width on the E088/E089 rows.
  • 成败判据：Success/Failure Test: 修正后宽度在每个 \(\hat{S}\) 箱、DET_ML 6–20 处都在统计误差内等于 1（我们的拟合器）；除不超过 2 个箱外都在 2\(\sigma\) 内算 ambiguous；否则 FAIL，由分解说明原因。The corrected width equals 1 within statistical error in every \(\hat{S}\) bin at DET_ML 6–20 (our fitter); ambiguous if within 2\(\sigma\) in all but at most 2 bins; otherwise FAIL, with the decomposition saying why.
  • 时间：Time: 一个理论任务（两条臂可选）。One theory task (two arms optional).
  • 先读：Read First: 04_DECISION_LOG.md 2026-10-21 06:20 EDT（PI 批准）→ .agent/jobs/2026-10-21-task-width-2nd/BRIEF.md → 07 R064、R068、R077、R078、R086 → cognition-os/experiments/E088、E090。04_DECISION_LOG.md 2026-10-21 06:20 EDT (PI approval) → .agent/jobs/2026-10-21-task-width-2nd/BRIEF.md → 07 R064, R068, R077, R078, R086 → cognition-os/experiments/E088, E090。
  • 结果：Result: 已完成，预注册判据 FAIL（R121，2026-10-21 14:16 EDT，已复算）：宽度亏缺被归因为条件方差 \(\mathrm{Var}[\Delta C \mid \hat{\theta}] < V_{3p}(\hat{\theta})\)——箱内 \(S_{true}\) 混合可忽略且只会增大方差（0.000–0.014），条件均值项同样可忽略（0.000–0.010）。新的 \(O(n^{-1})\) 闭式因子 \(c = 1 - \mathrm{tr}(I^{-1}Q)/(2V_{3p}) - \mathrm{tr}(I^{-1}P)^2/(4V_{3p})\) 与暴力 Edgeworth 展开吻合到小数点后四位；在 \(b = 0.01\) 处精确（对数误差均方根 3.1 %，DET_ML 6–20 的修正宽度在 1 的 1.6\(\sigma\) 内），但在 \(b = 0.04\) 欠修（0.952 \(\pm\) 0.013，−3.6\(\sigma\)）、\(b = 0.16\) 更差（0.926 \(\pm\) 0.012，−6.1\(\sigma\)）——8 个箱里 3 个超出 2\(\sigma\)，判据 FAIL。残差跟随 \(V_{3p}\) 的相消因子 \(4\sum\mu L^2/V_{3p}\)（\(R^2\) 0.89），不跟随 \(n_{\rm eff}\)；猜想是高 \(b\) 处 \(L \approx (S/b)p\) 与 S-score 近平行，\(O(n^{-1})\) 展开失效。Done, pre-registered verdict FAIL (R121, 2026-10-21 14:16 EDT, replayed): the width deficit is re-attributed to the conditional variance \(\mathrm{Var}[\Delta C \mid \hat{\theta}] < V_{3p}(\hat{\theta})\) — within-bin \(S_{true}\) mixing is negligible and only adds variance (0.000–0.014), and the conditional-mean term is likewise negligible (0.000–0.010). A new \(O(n^{-1})\) closed-form factor \(c = 1 - \mathrm{tr}(I^{-1}Q)/(2V_{3p}) - \mathrm{tr}(I^{-1}P)^2/(4V_{3p})\) matches a brute-force Edgeworth expansion to four decimals; it is exact at \(b = 0.01\) (rms log error 3.1 %, the corrected width at DET_ML 6–20 within 1.6\(\sigma\) of 1) but under-corrects at \(b = 0.04\) (0.952 \(\pm\) 0.013, −3.6\(\sigma\)) and more so at \(b = 0.16\) (0.926 \(\pm\) 0.012, −6.1\(\sigma\)) — 3 of 8 bins beyond 2\(\sigma\), so the key FAILS. The residual tracks the \(V_{3p}\) cancellation factor \(4\sum\mu L^2/V_{3p}\) (\(R^2\) 0.89), not \(n_{\rm eff}\); the hypothesis is that at high \(b\), \(L \approx (S/b)p\) is nearly parallel to the S-score, so the \(O(n^{-1})\) expansion breaks down there.