历史：前世今生History: The Full Story

时间线总览Timeline Overview

来源：cognition-os/04_DECISION_LOG.md（完整决策日志，每条都写明背景与理由）。Source: cognition-os/04_DECISION_LOG.md (the full decision log; every entry records its background and rationale).

阶段 1：玩具世界（2026 年 3–4 月）Stage 1: Toy World (2026 months 3–4)

决定（2026-03-28）：先建"玩具世界"——2D 高斯 PSF + 常数背景 + 单能段。理由：解析上可处理、模拟快、可以在碰真实管线复杂度之前建立直觉。Decision (2026-03-28): build a "toy world" first — 2-D Gaussian PSF + constant background + single energy band. Rationale: analytically tractable, fast to simulate, and builds intuition before touching real-pipeline complexity.

4 月的 44.6M 样本蒙特卡洛给出了项目的第一个中心结论——因子化假设（factorization hypothesis）：Month-4's 44.6M-sample Monte Carlo delivered the project's first central result — the factorization hypothesis:

p(DET_ML, Ŝ | S, b) ≈ p(DET_ML | Ŝ, b) · p(Ŝ | S, b)

即：给定测得流量 Ŝ 和背景 b，DET_ML 基本与真实流量 S 无关。这让贝叶斯 deboosting 变得可行。它至今成立（在玩具世界里），但后来发现"基本无关"不等于"完全无关"——这正是 v2 重构的伏笔。That is: given the measured flux Ŝ and background b, DET_ML is essentially independent of the true flux S. This makes Bayesian deboosting feasible. It still holds (in the toy world), but "essentially independent" later turned out not to be "completely independent" — the seed of the v2 restructuring.

同一时期也立下了观测条件化的方法论（T40）：理论必须条件于可观测量 (Ŝ, b)，S_true 只是模拟标签，永远不能当输入。The same period fixed the observable conditioning methodology (T40): theory must condition on the observables (Ŝ, b); S_true is only a simulation label and can never be an input.

阶段 2：真实 PSF 大战（2026 年 5 月）Stage 2: The Real-PSF Battle (2026 month 5)

5 月是"踩坑月"。为了和真实 emldetect 对比，项目把玩具 PSF 换成 XMM 官方的 ELLBETA 模型，结果连续发现：Month 5 was "pitfall month". To compare against real emldetect, the project swapped the toy PSF for XMM's official ELLBETA model, and discovered in succession:

• scut/ecut 之谜（12% 宽度差）。The scut/ecut mystery (12% width gap).emldetect 默认的拟合半径参数 ecut = 0.68 只覆盖 961 像素中的 37 个，导致 ΔC 宽度被压缩 12–18%。对齐官方约定（scut = 0.9, ecut = 15，即 60 角秒）后消失。教训：先查约定，再怀疑物理（决策日志 2026-05-02）。emldetect's default fit-radius parameter ecut = 0.68 covers only 37 of 961 pixels, compressing the ΔC width by 12–18%. It vanished after aligning to the official convention (scut = 0.9, ecut = 15, i.e. 60 arcsec). Lesson: check conventions before doubting physics (decision log 2026-05-02).
• 自研 PSF 的 5 个 bug。5 bugs in the in-house PSF.AnalyticEllbetaPSF 的辐条角度偏移、振幅偏大 2.7–2.9 倍等。所有依赖它的结果作废重跑（2026-05-11/12）。AnalyticEllbetaPSF's spoke-angle offset, amplitude 2.7–2.9× too large, and more. Every result depending on it was invalidated and re-run (2026-05-11/12).
• 优化器陷阱。The optimizer trap.L-BFGS-B 在宽 PSF 的平坦 Cash 地形上会卡住，产生负 ΔC；必须用 Nelder-Mead（2026-05-12）。L-BFGS-B stalls on the flat Cash landscape of a wide PSF and produces negative ΔC; Nelder-Mead is required (2026-05-12).
• DET_ML 公式的 shape 参数。The shape parameter of the DET_ML formula.单图像 3 参数拟合是 shape = ν/2 = 1.5，不是 2.5；gammaincc 的第二参数要除以 2（I040/I041）。The single-image 3-parameter fit has shape = ν/2 = 1.5, not 2.5; gammaincc's second argument must be divided by 2 (I040/I041).

这一阶段定下一个重要策略（2026-05-13）：自洽性优先于外部一致性——理论先要在自己的管线里严格成立，再去追 emldetect 的工程细节。This stage fixed an important strategy (2026-05-13): self-consistency before external agreement — the theory must hold rigorously in its own pipeline before chasing emldetect's engineering details.

阶段 3：emldetect 生产（2026 年 6–7 月）Stage 3: emldetect Production (2026 months 6–7)

6 月起项目开始大规模生产：3 相机（PN、MOS1、MOS2）× 5 波段 × 6 种背景 × psfgen 逐位置 PSF，用 SAS 的 emldetect 拟合。规模到了月级，于是立下整套运行纪律：CPU ≤ 60%、PID 文件、双主机校验（Lang2/Lang3 位元级一致）、发射门控（launch-only supervisor）。From month 6 the project went into large-scale production: 3 cameras (PN, MOS1, MOS2) × 5 bands × 6 backgrounds × psfgen per-position PSFs, fitted with SAS's emldetect. With month-scale runs came the full operational discipline: CPU ≤ 60%, PID files, dual-host verification (bit-exact Lang2/Lang3 agreement), and launch-only supervision.

7 月的低流量研究（E047–E050）发现暗源残差由质心漂移（centroid relocation）主导：拟合器在噪声里挪动位置，把 ΔC 抬高。位置优化项 B_relocate = +2.10，比振幅项大一倍多。Month-7's low-flux study (E047–E050) found that the faint-source residual is dominated by centroid relocation: the fitter moves the position within the noise, raising ΔC. The position-optimization term B_relocate = +2.10, more than twice the amplitude term.

阶段 4：低流量核与解析理论（7–9 月）Stage 4: Low-Flux Kernel & Analytic Theory (months 7–9)

这是理论收获最大的三个月。一系列实验（E086 → E096）逐步把"条件均值修正 δ"从经验公式推成解析公式：These three months yielded the most theory. A series of experiments (E086 → E096) turned the "conditional-mean correction δ" from an empirical formula into an analytic one:

1. E086：二阶展开 δ = −½Σrᵢ(I⁻¹)ₐᵦ[μᵢₐᵦ − μᵢₐμᵢᵦ/μᵢ]，方向和量级对，但精度不够；E086: second-order expansion δ = −½Σrᵢ(I⁻¹)ₐᵦ[μᵢₐᵦ − μᵢₐμᵢᵦ/μᵢ], right direction and magnitude but insufficient precision;
2. E088：理论匹配的连续拟合器上，δ₀ = 3 − ½tr(H_μ I⁻¹) − ∇μ₃p·bias_CS 在 μ₃p ≥ 25 完全闭合；E088: on the theory-matched continuous fitter, δ₀ = 3 − ½tr(H_μ I⁻¹) − ∇μ₃p·bias_CS closes completely at μ₃p ≥ 25;
3. E090：加上 Lawley/Bartlett 项 ε，δ₁ 在 μ₃p ≥ 13.5 就闭合，零拟合系数；E090: adding the Lawley/Bartlett term ε, δ₁ closes already at μ₃p ≥ 13.5, with zero fitted coefficients;
4. E092：FFT 相位平移模板族证明 emldetect 的 PSF 对齐方式属于公式成立的类别；E092: the FFT phase-shift template family proves emldetect's PSF alignment is in the class where the formula holds;
5. E096：在全部 17 个 PN 波段 4 位置上，δ₁ 都闭合（每箱精度 3%）。E096: δ₁ closes on all 17 PN band-4 positions (3% per-bin precision).

同时 E140 派生出零参数宽度律：G² = 1 − c20/μ₃p，c20 = 1.794（推导值，非拟合）。Meanwhile E140 derived the zero-parameter width law: G² = 1 − c20/μ₃p with c20 = 1.794 (derived, not fitted).

来源：cognition-os/experiments/E140_local_width_zero_parameter.md（c20 = 1.794461）；results/v2/T11/benchmark_deltac_T11S3_iso60_M2b4_c20_2D.csv。Source: cognition-os/experiments/E140_local_width_zero_parameter.md (c20 = 1.794461); results/v2/T11/benchmark_deltac_T11S3_iso60_M2b4_c20_2D.csv.

阶段 5：catalog 桥与 5 参数（8–9 月）Stage 5: Catalogue Bridge & 5 Parameters (months 8–9)

month-8 的 catalog 工作（E075–E085）证明：单靠 CTS 和背景两个量，无法用一个全局公式描述 DR14 目录；但按（相机、波段、离轴角）分层后，43 层查找表能把均值 RMSE 从 15.0 降到 3.3。E116 进一步把 15 个配置压缩到 5 个参数：η(S) = η₀(S/10)^q，其中 η₀ = 0.924（MOS）/ 0.955（PN），波段间几乎不变。Month-8's catalogue work (E075–E085) showed that CTS and background alone cannot describe the DR14 catalogue with one global formula; but after stratifying by (camera, band, off-axis angle), a 43-layer lookup table reduces the mean RMSE from 15.0 to 3.3. E116 further compressed 15 configurations into 5 parameters: η(S) = η₀(S/10)^q with η₀ = 0.924 (MOS) / 0.955 (PN), nearly unchanged across bands.

这个 η₀ 的物理起源成了悬案：它是渐晕（vignetting）约定吗？E158（T7）检验并否定了渐晕假设：RATE × EXP 本来就等于图像里拟合的计数。真正的答案在 10 月才出现（见阶段 7）。The physical origin of η₀ became an open question: was it a vignetting convention? E158 (T7) tested and refuted the vignetting hypothesis: RATE × EXP already equals the counts fitted in the image. The real answer only arrived in month 10 (see Stage 7).

阶段 6：Pareto 基准 v1（2026-09-24）Stage 6: Pareto Benchmark v1 (2026-09-24)

用户决定：把"ΔC 条件分布"问题变成一个多目标基准（benchmark）。任何理论的每一轮改进都要在同一把尺子下比较：均值偏差、宽度偏差、经验参数个数。规则写入 cognition-os/PARETO_BENCHMARK.md，成绩公开在网站上。The user decided to turn the "ΔC conditional distribution" problem into a multi-objective benchmark. Every round of theory improvement must be measured with the same ruler: mean bias, width bias, number of empirical parameters. The rules went into cognition-os/PARETO_BENCHMARK.md; scores are public on the website.

v1 运行了两周就暴露出自己的问题（E151，2026-10-09）：v1 ran for two weeks before exposing its own problems (E151, 2026-10-09):

• 评分隐含地依赖模拟设计里的源计数先验（α ≈ 1），换先验均值会移动；the scores implicitly depended on the simulation's source-count prior (α ≈ 1); changing the prior moved the means;
• 低于噪声底的分数差异也被拿来排名；score differences below the noise floor were still used for ranking;
• 重采样单元错了（应该是图像/观测，不是分箱）；the resampling unit was wrong (should be image/observation, not bin);
• 没有评"过阈值概率"——目录用户最关心的量。"threshold-exceedance probability" — the quantity catalogue users care most about — was not scored.

阶段 7：基准 v2 与二维位置损失（2026-10）Stage 7: Benchmark v2 & the two-dimensional Position Loss (2026 month 10)

2026-10-09 基准整体重设为 v2（规则见基准设计页）。month 10 上旬的突破是二维位置损失：On 2026-10-09 the benchmark was fully redesigned as v2 (rules on the Benchmark Design page). The early-month-10 breakthrough was the two-dimensional position loss:

emldetect 对 15 幅图像（3 相机 × 5 波段）同时拟合 1 个源位置。每幅图像的 ΔC 因此要在"共同的位置"上评估，而不是它自己最优的位置——这造成一个损失。旧模型只减了 (1−w)（w 是该图像在位置信息中的占比），但位置有 2 个维度（x 和 y），精确的二阶形式是：emldetect fits 1 source position simultaneously across 15 images (3 cameras × 5 bands). Each image's ΔC is therefore evaluated at the shared position rather than its own optimum — a loss. The old model subtracted only (1−w) (w is that image's share of the position information), but position has 2 dimensions (x and y); the exact second-order form is:

L_2D = tr(Φᵢ(Φᵢ⁻¹ − Φ_tot⁻¹)) ≈ 2(1 − wᵢ)

模拟里实测损失与这个公式的比值当年读作 1.04–1.06、与 S 无关（T11_013；后由 07 R013 放宽为跨 Ŝ 分箱 1.01–1.25，"与 S 无关"的措辞已撤回）。补上这一项后，catalog 的 MOS 均值亏缺一度看起来被零参数地闭合（在 6–10 窗口）：M2 波段 4 逐箱 A1 从 0.324 降到 0.059；两个零参数条目于 2026-10-13 正式登记。该闭合读法已于 2026-10-16 晚间撤回：按模拟自己的口径目录亏缺是 −9.282 % [−9.350, −9.219] 而非 −10.4 %（后者不可复现），联合模拟试点只到 −5.9 % / −7.4 % 且与目录不匹配，Q1 现为 open（闭合未证，07 R019）。In simulation, the ratio of the measured loss to this formula was read at the time as 1.04–1.06, independent of S (T11_013; later widened by 07 R013 to 1.01–1.25 across Ŝ bins, and the "independent of S" wording is withdrawn). With this term added, the catalogue's MOS mean deficit once appeared closed zero-parameter (in the 6–10 window): M2 band 4 per-bin A1 fell from 0.324 to 0.059, and two zero-parameter entries were formally registered on 2026-10-13. That closure reading was withdrawn on the evening of 2026-10-16: under the simulation's own convention the catalogue deficit is −9.282 % [−9.350, −9.219], not −10.4 % (the latter is not reproducible), the joint-simulation pilot reaches only −5.9 % / −7.4 % and is unmatched to the catalogue, and Q1 is now open (closure not shown, 07 R019).

但当天晚些时候的 T11_020 窗口平坦性检验发现：二维损失在 3 个 ML 窗口不平坦，不是闭合项；PN 上 (1−w) 反而闭合，L_2d 在 PN 上扣多了约 +0.06。联合拟合损失的正确形式仍是开放问题。But later the same day, the T11_020 window-flatness test found: the two-dimensional loss is not flat across the 3 ML windows, so it is not the closing term; on PN (1−w) closes instead, and L_2d over-subtracts on PN by ~+0.06. The correct form of the joint-fit loss remains open.

来源：benchmark/models/v2_T3_joint_2D.py 的推导注释；T11_012/T11_013 检查点与评审。Source: the derivation comments in benchmark/models/v2_T3_joint_2D.py; checkpoints and reviews T11_012/T11_013.

PN 相机是当前的例外：二维损失在 PN 上过冲 +5.5 到 +6.7%。排查已经定位到 PN 的条件均值修正项 δ（模板等级、波段内能量、离轴形状、光谱加权都已排除，见 E162 阶段 15–16），判定用的冻结蒙特卡洛网格已于 2026-10-13 跑完。The PN camera is the current exception: the two-dimensional loss overshoots on PN by +5.5 to +6.7%. The investigation localized it to PN's conditional-mean correction δ (template grade, in-band energy, off-axis shape, spectral weighting all excluded; see E162 stages 15–16), and the frozen Monte Carlo grid for adjudication finished on 2026-10-13.

阶段 19（T11_019）判定：固定 S_true 的 MC 超出是局部拟合器的低信噪比截断，不是 PN 响应；但口径（S_true 条件）与目录评分（Ŝ 条件）不匹配，不能直接搬到目录。阶段 20（T11_020）随后用窗口平坦性检验发现：二维损失在 3 个 ML 窗口不平坦，不是闭合项；PN 上 (1−w) 反而闭合，L_2d 在 PN 上扣多了约 +0.06。Stage 19 (T11_019) ruled: the fixed-S_true MC excess is a low-SNR truncation artefact of the local fitter, not a PN response; but the conditioning (S_true) does not match catalogue scoring (Ŝ), so it cannot be transferred to the catalogue. Stage 20 (T11_020) then found via the window-flatness test: the two-dimensional loss is not flat across the 3 ML windows, so it is not the closing term; on PN (1−w) closes instead, and L_2d over-subtracts on PN by ~+0.06.

2026-10-17：PI 把主攻方向转到模拟宽度（新问题 Q7）。07:10Z 的决定：目录问题（Q1 均值亏损、Q2 高计数 m、Q3 目录宽度、Q5 M31 图）降为附加讨论——不新派任务，在跑的跑完轻量分诊，留着不解决也可以接受；力气用在我们自己的模拟的宽度上：零参数插入式 \(\sqrt{V}\) 能不能描述给定 \(\hat{S}\) 的 \(\Delta C\) 宽度，在哪里失效。同一天提出理论任务的派发规则：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，复核判据脚本后才接受公式。2026-10-17: the PI moved the main line to the simulation width (new question Q7). The 07:10Z decision: the catalogue questions (Q1 mean deficit, Q2 high-count m, Q3 catalogue width, Q5 M31 map) are demoted to a bonus discussion — dispatch no new work, triage the running ones lightly, and leaving them unresolved is acceptable; the effort goes to the width of our own simulations: does the zero-parameter plug-in \(\sqrt{V}\) describe the \(\Delta C\) width given \(\hat{S}\), and where does it fail. The same day raised the theory-dispatch rule: write the known results and the executable key first, two independent arms in parallel, red teams attack the derivations with numerical counterexamples, and replay the key script before accepting any formula.

当天交付并评审的四条。① W1 玩具理论（T229/E170，Provisional，07 R037）：位置固定的 Fisher 宽度 \(V_{3p}\) 在 DET_ML ≳ 20 与低 b 正确（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）；条件均值偏低 +0.8…+1.7 \(\Delta C\)，无公式通过；截断支撑的矩全部通过。两臂结论相反，仲裁运行（项目自己的拟合器）支持 codexnim 臂、拒绝 Hermes 臂。② W2 宽度梯子（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、mean z −0.05…−0.07（用 emldetect 实际得到的背景 BG_MAP_EML 打分；b_true 打分下的 1.02–1.12 与 +0.28…+0.47 已由 RT-45 分诊撤回）；L0 的 1.45 / 1.52 是两种总体的混合——拟合位置远离中心的行 \(\Delta C\) 在插入式下方（mean z −1.70 / −4.03），“look-elsewhere 抬高 \(\Delta C\)”的读法符号相反、已撤回。③ SAS 22.1 对 20 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 生产不需要 SAS 版本注记。④ Q1 目标/L2d（07 R035，bonus）：目录目标冻结 −9.282 % [−9.350, −9.219]；试点的“L2d/\(\mu\) 4.9 %”是代码伪迹（更正为 10.4 %）。Four things reported and were reviewed that day. ① W1 toy theory (T229/E170, Provisional, 07 R037): the position-fixed Fisher width \(V_{3p}\) is right at DET_ML ≳ 20 and at low b (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 is low by +0.8…+1.7 \(\Delta C\) with no passing formula; the truncated-support moments all pass. The two arms disagreed and an arbiter run (the project's own fitter) sided with the codexnim arm and rejected the Hermes arm. ② The W2 width ladder (T228, Provisional, 07 R036, R039; the L1 numbers replaced by 07 R041): L1 emldetect std z 0.95 / 0.975 / 1.00 / 1.02 with mean z −0.05…−0.07, scored with the background emldetect was actually given (BG_MAP_EML; the b_true-scored 1.02–1.12 and +0.28…+0.47 are withdrawn by RT-45's triage); L0's 1.45 / 1.52 is a mixture of two populations — the rows whose fitted position wandered far from the centre have \(\Delta C\) below the plug-in (mean z −1.70 / −4.03), so the "look-elsewhere raises \(\Delta C\)" reading has the opposite sign and is withdrawn. ③ The SAS 22.1 vs 20 replay (T227/E169, Provisional, 07 R034, R038): emldetect and eboxdetect are bit-identical, only esplinemap changes; the truth-background L1/J production needs no SAS-version caveat. ④ Q1's target/L2d (07 R035, bonus): the catalogue target is frozen at −9.282 % [−9.350, −9.219]; the pilot's "L2d/\(\mu\) 4.9 %" was a code artefact (corrected to 10.4 %).

T13 线当天关闭（07 R032，bonus）：第 7b 阶段用真实邻居模板重做（282/282，约扣除 91 % 的邻居计数），大偏移标记比例下降（M2 0.925→0.725、M1 0.972→0.778）但振幅亏损没有恢复（big\(\div\)control \(S_A\) M2 +0.003、M1 −0.054）；成因未定，不再派 stage 7c。仍在跑：W3（全局搜索拟合器的条件律，双臂）与红队 RT-46、RT-47（RT-43/44/45 已分诊：RT-43 把 W1 宽度失效的一半降为待质疑（R042 定为 Questioned），R044 在调度方复跑后废弃该记录的失效，R045 记录 (15.32, 0.08) 处 0.975 的窄侧偏离；RT-44 保留 ecut 0.68 的截断机制但撤回"0.802 对 0.833"的读法（R040）；RT-45 把 L1 改用 BG_MAP_EML 打分（R041））。The T13 track closed that day (07 R032, bonus): stage 7b redid the neighbour removal with real templates (282/282, about 91 % of the neighbour's counts removed); the big-shift flag fraction falls (M2 0.925→0.725, M1 0.972→0.778) but the amplitude loss is not recovered (big\(\div\)control \(S_A\) M2 +0.003, M1 −0.054); the cause is not identified and no stage 7c is dispatched. Still running: W3 (the conditional law for the global-search fitter, two arms) and red teams RT-46, RT-47 (RT-43, RT-44 and RT-45 are triaged: RT-43 downgraded the W1 width-failure half to Questioned as a fitter clip artefact (R042); R044 deprecated the recorded failure after the dispatcher's re-run; R045 recorded the 0.975 narrow-side departure at (15.32, 0.08); RT-44 kept the ecut 0.68 truncation mechanism but withdrew the "0.802 against 0.833" reading, R040; RT-45 rescored L1 with BG_MAP_EML, R041).

2026-10-18：宽度主线的一天，W3 与 W4 收尾、W5 发出。七条修订记录（R047–R052），其中最重要的一更是调度方撤回自己一小时前的结论。2026-10-18: a day on the width main line; W3 and W4 closed, W5 launched. Seven revision records (R047–R052), the most important of which is the dispatcher withdrawing its own hour-old conclusion.

玩具均值被 E088 \(\delta_0\) 关闭（R047，红队检查的分诊）。"没有任何均值公式通过"撤回：项目自带的无参数修正 \(\delta_0\)（模板族的精确位置导数）在 6 个判据格中 4 个把 mean z 落到 −0.002…+0.018（N = 10,000），只有两个 DET_ML \(\approx\) 8 的格子不通过（+0.058、+0.070）；随库默认的 np.gradient 导数场少修正 0.09–0.15 \(\Delta C\)。The toy mean is closed by E088 \(\delta_0\) (R047, a red-team triage). "No mean formula passes" is withdrawn: the project's own parameter-free \(\delta_0\) (exact position derivatives of the template family) brings mean z to −0.002…+0.018 at 4 of the 6 key cells (N = 10,000), failing only at the two DET_ML \(\approx\) 8 cells (+0.058, +0.070); the library default np.gradient fields under-correct by 0.09–0.15 \(\Delta C\).

截断支撑矩重新限定范围（R048）。"掩膜矩在 27 格全过"只对位置固定成立；位置自由时按居中模板打分失败（mean −2.2…−3.4），改在拟合位置渲染模板后 mean z +0.10…+0.16、std z 1.03–1.05（N = 1000，开放）；背景盒均值只在 \(N_{bg} \geq 1000\) 修正。The truncated-support moments are re-scoped (R048). "The masked moments pass all 27 cells" holds with the position fixed only; with the position free, centred-template scoring fails (mean −2.2…−3.4) while the template rendered at the fitted position gives mean z +0.10…+0.16, std z 1.03–1.05 (N = 1000, open); the background-box mean is fixed only for \(N_{bg} \geq 1000\).

L1 的"宽度随 b 变化"降为 Questioned（R049）。用 E074 有效模板加 PSF 加权或像素背景后，6–10 的宽度 0.929、b 梯度 +0.016（记录值 +0.086）；约 5–7 % 的偏窄成因未明，且不是 beyond-Fisher 位置项。The L1 "b-dependent width" is Questioned (R049). With the E074 effective template and a PSF-weighted or pixel-map background the 6–10 width is 0.929 and the b-gradient +0.016 (recorded +0.086); the \(\approx\) 5–7 % narrowing is unexplained and is not the beyond-Fisher position term.

L0 近行的"通过"是合并陈述（R050）。r̂ < 2 px 内部宽度 0.865 → 0.751、均值 +0.33 → −0.27（6–10）；E088 精确 \(\delta_0\) 把近行均值残差降到 +0.088 / +0.010；\(\sigma_{ratio}\) 的转移本身复算到 2e-16。The L0 near-row "pass" is a pooled statement (R050). Inside r̂ < 2 px the width runs 0.865 → 0.751 and the mean +0.33 → −0.27 (6–10); the E088 exact \(\delta_0\) lowers the near-row mean residual to +0.088 / +0.010; the \(\sigma_{ratio}\) transfer itself reproduces to 2e-16.

R045 的"窄侧偏离"撤回，一小时后被取代（R051 → R052）。W4 codexnim 臂 N = 20,000 复跑 (15.32, 0.08) 得 0.997，重尾 SE 0.011（不是 0.007），低格含拟合器失败行——R045 改判"未确立"。R052 用两臂同一条打分代码给出真正的结构：可观测核 r̂ \(\leq\) 4 px 上 std z(\(\delta_0\)) = 0.984 / 0.974、0.965 / 0.964、0.961 / 0.961（两臂，SE 0.005），窄 2–4 %；全行 std 由 0.1–2.5 % 的远行（涨落锁定、拟合器失败）决定。V2、V2b 方向对但核上过校正到 1.09–1.65；V2b 就是 E171 的 \(\sigma_{ratio}\)，放到玩具核上给 1.024（(0.08, 25.23)，那里 \(V_{3p}\) 已是对的 0.998）与 1.10 / 1.18 / 1.30——T230 的 L0 近行通过因此在玩具上得不到支持，待 W5 判定。R045's "narrow-side departure" is withdrawn and superseded an hour later (R051 → R052). The W4 codexnim arm's N = 20,000 re-run at (15.32, 0.08) gives 0.997 with a heavy-tail SE of 0.011 (not 0.007), and the low cells contain fitter-failure rows — so R045 is re-graded "not established". R052, scoring both arms' rows with one code path, gives the real structure: on the observable core r̂ \(\leq\) 4 px std z(\(\delta_0\)) = 0.984 / 0.974, 0.965 / 0.964, 0.961 / 0.961 (two arms, SE 0.005), narrow by 2–4 %; the all-row std is set by 0.1–2.5 % of far rows (fluctuation lock-ons, fitter failures). V2 and V2b move the right way but over-correct the core to 1.09–1.65; V2b is E171's \(\sigma_{ratio}\), which applied to the toy core gives 1.024 at (0.08, 25.23) (where \(V_{3p}\) is already right at 0.998) and 1.10 / 1.18 / 1.30 — so T230's L0 near-row pass is not supported by the toy, pending W5.

W3 与 W4 关闭，W5 发出。两臂都不达判据：无参数公式都不行，\(\delta_1\) 两臂都排除，codexnim 的 \(\delta_0(\hat{S} + 4 \cdot bias_S)\) 六格全过但 k = 4 是在判据格上选的（经验锚，不是公式）。最佳均值律 \(\delta_0\) + \(V_{3p}\)，域 DET_ML \(\gtrsim\) 14。W3 Hermes 臂在全部 2.546 M 收敛行上确认近行偏窄是拟合器性质（未修正 0.87–0.90 于 6–10），但未给宽度公式；E171 的 \(\sigma_{ratio}\) 仍是最佳臂。W5（给定 \((\hat{S}, \hat{r})\) 的条件宽度，用精确导数重推 \(\sigma_{ratio}\)）06:58:41Z 发出，两臂；红队 RT-51、RT-52、RT-53 在跑。T12 第 20 阶段修复清单经审阅后部署为 c83b1ad3。W3 and W4 closed, W5 launched. Neither arm meets the key: no parameter-free formula works, \(\delta_1\) is ruled out by both arms, and codexnim's \(\delta_0(\hat{S} + 4 \cdot bias_S)\) passes all six cells but k = 4 was chosen on them (an in-sample empirical anchor, not a formula). The best mean law is \(\delta_0\) + \(V_{3p}\), domain DET_ML \(\gtrsim\) 14. The W3 Hermes arm confirms on all 2.546 M converged rows that the near-row narrowing is a property of the fitter (uncorrected 0.87–0.90 at 6–10) but gives no width formula; E171's \(\sigma_{ratio}\) remains the best arm. W5 (the conditional width given \((\hat{S}, \hat{r})\), re-deriving \(\sigma_{ratio}\) with exact derivatives) was launched at 06:58:41Z, two arms; red teams RT-51, RT-52 and RT-53 are running. The T12 stage-20 fix list was reviewed and deployed as c83b1ad3.

当天晚些时候：T226 的"只有截尾主体被校准"被改写（R053）。红队检查（攻击 T226 的 6–20 低计数区）发现"原始矩不存在"其实是玩具牛顿拟合器在近空图像上发散（\(\Delta C\) < 0、\(\hat{S} \to 10^{13}\)，0.03–12.7 % 的重复）；改用有界拟合后每个灾难行都消失，全量原始矩有限且校准：每个 b 匹配格 mean z 在 \(\pm\) 0.016 内（典型 DET_ML 6–20，两个种子）；b \(\geq\) 0.02 时 std 0.987–1.011，b = 0.005 在 DET_ML 6–10 偏窄 1–3 %；右尾短于正态（q99.9 % 2.3–3.0 对 3.09）——这些是固定 \(S_{true}\) 格子的格均值，R061 已把 T226 改写成可观测律（按 \(\hat{S}\) 分箱：b \(\geq\) 0.02 时 std z 在 1 的 2–3 % 内、mean z 在 \(\pm\) 0.09 内；b = 0.005 且 \(\hat{S} \lesssim\) 3 是离散格子，\(\hat{S} = 0\) 原子占 17–23 %）。来源：07 R053、R061；reviews/RT-51_triage_claude.md、reviews/RT-57_triage_claude.md。Later that day: T226's "only the trimmed bulk is calibrated" is rewritten (R053). A red-team check (attacking T226's low-count 6–20 region) found the "raw moments do not exist" was the toy Newton fitter diverging on near-empty images (\(\Delta C\) < 0, \(\hat{S} \to 10^{13}\), 0.03–12.7 % of reps); with a bounded fit every catastrophic row disappears and the full raw moments are finite and calibrated: mean z within \(\pm\) 0.016 at every b-matched cell (typical DET_ML 6–20, two seeds); std 0.987–1.011 at b \(\geq\) 0.02, and b = 0.005 at DET_ML 6–10 narrow by 1–3 %; the right tail is short of normal (q99.9 % 2.3–3.0 vs 3.09) — these are fixed-\(S_{true}\) cell averages, and R061 rewrites T226 as an observable law (binned in \(\hat{S}\): at b \(\geq\) 0.02 the std z is within 2–3 % of 1 and the mean within \(\pm\) 0.09; at b = 0.005 and \(\hat{S} \lesssim\) 3 it is a discrete lattice, and the \(\hat{S} = 0\) atom is 17–23 %). Sources: 07 R053, R061; reviews/RT-51_triage_claude.md, reviews/RT-57_triage_claude.md.

当天下午：上午的宽度结论几乎全部被改写（R054–R059）。W5 双臂收尾（E173）：把 \(\sigma_{ratio}\) 的位置导数换成拟合模板族的精确导数后，它在 L0 近行 DET_ML 6–10 过度修正（std z 1.67 / 1.83，两臂），玩具核也过度修正（1.05–2.18）——判据未达成。RT-54 用不同差分格式给出决定性证据（R057）：只有 np.gradient 的 3 点 h = 1 模板通过，5 点给 1.31、h = 2 给 0.81、精确导数给 1.68，所以原先"18/19 格通过"是离散化误差的偶然抵消；\(\sigma_{ratio}\) 因此降级为经验校准，T230 的 L0 近行通过标为 Questioned。RT-52 的分诊把 T219 重新限定（R056）：位置固定时逐行 \(V(\hat{A})\) 在 DET_ML 8–600 通过（std 0.995–1.007，N = 20,000），而记录的脚本除以 \(V(S_{true})\)（一个标签，调度方自己的 IRON-RULE 失误）才失败；resume 给出新记录 T234（R058）——位置自由时残差均值 +1.15…+1.73 \(\Delta C\) 等于解析 \(\delta_{3p}\)，宽度 std 0.914（DET_ML 8）、0.955（15）、0.992–1.000（30–600）。R060 限定：这些是固定 \(S_{true}\) 格子，用 \(V(S_{true})\) 归一后 std 是 1.03 / 1.01 / 1.02 / 1.00 / 1.00（DET_ML 8 / 15 / 30 / 60 / 600），偏窄只在逐行 \(V(\hat{S})\) 且位于格内低 \(\hat{S}\)；"玩具 / L0 / 真实 PSF 三处共同偏窄"的综合已撤回（行集不同）。RT-55 把玩具低计数偏窄改写成可观测律（R059）：按可观测 \(\hat{S}\) 分箱后核 r̂ \(\leq\) 4 px 的 std z 是 0.82–0.86（DET_ML 1–3）、0.93–0.95（\(\approx\) 5）、0.94–0.96（\(\approx\) 7.5）、0.993（\(\approx\) 12 及以上，1 SE 内）；固定 \(S_{true}\) 格子的"窄 2–4 % 于 8–14"只是格子效应。新缺陷 B009：lib/beyond_fisher 的方差项漏掉泊松 \(\kappa_4\)，其 mean_curvature 是错的。RT-56（低计数偏窄的尺寸预测）与 RT-57（可观测箱里的振幅律）15:44Z 发出。T12 第 23 阶段经审阅后部署为 935d5f5f。来源：07 R054–R059、reviews/W5_codexnim_claude.md、reviews/W5_hermes_claude.md、reviews/RT-52_triage_claude.md、reviews/RT-55_triage_claude.md、T234、B009。That afternoon: almost every morning width conclusion was rewritten (R054–R059). W5's two arms closed (E173): with \(\sigma_{ratio}\)'s position derivatives replaced by the fitter template family's exact derivatives it over-corrects L0 near rows at DET_ML 6–10 (std z 1.67 / 1.83, two arms) and over-corrects the toy core too (1.05–2.18) — the key is not met. RT-54 gave the decisive evidence with different stencils (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, so the old "18 of 19 cells pass" was an accidental cancellation of a discretisation error; \(\sigma_{ratio}\) is therefore downgraded to an empirical calibration and T230's L0 near-row pass is marked Questioned. The RT-52 triage rescoped T219 (R056): with the position fixed the per-row \(V(\hat{A})\) passes at DET_ML 8–600 (std 0.995–1.007, N = 20,000), while the recorded script's division by \(V(S_{true})\) (a label; the dispatcher's own IRON-RULE slip) is what failed; the resume gives the new record T234 (R058) — with the position free the residual mean +1.15…+1.73 \(\Delta C\) equals the analytic \(\delta_{3p}\) and the width std is 0.914 (DET_ML 8), 0.955 (15) and 0.992–1.000 (30–600). R060 qualifies it: these are fixed-\(S_{true}\) cells, normalised by \(V(S_{true})\) the std is 1.03 / 1.01 / 1.02 / 1.00 / 1.00 (DET_ML 8 / 15 / 30 / 60 / 600), the narrowing appears only with the per-row \(V(\hat{S})\) and sits at low within-cell \(\hat{S}\); the "narrow in the toy, L0 and the real PSF alike" synthesis is withdrawn (the row sets differ). RT-55 rewrote the toy low-count narrowing as an observable law (R059): binned by the observable \(\hat{S}\), the core r̂ \(\leq\) 4 px std z is 0.82–0.86 (DET_ML 1–3), 0.93–0.95 (\(\approx\) 5), 0.94–0.96 (\(\approx\) 7.5) and 0.993 (\(\approx\) 12 and above, within 1 SE); the fixed-\(S_{true}\) cells' "narrow by 2–4 % at 8–14" is only a cell effect. New defect B009: lib/beyond_fisher's variance term omits the Poisson \(\kappa_4\) and its mean_curvature is wrong. RT-56 (a size prediction for the low-count narrowing) and RT-57 (the amplitude law in observable bins) were launched at 15:44Z. The T12 stage-23 refresh was reviewed and deployed as 935d5f5f. Sources: 07 R054–R059, reviews/W5_codexnim_claude.md, reviews/W5_hermes_claude.md, reviews/RT-52_triage_claude.md, reviews/RT-55_triage_claude.md, T234, B009.

当天深夜（17:31Z 之后）：真实 PSF 的宽度律被重写（R062、R063）。红队 RT-58（codexnim 跑，16:36Z 启动，约 4.9 h，virgo 约 170 ms/行）把 T234 的固定 S_true 分箱表换成可观测 \(\hat{S}\) 分箱：std z 在 DET_ML \(\approx\) 7 / 9 / 12 / 18 为 0.88 / 0.92 / 0.94 / 0.96，约 25–40 才回到 1，低端深到 0.73（DET_ML 0.75），偏窄持续到目录区间，不像玩具（R059：DET_ML \(\approx\) 12 就消失）。Claude 分诊确认了 R062，并用先验边界/斜率检验（S_true \(\geq\) 7 / \(\geq\) 8.5、斜率 −2.5 \(\to\) −1.5）证明这条律与先验无关。R063 则给出先验作用域：宽度与源计数先验无关，均值只在 −2.5 先验下成立（DET_ML \(\lesssim\) 17 以下随先验移动）。RT-60 与 RT-61 随后发出。T12 第 27–28 阶段刷新后部署为 663838ca（数据截至 17:31 UTC）。来源：07 R062、R063及两篇分诊评审。

Late that night (after 17:31Z): the real-PSF width law was rewritten (R062, R063). The red team RT-58 (run by codexnim, started 16:36Z, about 4.9 h, about 170 ms/row on virgo) replaced T234's fixed-S_true bin table with observable \(\hat{S}\) bins: std z is 0.88 / 0.92 / 0.94 / 0.96 at DET_ML \(\approx\) 7 / 9 / 12 / 18, returning to 1 only by about 25–40, and deepening to 0.73 at the low end (DET_ML 0.75), with the deficit lasting into the catalogue regime, unlike the toy (R059: gone by DET_ML \(\approx\) 12). Claude's triage confirmed R062 and, with the prior edge/slope test (S_true \(\geq\) 7 / \(\geq\) 8.5, slope −2.5 \(\to\) −1.5), showed the law is prior-independent. R063 then gave the prior scope: the width is prior-independent, while the mean holds only for the −2.5 prior (moving with the prior below DET_ML \(\lesssim\) 17). RT-60 and RT-61 were launched after. The T12 stage-27/28 refresh was deployed as 663838ca (data as of 17:31 UTC). Sources: 07 R062, R063 and the two triage reviews.

2026-10-19：宽度红队的一天，机制被重新标注并两次改写（R064–R080）。二十二条修订记录（R064–R085）全部来自宽度红队的分诊，且全部被复算；手册五次部署。当天的主线是机制：真实 PSF 的插入式公式没有错，错的是把它放在偏高的测量振幅上。2026-10-19: a day of width red teams, with the mechanism re-labelled and rewritten twice (R064–R080). Twenty-two revision records (R064–R085), all from width red-team triages and all replayed; the handbook was deployed five times. The day's main line was the mechanism: the real-PSF plug-in formula is not wrong — what is wrong is evaluating it at a biased-high measured amplitude.

2026-10-20：主线转回 emldetect 的一天，机制先被限定、后被统一，晚上规律被逐个检验并复核（R081–R107）。五支宽度红队当天分诊：计数依赖修订（R081：共同掩膜上 +0.80 pp，R080 的 +0.86 pp 是掩膜假象）、L0 低计数中心项超额是 L0 流水线自己的性质（R082/R084/R085：Ŝ 分箱 +0.389 / +0.14–0.18 / +0.058 \(\pm\) 0.007；72/28 归因后来经 R087 撤回，成因开放）、penalty 错配机制先被写成不解释 L1 宽度（R083），当晚被 R086 推翻——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]（预测 std z 0.919，实测 0.918），同一 \(V_{3p}\) 机制同时决定 L0 与 emldetect 的宽度；高 Ŝ 偏移是对照截断假象（R085）。PI 当天指示主线回到 emldetect，并在 04:51 EDT 定下终点线：单图 emldetect 条件律 p(DET_ML | Ŝ, b, index)。专家（astra）结论“目标未达成”，接受 P0（0b 验收标准）与 P1（0b 试点后的验收清单、emldetect 条件律表）。当天晚些时候再分诊三支：任务 0b 完成（R089）——emldetect 与 L0 高计数一致、DET_ML ≲ 7 处 emldetect 中心低 0.16–0.53 z、宽 0.05–0.14，所以 L0 低计数中心项超额不是 emldetect 的性质，单图 emldetect 规律必须在 emldetect 输出上建立；RT-85/RT-86（R088/R090）——公共 2 参数 index 规律被否定，须建显式含 b 的 p(DET_ML | Ŝ, b, index)；RT-84（R091）——按计数修正的 V₃p 只在 emldetect 输出上校准时成立（留出 0.976 [0.956, 0.993]），换独立模拟过修。主线下一步（T-LAW）在 L1 输出上建条件律。手册部署为 8b796a9d。下午规律阶段（R092–R104）：RT-87 把 R089 的两条读数收窄（中心差距在 Ŝ 2.2–9.5 成立，"宽度更大"降级为建议，R092）。四支候选规律在 M2 能段 4 上依次被检验：LAW-1 C 10/72（R093）、LAW-2 D 48/72（R095）、LAW-3 DQ 42/72（R096）、LAW-4 MIX3d 35/72（R101）；能段 1 上 C 4/72（R097）、D 重拟合 37/72 而能段 4 系数原样迁移只有 17/72（R102）。打分手本身错了两次并修好：R094 给出噪声底（一条完美规律平均也只通过约 61/72，第 5 百分位 48），R100 撤回 R096 的"D 以 p = 0 被全局拒绝"（零假设是单一 index 的 72 项而观测 Q 是 288 项，对合并零假设 p = 0.75 或 0.0025）——失败的符号格局成立、显著性开放。最后一支独立的评分器（NULL-1，R103）在"恰好为真"的数据上验证后给出唯一结论：四条候选全部被拒（D 最好，48 对零假设均值 63.3，p_pass = 0.0015），而 eval_lib.gof() 的位置重采样零假设连 10/72 的 C 都接受（p = 0.73），不再用作验收零假设。RT-91b（R104）把 D 拆开：宽度这一半可以推广，中心这一半不能——24 个均值失败全部为负、集中在中计数、随 index 增长；12/12 的高计数锚完全取决于位置 14 被分到哪一半。CENTRE-POS 给出否定答案：平滑的位置项把留出失败从 24 个推到 58–72 个，位置不是缺失的中心变量。PI 于 18:47 EDT 批准符号回归试点（PySR + SRHarness，LLM 后端限 NIM 模型）。当天最后三条记录是复核：R105（CENTRE-POS，已复算）驳回中心项的位置假设——平滑位置协变量让留出均值 z 更差（失败 24 \(\to\) 58–72），唯一有用的是逐格中心截距（18 个参数，失败 24 \(\to\) 11、负和 −2.28 \(\to\) −0.80），离轴斜率不加分；R106（RT-93 分诊）指出 LAW-4 的 Q 部分重复 R100 的尺度错误，而 MIX3d 的亏缺在宽度/覆盖率、其 index 层面的中心行为优于 D；R107（RT-94 分诊）确认 NULL-1 的判定全部成立，但把它的验证加上限定（pos_sd 恢复是循环论证；跨格相关 +0.44 未入零假设，通过数零假设 sd 小约 30 %，D 的 p_pass 变约 10 倍但无判定翻转；p_q 是 < 1/2000 而不是 0）。随后派出混合规律任务（D 的宽度 + MIX3d 的中心）与 RT-95。2026-10-20: the day the main line turned back to emldetect — the mechanism first scoped, then unified, and the candidate laws tested one by one and re-checked in the evening (R081–R107). Five width red teams were triaged that day: the count dependence revised (R081: +0.80 pp on a common mask, R080's +0.86 pp a mask artefact), the L0 low-count centre excess a property of the L0 pipeline itself (R082/R084/R085: Ŝ-bin gaps +0.389 / +0.14–0.18 / +0.058 \(\pm\) 0.007; the 72/28 attribution was later withdrawn by R087, cause open), the penalty-mismatch mechanism first written up as not explaining the L1 width (R083) and reversed that night by R086 — 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 (predicted std z 0.919, observed 0.918), one \(V_{3p}\) mechanism carrying both the L0 and the emldetect widths; and the high-Ŝ offset a control-truncation artefact (R085). The PI redirected the main line back to emldetect that day and set the finish line at 04:51 EDT: the single-image emldetect conditional law p(DET_ML | S-hat, b, index). The expert (astra) verdict was “goals not reached”, accepting P0 (0b acceptance criteria) and P1 (the post-0b acceptance manifest and the emldetect conditional-law table). Three more were triaged later that day: task 0b completed (R089) — 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 centre excess is not an emldetect property and the single-image emldetect law must be built on emldetect output; RT-85/RT-86 (R088/R090) — the common 2-parameter index law is refuted and the law must be p(DET_ML | Ŝ, b, index) with explicit b; RT-84 (R091) — the count-corrected V₃p works only when calibrated on emldetect output (held-out 0.976 [0.956, 0.993]) and over-corrects when transferred from an independent simulation. The next main-line step (T-LAW) builds the conditional law on L1 output. The handbook was deployed as 8b796a9d.The afternoon law stage (R092–R104): RT-87 narrowed two readings of R089 (the centre gap holds at Ŝ 2.2–9.5; the “larger width” clause is downgraded to a suggestion, R092). Four candidate laws were tested in turn on M2 band 4: LAW-1 C 10/72 (R093), LAW-2 D 48/72 (R095), LAW-3 DQ 42/72 (R096) and LAW-4 MIX3d 35/72 (R101); on band 1, C 4/72 (R097) and the D refit 37/72 with the unchanged band-4 coefficients transferring at only 17/72 (R102). The scoring itself was wrong twice and fixed: R094 gives the noise floor (a perfect law passes only about 61/72 on average, 5th percentile 48) and R100 withdraws R096’s “D globally rejected at p = 0” (that null was a single-index 72 terms against a 288-term observed Q; against a pooled null p = 0.75 or 0.0025) — the sign pattern of the failures stands, its significance is open. One independent scorer (NULL-1, R103), validated on exactly-true data, gives the single verdict: every candidate is rejected (D is best, 48 against a null mean of 63.3, p_pass = 0.0015), while eval_lib.gof()’s position-resampling null accepts even the 10/72 candidate C (p = 0.73) and is no longer used as the acceptance null. RT-91b (R104) takes D apart: the width half generalises, the centre half does not — the 24 mean failures are all negative, concentrated at mid count and grow with index, and the 12/12 high-count anchor hinges entirely on which half position 14 falls into. CENTRE-POS gives a negative answer: a smooth position term pushes the held-out failures from 24 to 58–72, so position is not the missing centre variable. The PI approved a symbolic-regression trial at 18:47 EDT (PySR + SRHarness, LLM backend restricted to NIM models).The day's last three records are the re-checks: R105 (CENTRE-POS, replayed) refutes the position hypothesis for the centre term — smooth position covariates make the held-out mean z worse (failures 24 \(\to\) 58–72), and the only form that helps is a per-cell centre intercept (18 parameters, failures 24 \(\to\) 11, negative sum −2.28 \(\to\) −0.80), with the off-axis slope adding nothing; R106 (the RT-93 triage) notes that LAW-4's Q part repeats R100's scale error, while MIX3d's deficit is width/coverage and its index-level centre behaviour is better than D's; and R107 (the RT-94 triage) confirms NULL-1's verdicts all stand but qualifies its validation (the pos_sd recovery is circular; the cross-cell correlation +0.44 is not in the null, the pass-count null sd is about 30 % too small, D's p_pass changes by about 10\(\times\) with no verdict flipping; p_q is < 1/2000, not 0). The hybrid-law task (D's width + MIX3d's centre) and RT-95 were then dispatched.

2026-10-21：规律阶段被一次专家复核改写（R108–R118）。凌晨四条记录先把昨天的结论各自收紧：R108（RT-95 分诊）把 R105 的"位置不是缺失变量"加上限定——协变量是以 R 单位进的而检验是 z 单位的，那个拟合项约 90 % 只是一个与协变量无关的 +0.44 R 单位常数；视场边缘异常值是真实的（PSF 外插 / 低支撑），逐格截距不能跨 12 个位置分半推广（28.0 \(\to\) 33–34 失败）。R109（SR-pysr / SR-srharness）给出符号回归试点的答案：没有公式配得上 D，最好的 SRHarness 三常数式 28/72，迁到能段 1 是 22 而 D 自己的重拟合是 37。R110/R112（LAW-HYBRID 两臂一致）：宽度没问题，H2（D + 12 系数 index 无关中心多项式）是唯一有用的中心修补，部分且不跨分半稳健。R111（CENTRE-Z）：把 D 的中心改成 z 单位重拟合（288 参数）后留出判定一格不变（24 \(\to\) 24，负和 −2.28 \(\to\) −2.28），驳回 R108 的单位假象说。R113（RT-97）：SR 试点的目标、评分器、只用拟合半三点确认，拟合半增益不迁移，"index 无关所以 \(\beta\) 不必要"被驳回。白天，专家（astra）的报告被采纳为 R114，规律阶段的读法从此改写：旧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，零假设的位置散布还从候选自己的留出残差估计。于是"宽度可以推广"只能读作位置内；72 格是 18 个可观测格 \(\times\) 4 个 index，约 19 个有效实验（跨 index 相关 +0.81…+0.92）；Monte-Carlo p 必须写成 k/N 加一个界（0/2000 \(\to\) \(\leq\) 0.0015 单侧 95 %）。R117（RT-98，已复算）把这一点做到精确：位置内二阶矩对任何常数位置偏移恒等不变（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；自适应零假设在这批数据上只让 D 的 Q 宽约 6.5 %（不是 25–93 % 最坏情形）。R118（SCORER-CONTRACT v2，已复算）给出新打分手，两个目标分开：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），重尾 df \(\approx\) 3。中心项也有了新读数（R116，CENTRE-DECOMP + RT-99）：主导成分是 D 表自身的样本内表征误差——同一位置内失败 24 \(\to\) 3、未见位置 21 \(\to\) 8（72 参数诊断修正），所以"中心差距是位置泛化缺口"这句不再成立；专家那条先验混合恒等式作为代数成立（闭合 1e-9），但其协方差项在失败格里比误差大 2–8 倍且已被 D 自己的 \(\beta\) 斜率抵消。宽度读数被限定（R115，RT-96b）：0.976 的留出宽度只是一次偏低的分半，40 个按半径分层的分半平均 0.998；6 参数修正只通过 25/72 个宽度格。PI 当天三次修订进度指标（04）：05:39 EDT 定下验收目标——不是每个位置，而是大多数位置逐位置准确，位置域越大越好（如离轴 14′ 内全部位置），允许 MASKFRAC 之类的质量截断，下一个测量就是逐位置通过比例对离轴半径；11:42 EDT 改为位置上没有通过/不通过阈值——进度指标是已解决位置的比例（与面积），可以从 0 开始，任何增长都算进展，90 % 默认值撤回；06:20 EDT 批准 WIDTH-2ND（分解可观测 \(\hat{S}\) 箱的宽度亏缺 + 推导 \(\mathrm{Var}[\Delta C \mid \hat{\theta}]\) 的二阶修正）。RT-100 与 RT-101 当天分诊：RT-100（R120）扫 POS-CAL 的域——已指出 MASKFRAC \(\geq\) 0.9 在现有行上近乎空操作（99.92 % 已满足）、否定 RT-98 的 tau 地板塌陷，并给出进度曲线的第一个点：能段 4 的 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，离轴三分位 0.767 / 0.605 / 0.542，失败位置散布而非视场外缘效应；RT-101（R119）审 CENTRE-DECOMP 的三条主论断（已指出"新重复"臂共用源表、三臂读数并非几乎相同、几何残差只是 1.3–2.1\(\sigma\)）。当天晚些时候 WIDTH-2ND 完成（R121）：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 \(S_{true}\) 混合与条件均值项都可忽略），新的 \(O(n^{-1})\) 闭式因子在 \(b = 0.01\) 精确、在 \(b = 0.04/0.16\) 欠修，预注册判据 FAIL（8 个箱里 3 个超出 2\(\sigma\)）；残差跟随 \(V_{3p}\) 的相消因子（\(R^2\) 0.89），下一步是沿 S-score 方向的重求和/精确条件方差。2026-10-21: the law stage rewritten by one expert review (R108–R118). Four early-morning records tightened yesterday's conclusions: R108 (the RT-95 triage) qualifies R105's "position is not the missing variable" — the covariates were entered in R units against a z-unit test, the fitted term is about 90 % a covariate-independent +0.44 R-unit constant, the FoV-edge outliers are real (PSF extrapolation / low support), and the per-cell intercept does not generalise across 12 position splits (28.0 \(\to\) 33–34 failures). R109 (SR-pysr / SR-srharness) answers the symbolic-regression trial: no formula matches D, the best SRHarness three-constant form gives 28/72 and transfers to band 1 at 22 against D's own refit of 37. R110/R112 (LAW-HYBRID, both arms agreeing): the width is fine and H2 (D plus a 12-coefficient index-free centre polynomial) is the only centre repair that helps, partial and not robust across splits. R111 (CENTRE-Z): refitting D's centre in z units (288 parameters) leaves the held-out verdict unchanged in every cell (24 \(\to\) 24, negative sum −2.28 \(\to\) −2.28), refuting R108's units-artefact reading. R113 (RT-97): the SR trial's target, scorer and fit-half-only are all confirmed, the fit-half gain does not transfer, and "index-free, so \(\beta\) unnecessary" is refuted. In the daytime the expert (astra)'s report was accepted as R114, and the law stage's reading was rewritten from there: the old scorer mixed three estimands — the width is a within-position weighted std, the mean and coverage pool rows across positions, and the null's position scatter is estimated from the candidate's own held residuals. So "the width generalises" can only be read as within position; the 72 cells are 18 observable cells \(\times\) 4 indices, about 19 effective experiments (cross-index correlation +0.81…+0.92); and a Monte-Carlo p must be written as k/N with a bound (0/2000 \(\to\) \(\leq\) 0.0015 one-sided 95 %). R117 (RT-98, replayed) made this exact: the within-position second moment is identically invariant (1e-15) to any constant per-position offset, so a law carrying a 0.30 z per-position offset still passes width 70/72 while cov68 failures rise from 4 to 29; the adaptive null widens D's Q by only about 6.5 % on this dataset (not the 25–93 % worst case). R118 (SCORER-CONTRACT v2, replayed) gives the new scorer, with two targets separated: 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 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), heavy-tailed with df \(\approx\) 3. The centre term has a new reading too (R116, CENTRE-DECOMP + RT-99): the dominant component is an in-sample representation error of the D table itself — within position the failures go 24 \(\to\) 3 and at unseen positions 21 \(\to\) 8 (a 72-parameter diagnostic correction), so "the centre gap is a position-generalisation gap" no longer holds; the expert's prior-mixing identity holds as algebra (closing to 1e-9), but its covariance term is 2–8\(\times\) the error in the failing cells and is already cancelled by D's own \(\beta\) slope. The width reading is qualified (R115, RT-96b): the 0.976 held-out width was one low split, 0.998 averaged over 40 radius-stratified splits, and the 6-parameter correction passes only 25/72 width cells. The PI amended the progress metric three times that day (04): at 05:39 EDT the acceptance target — not every position, but per-position accuracy at most positions, over the largest possible position domain (e.g. all positions within 14′ off-axis), with quality cuts such as MASKFRAC allowed, so the next measurement is the per-position pass fraction against the off-axis radius; at 11:42 EDT it became 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; and at 06:20 EDT WIDTH-2ND was approved (decompose the observable-\(\hat{S}\)-bin width deficit and derive the second-order correction to \(\mathrm{Var}[\Delta C \mid \hat{\theta}]\)). RT-100 and RT-101 were triaged the same day: RT-100 (R120) swept the POS-CAL domain — noting that MASKFRAC \(\geq\) 0.9 is nearly vacuous on the existing rows (99.92 % already satisfy it), refuting RT-98's tau-floor collapse, and giving 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), 1/33 on band 1, off-axis terciles 0.767 / 0.605 / 0.542, with the failing positions scattered rather than at the outer field of view; RT-101 (R119) audited CENTRE-DECOMP's three load-bearing claims (noting that the "new reps" arm shares the source list, the three arms' readings are not nearly identical, and the geometry residual is only 1.3–2.1\(\sigma\)). Later that day WIDTH-2ND finished (R121): the width deficit is re-attributed to the conditional variance \(\mathrm{Var}[\Delta C \mid \hat{\theta}] < V_{3p}(\hat{\theta})\) (both within-bin \(S_{true}\) mixing and the conditional-mean term are negligible), the new \(O(n^{-1})\) closed-form factor is exact at \(b = 0.01\) and under-corrects at \(b = 0.04/0.16\), and the pre-registered verdict is FAIL (3 of 8 bins beyond 2\(\sigma\)); the residual tracks the \(V_{3p}\) cancellation factor (\(R^2\) 0.89), and the next step is a resummed or exact conditional variance along the S-score direction.

2026-10-22：规律阶段的判据、位置项与宽度规则同时被改写（R125–R152）。当天的第一条主线是进度判据本身："每个条目都在容差 \(\varepsilon\) 内"被证明是噪声受限的（R128/R129/R133，最好候选的天花板只有 0.21–0.27），改为每个位置一个合并偏移（R131）后天花板回到 1.000；PI 于当天定下 \(\varepsilon\) 0.05 用于比较规律、0.10 用于跟踪进度（R137，在 DET_ML 6–10 窗口约合 0.12），打分手冻结为 v1（R147）。R131 的"规律就压在位置盲地板上"随后被 R138 的地板改正推翻。第二条主线是位置效应：它不是源表的性质（R126/R132），而是几何的、且两个能段共享（r 0.936，R134/R145）；R144 读出"把注入点移到像素中心会把偏移放大 2 倍"，该读数已由 R150 撤回（打分手伪迹，比值 1.16）；第一个拟合前位置项（拟合盒背景水平）把最好候选从 0.545 抬到 0.727（p 0.003–0.004），尚未在新位置上检验（R139/R148）。第三条主线是宽度规则：一条完美的规律在严格的"8 箱全过"规则下只有 66.8 % 的通过率（R143），PI 当天批准改为8 箱允许 1 箱失败（约 95 %）；此后第一条可观测箱级规律 V̄ 在 \(\beta\) −1.5 处通过（1/8），但 R152 指出带宽是在测试半上选的，40 个分半里 \(P(\leq\) 1 次失败\()\) 只有 0.79；迁到 emldetect 全统计量行上仍失败 5/5 箱（R151）。当天最后定下普适性设计（R149）：Tier 0 先跑，Tier 1 两项随后。

2026-10-23：普适性 Tier 0 跑完、宽度与目录的读数被改写（R157–R185）。当天最重要的转折是一次口径错误被撤回：Tier 0 的第一次打分（A0-SCORE，R158）把 M2 能段 1 记为 FAIL（0.606，低于线 0.627）；红队 RT-117（R162）发现该打分手把位置项按行除进了 z，而冻结口径是从合并偏移中减掉——同一物理项放错分母，一条通过的定律就变成了"失败"。按冻结口径重打（零新参数）：M2 能段 1 通过（24/33 = 0.7273），但 M1 能段 1 是一个已分辨的中心失败（8/21 = 0.381，配对差 −0.333 [−0.619, −0.048]），M1 其余能段分辨不了（MDS 约 0.30），高计数锚在 M1 能段 4 为 0/12（R169，A0-RESCORE）。调度方把 R158 记成 FAIL 时没有检查口径——这是调度方自己的错误，此后每个新打分手先在冻结基准上逐位复算。宽度侧：细锚方差律在 emldetect 行上达到可观测神谕地板（2/9 失败箱，神谕自身也失败 2/9，R164），经验合成的簇自助 95 % CI 为 [1, 7] %（R166）；"8 箱允许 1 箱失败"按实际箱数（9–10）重标定为 k* 2–3，零假设的选择留给 PI（R171）；低 b 高 \(\hat{S}\) 的宽度失误定位到少数边缘拟合位置上的探测器掩膜截断（R170），预注册规则"中位 MASKFRAC_EML < 0.98 的位置排除"在留出数据上去掉该失误，代价约 10 % 的位置（R180）。目录侧：冻结目录律在 \(\mu_0 \lesssim\) 250 通过（R160/R163）；250 以上 \((f\cdot\mu)^2\) 恢复 MOS 覆盖率但其 PSF 标定依据已撤回（R168/R174），逐相机 \(c\cdot\mu^p\) 在不相交 \(\mu_0\) 区域之间不迁移（R185）。位置项在 27 个新位置上的检验为 AMBIGUOUS（+0.148 [−0.037, +0.333]，R165），且 8/27 个新位置受源拥挤污染，因此不能读作位置项的检验（R173）；139 个隔离位置（NN \(\geq\) 45 px，约 500 virgo CPU-h）的设计已就绪（R177）。M1 能段 4 的失败随后定位到三个灾难位置（占锚行的 14.9 %；排除后格从 0/12 变 9/12；RT-126 安慰剂检验已跑——29/31/35 是全部 1330 个留出三位置排除中唯一达到 9/12 的，不是选择，而"占合并锚均值 98 %"的措辞已撤回（小数字之比、簇自助区间很宽），R186；该失败不是掩膜机制——这三个位置的 MASKFRAC 是 0.994，掩膜规则下 M1 能段 4 仍为 0/12，R187）。修订的 Tier 1 方案（5 臂，核心约 574 virgo CPU-h）待 PI 批准（R183）。2026-10-23: universality Tier 0 completed, and the width and catalogue readings were rewritten (R157–R185). The day's most important turn was the withdrawal of a convention error: Tier 0's first scoring run (A0-SCORE, R158) had recorded M2 band 1 as a FAIL (0.606, below the bar 0.627); red team RT-117 (R162) found that scorer had divided the position term row by row inside z, whereas the frozen convention subtracts it from the pooled offset — the same physical term placed in the wrong denominator turned a passing law into a "failure". Re-scored under the frozen convention (zero new parameters): M2 band 1 passes (24/33 = 0.7273), but M1 band 1 is a resolved centre failure (8/21 = 0.381, paired difference −0.333 [−0.619, −0.048]), the remaining M1 bands cannot be resolved (MDS about 0.30), and the high-count anchor on M1 band 4 is 0/12 (R169, A0-RESCORE). The dispatcher had recorded R158 as a FAIL without checking the convention — the dispatcher's own error; from then on every new scorer is first replayed bit-for-bit on the frozen baseline. On the width side: the fine-anchor variance law reaches the observable-oracle floor on emldetect rows (2/9 failing bins, the oracle itself failing 2/9, R164), with the empirical composition's cluster-bootstrap 95 % CI at [1, 7] % (R166); the "at most 1 failing bin of 8" rule was re-calibrated to the actual bin count (9–10) as k* 2–3, with the choice of null left to the PI (R171); the low-b high-\(\hat{S}\) width miss was localized to detector-mask truncation at a few edge fit positions (R170), and the pre-registered rule "exclude positions with median MASKFRAC_EML < 0.98" removes it on held data at a cost of about 10 % of positions (R180). On the catalogue side: the frozen catalogue law passes for \(\mu_0 \lesssim\) 250 (R160/R163); above 250 the \((f\cdot\mu)^2\) term restores MOS coverage but its PSF-calibration justification is withdrawn (R168/R174), and the per-camera \(c\cdot\mu^p\) form does not transfer across disjoint \(\mu_0\) regions (R185). The position term's test at 27 fresh positions is AMBIGUOUS (+0.148 [−0.037, +0.333], R165), and 8 of the 27 fresh positions are source-crowded, so it cannot be read as a test of the position term (R173); a design with 139 isolated positions (NN \(\geq\) 45 px, about 500 virgo CPU-h) is ready (R177). M1 band 4's failure was then localized to three catastrophic positions (14.9 % of anchor rows; excluding them moves the cells from 0/12 to 9/12; the RT-126 placebo test has run — 29/31/35 is the unique three-position drop of all 1330 held three-drops reaching 9/12, so it is not selection, and the "98 % of the pooled anchor mean" wording is withdrawn (a ratio of small numbers with a wide cluster-bootstrap CI), R186; the failure is not the mask mechanism — those three positions have MASKFRAC 0.994, and under the mask rule M1 band 4 stays at 0/12, R187). The revised Tier 1 plan (5 arms, core about 574 virgo CPU-h) awaits the PI (R183).

2026-10-24：Tier 1 第一臂打分、背景通道出定律缺口、代价锚改正（R189–R231）。当天最重要的转折是背景通道：T1-c 用 527 次重复对比真背景与 DR14 估计背景，冻结定律在后者的窗口偏移是 +0.174 z（簇区间 [0.165, 0.260] 不含 \(\varepsilon\) 0.10；不开窗口 0.151，不是选择伪影），红队 RT-139 排掉十条伪影通道后机制定为 \(z_B - z_A \approx 48.2 \times \Delta b_{\rm PSF}\)——定律需要一个命名的背景估计参数（R224/R227）。候选 C2（\(\gamma = 0.921 + 0.0041(\mathrm{off} - 3')\)）把 9 个留出位置的偏移从 +0.101 压到 −0.052 z，但 RT-140 的直接物理核对说物理尺度约 0.86 不是 0.92，且 C2 过修——暂定于一个观测，进定律前要在第二个观测上确认（R228/R230）。第一臂 T1-e：M2 能段 1 中心确认（24/33 = 0.727，区间 [−0.152, +0.061]），宽度 4 个失败箱对 \(k^*\) 2 先记 FAIL、红队复核后改记 AMBIGUOUS（超出真实、单侧约 2 %，但箱边移 −1/3 个箱就翻），锚 8/12；执行方的"整体 REFUTED"不是预注册组合，只记三条独立判定（R220/R229）。代价锚：实测 0.97 CPU-s/行（区间 0.86–1.3），旧 0.4 在算术上就是错的（2.4 倍）；T1-a 按 PI 选项 A 只跑 M2（N = 115，149–192 CPU-h）并已发射，PN b1/b4 不重跑（lang2 源树已不在，完整 py 输入要约 397 GB）——这是调度方自己的错误被写进记录的延续：记录一个"失败"前先核对口径，记录一个"成本"前先实测。星表律：DET_ML 有序残差确认为标注同义反复（实/零之比约 0.93）；断幂律过 5/6 主格但 M2 低 \(\mu_0\) 的补丁是事后选择；eta 振幅修好合并宽度（1.026/0.995/1.021）但逐格形状失败（最好 16/87，残差重尾 PN 中位超额峰度 +2.9），且逐格规则本身要按正确模型零假设重标（RT-141：正确模型下也只过 37/87）——下一步是尾部项。M1 坏位置：PSF 加权参考被预注册检验驳回（0/12 → 0/12，护栏 5/6 块失败），地图错配是伴随物不是原因，原因保持开放（R200/R205/R214）。2026-10-24: Tier 1's first arm scored, the background channel opened a gap in the law, and the cost anchor was corrected (R189–R231). The day's pivotal turn was the background channel: T1-c contrasted the truth background with the DR14-estimated background over 527 repeats, and the frozen law's window offset under the latter is +0.174 z (cluster interval [0.165, 0.260] excluding \(\varepsilon\) 0.10; 0.151 with no window, so not a selection artefact); after red team RT-139 excluded ten artefact channels the mechanism was fixed as \(z_B - z_A \approx 48.2 \times \Delta b_{\rm PSF}\) — the law needs a named background-estimation parameter (R224/R227). The candidate C2 (\(\gamma = 0.921 + 0.0041(\mathrm{off} - 3')\)) brings the offset over 9 held positions from +0.101 to −0.052 z, but RT-140's direct physical check puts the physical scale at about 0.86 rather than 0.92, and C2 over-corrects — provisional on one observation, with a second observation required before it enters the law (R228/R230). The first arm, T1-e: M2 band 1's centre is confirmed (24/33 = 0.727, interval [−0.152, +0.061]); the width, 4 failing bins against \(k^*\) 2, was first recorded FAIL and re-flagged AMBIGUOUS after the red-team re-check (the excess real, one-sided about 2 %, but a −1/3-bin edge shift flips it); the anchor is 8/12; the executor's "overall REFUTED" is not a pre-registered combination, so only the three separate verdicts are recorded (R220/R229). The cost anchor: the measured cost is 0.97 CPU-s/row (band 0.86–1.3), the old 0.4 being arithmetically broken (2.4\(\times\)); T1-a runs M2 only under the PI's option A (N = 115, 149–192 CPU-h) and has been launched, while PN b1/b4 are not re-run (the lang2 source tree is gone and a full py input would be about 397 GB) — a continuation of the dispatcher's own error being written into the record: check the convention before recording a "failure", and measure before recording a "cost". The catalogue law: the DET_ML-ordered residual is confirmed a labelling tautology (real/null about 0.93); the broken power law passes 5/6 primary cells but M2's low-\(\mu_0\) patch is a post-hoc choice; the eta amplitude restores the pooled width (1.026/0.995/1.021) while the per-cell shape fails (the best 16/87, the residual heavy-tailed with PN median excess kurtosis +2.9), and the per-cell rule itself must be re-calibrated against the correct-model null (RT-141: it passes only 37/87 even under a correct model) — the next candidate is a tail term. The M1 bad positions: the PSF-weighted reference is refuted by a pre-registered test (0/12 → 0/12, the guard failing on 5/6 blocks), the map mismatch being a correlate rather than the cause, so the cause stays open (R200/R205/R214).

T234 的机制被重新标注（R064，RT-60 分诊，已复算）。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，从不小于 1）——公式在真实参数处是对的，甚至略偏窄。可观测箱里的偏窄来自把凸的 \(V_{3p}\) 放在偏高的 \(\hat{S}\) 上计算（DET_ML 3 时 Ŝ 高 +14.5 %、30 时 +1.2 %，所以 \(E[V_{3p}(\hat{S})]/V_{3p}(S_{true})\) 从 1.63 降到 1.05），再加箱内 S_true 混合项；拟合器最优性已排除。“方差最多高估 1.9 倍”的说法删除（那是一个均值被移过的分布的标准差比）。T234's mechanism is re-labelled (R064, the RT-60 triage, replayed). 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 (DET_ML 3 / 5 / 9 / 15 / 30, never below 1) — the formula is right at the true parameters, if anything slightly narrow. The observable-bin narrowing comes from evaluating the convex \(V_{3p}\) at a biased-high \(\hat{S}\) (+14.5 % at DET_ML 3, +1.2 % at 30, so \(E[V_{3p}(\hat{S})]/V_{3p}(S_{true})\) runs from 1.63 down to 1.05), plus a within-bin S_true mixture term; fitter optimality is ruled out. The “variance over-stated by up to 1.9\(\times\)” wording is dropped (it was the std ratio of a shifted distribution).

R064 预言的修正被推翻（R068，RT-63 分诊，已复算）。把 \(V_{3p}\) 改在可观测的 Cox–Snell 偏差校正振幅 \(\hat{S} - b_S(\hat{S})\) 上求值，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。08 §7 第 0a 项以否定结论关闭：没有任何可观测的点修正能闭合真实 PSF 的宽度。The correction R064 predicted is refuted (R068, the RT-63 triage, replayed). 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. 08 §7 item 0a is closed negative: no observable point correction closes the real-PSF width.

这条规律跨位置成立；L1 没有它（R065、R066、R067，RT-62 / RT-61 / RT-64 分诊，均已复算）。L0 生产数据 67 个位置（离轴 0–13′，我们的全局搜索拟合器）上都成立，数值比单一位置低约 0.02，合并 std z 0.840 / 0.894 / 0.909 / 0.959 / 0.984 于 DET_ML \(\approx\) 4.6 / 7.9 / 11.1 / 22.8 / 36.8，到 DET_ML \(\approx\) 36–60 才回到 1。L1（emldetect + e114 核，名义位置）按可观测 S_eml 分箱后没有这条低端偏窄（std z 0.94–0.98 于 DET_ML 1–6、\(\approx\) 0.95 于 6–10、12–20 以上约 1，最低一箱反而宽 1.31）；T228 里用 b_true 标注的格子数值降为历史。匹配比较下 L1 − L0 = +0.24 / +0.19 / +0.12 / +0.08 / +0.07 于 DET_ML \(\approx\) 2 / 3 / 4 / 5 / 7，但那是两条流水线加两种打分与两个振幅估计量之间的差，不归因于拟合器——决定性实验（08 §7 第 0b 项）是在 L0 的真实 PSF 图像上跑 emldetect。The law holds across positions; L1 does not show it (R065, R066, R067, the RT-62 / RT-61 / RT-64 triages, all replayed). It holds at all 67 L0 production positions (off-axis 0–13', our global-over-support fitter), \(\approx\) 0.02 below the single position, with pooled std z 0.840 / 0.894 / 0.909 / 0.959 / 0.984 at DET_ML \(\approx\) 4.6 / 7.9 / 11.1 / 22.8 / 36.8, reaching 1 only at DET_ML \(\approx\) 36–60. L1 (emldetect + the e114 kernel, nominal position) binned by the observable S_eml shows no such low-end narrowing (std z 0.94–0.98 at DET_ML 1–6, \(\approx\) 0.95 at 6–10, about 1 above 12–20, with the lowest bin wide at 1.31); T228's b_true-labelled cell numbers are demoted to history. In the matched comparison L1 − L0 = +0.24 / +0.19 / +0.12 / +0.08 / +0.07 at DET_ML \(\approx\) 2 / 3 / 4 / 5 / 7, but that is a gap between two pipeline + scoring pairs and two amplitude estimands, not attributable to the fitter — the decisive run (08 §7 item 0b) is emldetect on the L0 real-stamp images.

手册两次部署。T12 第 29–30 阶段（数据截至 2026-10-18 23:58 UTC）经审阅（PASS with banner：图题方向反了、§15 一句假比较、图注行数错、62 处双反斜杠）后部署为 c742d5e6；第 31 阶段修完五项后部署为 f0f2b93d（数据截至仍为 2026-10-18 23:58 UTC，横幅只列 R064–R067）。RT-65 与 RT-66 发出，其后分诊为 R070 与 R069；RT-67 与 RT-68 接着发出，RT-68 分诊为 R071，RT-67 也分诊为 R072（更好的偏差修正路线关闭，残差与先验无关（\(\leq\) 0.6 SE），机制标签后来被 R077/R078 改写为 \(V_{3p}\) 的近乎全抵消）；RT-69 与 RT-70 接着发出，RT-69 又分诊为 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）。RT-70 至 RT-73 与 RT-75 随后也分诊：R074 把 0.5\(\times\) 背景下“闭合到 0.99–1.00”的读数废弃（单一种子且过度修正），R075 显示 L0 的宽度与覆盖率依赖打分，R076 判定 z_off 是 L0 打分（R065 的偏窄恢复为 L0 宽度律），R077/R078 把机制改写为“Fisher penalty 的 \(\approx\) 1 % 错配被 \(V_{3p}\) 的近乎全抵消放大”，且错配随 PSF 的中心集中度上升。RT-74 随后也分诊为 R079：L0 的低 DET_ML 中心项依赖拟合离轴 \(r\)（固定 \(\hat{S}\) 下 +0.71 \(\to\) +0.23 与 +0.10 \(\to\) −1.34），超过两个已验证的纯构造对照 +0.23…+0.50 z（DET_ML \(\lesssim\) 8），R076 的“近乎抵消”注记已撤回，归因开放。RT-76 随后也分诊为 R080：候选的二阶 penalty 修正未达成（Bartlett 项符号反了，beyond-Fisher 项过度修正 1.3–7.4\(\times\)），错配跟随 PSF 翼结构与计数（\(\hat{S}\), b）。五支宽度红队随后也分诊（R081–R085）：计数依赖修订（\(\hat{S}\) 约 5 的 +0.86 pp 是掩膜假象）、L0 低计数中心项超额是 L0 流水线自己的性质（72/28 归因后经 R087 撤回）、penalty 错配机制先记为不解释 L1 宽度（R083）、后被 R086 推翻（合并估计量下 L1 错配 +1.19 %，同一机制同时决定两条流水线的宽度）、高 \(\hat{S}\) 偏移是对照截断假象。PI 当天指示主线回到 emldetect 并定下终点线（单图条件律），任务 0b 与 RT-84 派出，晚间 0b 完成（R089）、RT-84/RT-85/RT-86 分诊（R091、R088/R090），主线下一步 T-LAW 在 L1 输出上建条件律。第 33 阶段（stage 32 的重启）折叠 R064–R071 后部署为 25620f8d，第 34 阶段折叠 R072–R073 后部署为 a94494ac，第 35+36 阶段折叠 R074–R080 后部署为 8b796a9d。Two handbook deploys. T12 stages 29–30 (data as of 2026-10-18 23:58 UTC) were reviewed (PASS with banner: a reversed figure title, a false §15 comparison sentence, a wrong caption and 62 doubled backslashes) and deployed as c742d5e6; stage 31 fixed all five items and deployed as f0f2b93d (data still as of 2026-10-18 23:58 UTC, the banner listing R064–R067 only). RT-65 and RT-66 were launched and were later triaged as R070 and R069; RT-67 and RT-68 were then launched; RT-68 was triaged as R071 and RT-67 as R072 (the better-bias-correction route is closed, the residual is prior-independent (\(\leq\) 0.6 SE) and R077/R078 later rewrote the mechanism label as \(V_{3p}\)'s near-total cancellation); RT-69 and RT-70 were then launched, and RT-69 was triaged as 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). RT-70 through RT-73 and RT-75 were then triaged too: R074 deprecates the reading that at 0.5\(\times\) background the calibration closes to 0.99–1.00 (one seed, and an over-correction), R075 shows the L0 width and coverage depend on the scoring, R076 rules that z_off is the L0 scoring (R065's narrowing restored as the L0 width law), and R077/R078 rewrite the mechanism as “a \(\approx\) 1 % Fisher-penalty mismatch amplified by \(V_{3p}\)'s near-total cancellation”, with the mismatch rising with the PSF's central concentration. RT-74 was then triaged as R079: the L0 low-DET_ML centre term depends on the fitted offset \(r\) (at fixed \(\hat{S}\), +0.71 \(\to\) +0.23 and +0.10 \(\to\) −1.34), exceeding two validated pure-construction controls by +0.23…+0.50 z (DET_ML \(\lesssim\) 8); R076's “near-cancellation” note is withdrawn and the attribution is open. RT-76 was then triaged as R080: the candidate second-order-penalty fix is not achieved (the Bartlett term has the wrong sign and the beyond-Fisher term over-corrects 1.3–7.4\(\times\)), and the mismatch tracks the PSF wing structure and the counts (\(\hat{S}\), b). Five width red teams were then triaged too (R081–R085): the count dependence revised (the +0.86 pp at \(\hat{S}\) about 5 was a mask artefact), the L0 centre excess a property of the L0 pipeline itself (the 72/28 attribution later withdrawn by R087), the penalty-mismatch mechanism first recorded as not explaining the L1 width (R083) and then reversed by R086 (pooled estimators give an L1 mismatch of +1.19 %, one mechanism carrying both pipelines' widths), and the high-\(\hat{S}\) offset a control-truncation artefact. The PI redirected the main line back to emldetect the same day and set the finish line (the single-image conditional law); task 0b and RT-84 were dispatched, with 0b completed in the evening (R089) and RT-84/RT-85/RT-86 triaged (R091, R088/R090); the next main-line step T-LAW builds the conditional law on L1 output. Stage 33 (the fresh restart of stage 32) folded R064–R071 and deployed as 25620f8d, stage 34 folded R072–R073 and deployed as a94494ac, and stages 35+36 folded R074–R080 and deployed as 8b796a9d.

死胡同清单（为什么失败）Dead Ends (and Why They Failed)

教训总结Lessons

1. 约定错误比物理错误常见得多。Convention errors are far more common than physics errors.ln vs log10、ΔC/2、ecut 默认值、BG 单位——每一个都造成过 2 到 50 倍的假信号。任何新数字先过一遍约定清单（CLAUDE.md）。ln vs log10, ΔC/2, ecut defaults, BG units — each caused a 2 to 50× spurious signal at some point. Run every new number through the convention checklist (CLAUDE.md) first.
2. 条件化决定一切。Conditioning decides everything.条件于 S_true 的诊断看起来很好，条件于 Ŝ 才是理论。按 ML 切样本永远产生选择偏差。Diagnostics conditioned on S_true look great; theory conditions on Ŝ. Cutting samples by ML always produces selection bias.
3. 零参数是硬约束也是荣誉。Zero parameters is both a hard constraint and a badge of honour.每往公式里加一个拟合数，就要在基准上付出"参数个数"的代价；而一个零参数公式如果赢了，说明物理真的被抓到了。Every fitted number added to a formula costs "parameter count" on the benchmark; if a zero-parameter formula wins, the physics has genuinely been captured.
4. 负结果要登记。Negative results must be recorded.被支配的条目、失败的迁移、被否定的假设都写进实验笔记。它们防止下一个人重走死路。Dominated entries, failed transfers, refuted hypotheses all go into the experiment notes. They stop the next person from repeating the dead end.
5. 先冻结协议再看数据。Freeze the protocol before looking at the data.判定规则、网格、权重都在跑之前写下来并记哈希（E162 阶段 16–18 的做法），否则就是拟合。Adjudication rules, grids, weights are all written down and hashed before running (the practice of E162 stages 16–18); otherwise it is fitting.
6. 端到端闭合检验。End-to-end closure tests.中间链条每环都对不等于整体对（背景误差两起两落）；最后一定要做一次"从真值到 catalog"的完整对照。Every link being right does not mean the whole chain is right (the background-error case rose and fell twice); a full truth-to-catalogue comparison must be done at the end.