现状：2026-10-19 时的项目Current State: The Project as of 2026-10-19

m 的现状Where m Stands

来自 reviews/T11_026_claude.md 的总结表（评审口径）。Summary table from reviews/T11_026_claude.md (review scope).

来源：figures/v2/T11/t11_s25_m_date.png（T11_025，MJD_START 按 FITS TCOMM 引用；Δχ² 判读已存入 analyze_t11_s26_*.py）。Source: figures/v2/T11/t11_s25_m_date.png (T11_025; MJD_START quoted from the FITS TCOMM; the Δχ² code is saved in analyze_t11_s26_*.py).

已确立的结论Established Conclusions

1. 单图像理论在理想拟合器上闭合。Single-image theory closes on the ideal fitter.L0 前缀（约 300–500 窗口行/列表）上，E140、E141 和 V1–V4 在每个 α 列表和天空先验下 A1 都在噪声底以内；不含 δ 项的 V0 的 A1 是 0.50。（来源：BENCHMARK_V2_TASKS.md 2026-10-11 状态段，T5_006。）On the L0 prefix (~300–500 window rows per list), E140, E141 and V1–V4 have A1 inside the noise floor for every α list and sky prior; V0 without δ has A1 = 0.50. (Source: BENCHMARK_V2_TASKS.md 2026-10-11 state section, T5_006.)
2. 联合拟合损失是二维的。The joint-fit loss is two-dimensional.实测局部联合损失与精确 2-D 式 tr(Φᵢ(Φᵢ⁻¹ − Φ_tot⁻¹)) ≈ 2(1−w) 之比跨 Ŝ 分箱为 1.01–1.25（07 R013 放宽；早期读数 1.04–1.06 只是中心值，"与 S 无关"的措辞已撤回）。（来源：reviews/T11_013_claude.md，PASS；07 R013。）The ratio of the measured local joint-fit loss to the exact 2-D formula tr(Φᵢ(Φᵢ⁻¹ − Φ_tot⁻¹)) ≈ 2(1−w) is 1.01–1.25 across Ŝ bins (widened by 07 R013; the early reading 1.04–1.06 is only the central value and its "independent of S" wording is withdrawn). (Sources: reviews/T11_013_claude.md, PASS; 07 R013.)
3. 2-D 位置损失本身 Provisional（量级对）；"它单独闭合目录" Questioned，Q1 现为 open（闭合未证）。The 2-D position loss itself is Provisional (right order); "it closes the catalogue alone" is Questioned, and Q1 is now open (closure not shown).2026-10-13 的读数曾把所有 MOS 波段的窗口比例从 −2.5 到 −4.9% 变为 +0.6 到 +3.1%（M2 b4 逐箱 A1 从 0.324 到 0.059），当时被写成"零参数地闭合了 catalog 均值亏缺"。该闭合读法已于 07 R019 撤回：按模拟自己的口径，目录亏缺是 −9.282 % [−9.350, −9.219]（11,401 个观测）而不是 −10.4 %（后者不可复现，来源 = 13c C-iso60 表，行集未重新导出）；15 幅图像联合模拟试点用估计背景只给到 A 臂 −5.9 %、B 臂 −7.4 %（差 3.39 % / 1.85 %），且试点与目录不匹配——b 0.028 恒定对 0.02→0.14，两个 6–10 窗口落在不同 Ŝ（7.3–8.5 对 9.5–13），匹配 Ŝ 后差 4–6 %，试点亏损到 Ŝ > 30 消失而目录停在 −7 %，试点 L2d/μ 5.0 % 对目录 9.8 %。T11_021 的"大小巧合"假说仍未证实（T11_022：冻结 m_cam 只闭合 12 个配置中的 4 个）。（来源：07 R015、07 R019；reviews/T11_013_claude.md、reviews/T11_021_claude.md、reviews/RT-26_triage_claude.md。）The 2026-10-13 reading moved all MOS bands' window fractions from −2.5 to −4.9% to +0.6 to +3.1% (M2 b4 per-bin A1 from 0.324 to 0.059) and was then written up as "the two-dimensional loss closes the MOS catalogue mean deficit zero-parameter". That closure reading is withdrawn by 07 R019: under the simulation's own convention the catalogue deficit is −9.282 % [−9.350, −9.219] (11,401 observations), not −10.4 % (the latter is not reproducible; origin = the 13c C-iso60 table, row set not re-derived); and the 15-image joint-simulation pilot, scored with the estimated background, reaches only arm A −5.9 % and arm B −7.4 % (gap 3.39 % / 1.85 %), and is not matched to the catalogue — b 0.028 flat against 0.02→0.14, the two 6–10 windows at different Ŝ (7.3–8.5 against 9.5–13), 4–6 % at matched Ŝ, the pilot's deficit gone by Ŝ > 30 while the catalogue stays at −7 %, and L2d/μ 5.0 % against 9.8 %. The T11_021 "coincidence of size" hypothesis remains unconfirmed (T11_022: the frozen m_cam closes only 4 of 12 configs). (Sources: 07 R015, 07 R019; reviews/T11_013_claude.md, reviews/T11_021_claude.md, reviews/RT-26_triage_claude.md.)
4. 两个零参数条目已正式登记Two zero-parameter entries formally registered（T3_joint_2D、T3_E147mix_2D，C-all 与 C-iso 各一份），前沿重算并部署到网站。（来源：04_DECISION_LOG.md 2026-10-13 03:30 条目。）(T3_joint_2D, T3_E147mix_2D, one on C-all and one on C-iso); fronts recomputed and deployed. (Source: 04_DECISION_LOG.md 2026-10-13 03:30 entry.)
5. η₀ 不是渐晕约定。η₀ is not a vignetting convention.RATE × EXP 已等于图像拟合计数；用 S × VIG 重评会过度修正。（来源：E158。）RATE × EXP already equals the image fit counts; re-evaluating with S × VIG over-corrects. (Source: E158.)
6. 最内环亏缺是环境效应。The innermost-ring deficit is an environment effect.有延展邻居时 −1.0 到 −2.0σ，没有时 −0.2 到 −0.8σ；这就是 C-iso 层存在的原因。（来源：E162 第 2 阶段。）With an extended neighbour −1.0 to −2.0σ, without −0.2 to −0.8σ; that is why the C-iso layer exists. (Source: E162 stage 2.)
7. PN 的宽度超出与背景估计误差一致。PN's width excess is consistent with background-estimation error.样本外 R² 0.73–0.88，σ_ε ≈ 0.035–0.08。（来源：T11_004。）Out-of-sample R² 0.73–0.88, σ_ε ≈ 0.035–0.08. (Source: T11_004.)
8. 背景误差不是 MOS 均值亏缺的解释。Background error is not the explanation for the MOS mean deficit.隔离匹配的模拟只对上了 6–10 窗口的水平（−0.039 vs catalog −0.046），r 斜率符号相反、更亮窗口不匹配。（来源：T4_007，NOT MATCHED。）Isolated matching simulation only reaches the 6–10 window level (−0.039 vs catalogue −0.046); r slope has opposite sign, brighter windows don't match. (Source: T4_007, NOT MATCHED.)
9. 高计数乘性亏损属于真实数据。The high-count multiplicative deficit belongs to the real data.L0 对照 m_L0 = 0.0038（不截断主读数，RT-13 分诊更正：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]，区间不含 0（T5_009 评审接受；RT-13/RT-03：结论不变——L0 的底远小于星表的 m ≈ 0.06）；C-all 上 m 不随离轴环、不随背景变化，共同行上 C-all 与 C-iso 相同（+0.0002 [−0.0004, +0.0008]，更正 T5_008 评审第 4 条）。（来源：reviews/T5_009_claude.md。）L0 control m_L0 = 0.0038 (the uncut primary reading; the RT-13 triage corrected the record: 0.0077 is the ML < 690 cut value, a sensitivity only), catalogue m minus L0 m = 0.058 [0.056, 0.061], interval excludes 0 (T5_009 review accepted; RT-13/RT-03: the verdict is unchanged — L0's floor is far below the catalogue's m ≈ 0.06); on C-all m has no ring and no background dependence, and on common rows C-all and C-iso m are identical (+0.0002 [−0.0004, +0.0008], correcting T5_008 review item 4). (Source: reviews/T5_009_claude.md.)
10. 核函数高 S 失效已修复。The kernel high-S failure is fixed.全部 15 个 E114 网格已扩展到 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的默认核，前沿不变。剩约 0.6% 的乘性底线（L0 m = 0.004–0.008），是"闭合"类结论的精度下限。T11 的星表总体从未超过 S = 300（加载器为留在旧网格内而截断），所以 T11 的 m_cam 结论不受影响（T5_011 评审更正了 T5_010 评审的担忧）。（来源：reviews/T5_010_claude.md、reviews/T5_011_claude.md。）All 15 E114 grids 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), the precision floor for "closed" claims. T11's catalogue populations never exceed S = 300 (the loader caps S to stay inside the old grid), so T11's m_cam conclusions are unaffected (the T5_011 review corrects the T5_010 review's concern). (Sources: reviews/T5_010_claude.md, reviews/T5_011_claude.md.)
11. PN 的隐含损失 ℓ_impl ≈ 0.8·(1−w)/μ。PN's implied loss ℓ_impl ≈ 0.8·(1−w)/μ.措辞更正：接近而非吻合（比值 0.78–0.87，区间不含 (1−w)/μ）；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。（来源：reviews/T11_021_claude.md。）Wording correction: closer to, not matches (ratio 0.78–0.87, intervals exclude (1−w)/μ); 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. (Source: reviews/T11_021_claude.md.)
12. "每相机一个常数 m" 假说被否定。The "one constant m per camera" hypothesis is refuted.高计数冻结的 m_cam 预测低窗口只闭合 12 个配置中的 4 个（未闭合）；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。（来源：reviews/T11_022_claude.md。）A high-count-frozen m_cam closes only 4 of 12 configs in the low windows (not closed); b3/b4 low-window residuals are positive (deficit grows with S); m grows with band (b4−b1 ≈ +0.034), PN b1 negative. (Source: reviews/T11_022_claude.md.)
13. MOS 的隐含损失：一份 (1−w) 加真实数据的 m。MOS's implied loss: one (1−w) plus the real-data m.M2 b4 上 ℓ_impl = 0.1124 [0.1111, 0.1136]。比值：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"，不是 2(1−w)——T11_021 的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。（来源：reviews/T11_023_claude.md。）On M2 b4, ℓ_impl = 0.1124 [0.1111, 0.1136]. Ratios: 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", not 2(1−w) — the T11_021 hypothesis is supported on the camera where the two-dimensional form looked right. The fitter-level term r_MC = +0.017 is recorded. (Source: reviews/T11_023_claude.md.)
14. m 的拆分：随年代下降；MASKFRAC 是代理。The splits of m: falls with epoch; MASKFRAC is a proxy.全部 12 个配置的 m 都随观测年代下降（这是更普遍的信号）。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（4XMM 中是探测掩膜的 PSF 加权平均；M2 b4 中位数 0.981，IQR 0.961–0.986），"高 MASKFRAC 有亏损"意味着完全覆盖的源有亏损、部分遮挡的源 m ≈ 0；两组只差约 0.02，不足以直接造成 0.05 的 m 差异——它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。（来源：reviews/T11_023_claude.md；RT-11 分诊：M2 b4 上"部分遮挡的源 m ≈ 0"撤回——低 MASKFRAC 一半 m = 0.064（高的一半 0.068，差 −0.004 [−0.008, +0.0005]）；分组表把修正前的数值标成了"mask-aware"（脚本 bug），修正后为 0.024（低）/ 0.051（高）；掩膜臂是代数推出的反事实，不是因果证据。4XMM 的 MASKFRAC 定义与"加入 MASKFRAC 后 M1 阶跃仍在"两条成立。）m falls with epoch on all 12 configs (the more general signal). 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 (in 4XMM it is the PSF-weighted mean of the detection mask; on M2 b4 the median is 0.981, IQR 0.961–0.986), so "high MASKFRAC carries the deficit" means fully covered sources have the deficit and partly masked sources have m ≈ 0; the two groups differ by only ~0.02, too small to produce a 0.05 difference in m directly — it is probably a proxy for epoch or detector position. EP_EXTENT and SUM_FLAG could not be tested (the non-zero groups are too small). (Source: reviews/T11_023_claude.md; RT-11 triage: "partly masked sources have m ≈ 0" is withdrawn on M2 b4 — the low-MASKFRAC half has m = 0.064 [high half 0.068, difference −0.004 [−0.008, +0.0005]]; the split table's "mask-aware" labels were pre-arm values, a script bug — corrected halves 0.024 (low) / 0.051 (high); the mask-aware arm is an algebraic counterfactual, not causal evidence. The 4XMM MASKFRAC definition and the M1 step's survival with MASKFRAC stand.)
15. m 和 a 成为高计数报告列（用户决定）。m and a become high-count report columns (user decision).两者目标值都是 0，报告用，不进前沿；漂移量只作派生列。高计数报告页已上线：v2/highcount.html。（来源：04_DECISION_LOG.md 2026-10-13 20:45 条目。）Both with target 0, report-only, not entering the front; the drift is a derived column. The high-count report page is live: v2/highcount.html. (Source: 04_DECISION_LOG.md 2026-10-13 20:45 entry.)
16. ML→ΔC 溢出已修复。The ML→ΔC overflow is fixed.ΔC = 2·gammainccinv(1.5, exp(−ML)) 在 ML ≳ 745 处 exp(−ML) 下溢为 inf；对 ν = 3 改用精确恒等式 ln Q(3/2,x) = −x + ln((2/√π)√x + erfcx(√x)) 反解（brentq，往返精度 4.4e-13，ML < 600 处与旧代码一致到 2e-13）；lib/detml.py:75,104-105 有同样模式，已报告未改动。评审更正：17,958 → 467 行的流失来自 EP_EXTENT==0（−12,514）、SUM_FLAG==0（−4,347）和公开切分，不是溢出删行；溢出让 467 行中 402 行的 ΔC 失真。ML 截断曾让参照 m 偏高约 0.002（0.0660 → 0.0643）。（来源：reviews/T5_012_claude.md、reviews/T5_013_claude.md。）ΔC = 2·gammainccinv(1.5, exp(−ML)) underflows to inf at ML ≳ 745; for ν = 3 replaced by inverting the exact identity ln Q(3/2,x) = −x + ln((2/√π)√x + erfcx(√x)) (brentq, round-trip accuracy 4.4e-13, agrees with the old code below ML 600 to 2e-13); lib/detml.py:75,104-105 has the same pattern, reported without editing. Review correction: the 17,958 → 467 attrition comes from EP_EXTENT==0 (−12,514), SUM_FLAG==0 (−4,347) and the public split, not rows dropped by the overflow; the overflow distorted ΔC for 402 of the 467 rows. The ML clip had biased the reference m up by ~0.002 (0.0660 → 0.0643). (Sources: reviews/T5_012_claude.md, reviews/T5_013_claude.md.)
17. 乘性亏损 m 延续到亮源。The multiplicative deficit m extends to bright sources.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）（T5_013）。全部 6 个零参数 T3 条目同样延续：|Δm| < 0.01（C-all 上 −0.006 到 +0.0002，C-iso 上 +0.001 到 +0.004）；在 >300 背景窗口它们的配对 A1 差全部为负且区间不含 0（T5_014）。约 6% 的乘性亏损从 S ≈ 100 延续到 S ≈ 4000——这是 m 是真实乘性项的最强证据。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) (T5_013). All 6 zero-parameter T3 entries extend too: |Δm| < 0.01 (−0.006 to +0.0002 on C-all, +0.001 to +0.004 on C-iso); in the >300 context window their paired A1 differences are all negative with intervals excluding 0 (T5_014). The ~6% multiplicative deficit holds from S ≈ 100 to S ≈ 4000 — the strongest evidence yet that m is a true multiplicative term.
18. V5c_eta 在亮源上比不修正更差；区间根因已修复。V5c_eta is worse than no correction on bright sources; the interval root cause is fixed.（区间修正后，T5_015 评审口径。）在 >300：其 |frac| 差在 C-all 和 C-iso 两个层上、所有区间方法下都不含 0（C-all 上 [+0.033, +0.034]）；A1 差在 C-iso 上所有方法都不含 0（BCa 主区间 [+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，分布偏斜）。"为了修正而修正"的实例。区间缺陷的根因：bootstrap 重抽样的是原始 A1，而点估计是噪声校正后的 A1_nc——两个不同的统计量；修复后 bootstrap 直接计算所报告的统计量，440/440 主区间包含点估计，70/70 个标记判定无一改变（T5_015）。现行标记只覆盖 ≤300 的窗口；是否扩展到 >300 待用户决定（若扩展，只有 V5c_eta 被标记，前沿不受影响）。高计数报告页已发布：../v2/highcount.html。（来源：reviews/T5_014_claude.md、reviews/T5_015_claude.md。）(After the interval fix, per the T5_015 review.) In >300: its |frac| difference excludes 0 on both C-all and C-iso under every interval method ([+0.033, +0.034] on C-all); its A1 difference excludes 0 on C-iso under every method (BCa headline [+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, skewed distribution). The "correction for its own sake" case. The interval defect's root cause: the bootstrap resampled raw A1 while the point estimate was the noise-corrected A1_nc — two different statistics; after the fix the bootstrap evaluates the reported statistic, 440/440 headline intervals contain their point, no flag verdict changed (70/70, T5_015). The current flag covers windows ≤ 300; extending it to >300 is a pending user decision (if extended, only V5c_eta is flagged, the front is unaffected). The high-count report page is published: ../v2/highcount.html. (Sources: reviews/T5_014_claude.md, reviews/T5_015_claude.md.)
19. T13 第 1 阶段：产品路线可行；发现管线版本陷阱。T13 stage 1: the product route works; a pipeline-version trap found.NXSA 的 PPS 产品（4 个产品 + OBLSLI）、EXPID 匹配、共享 WCS 全部验证（X_IMA = WCS 像素 + 1，误差 < 0.002 px）；冻结拟合器原样可用。3 个观测上 S_local/S_cat 为 1.05–1.11；孤立场的 ΔC 排序相关 ρ 为 0.81–0.98（拥挤的 2001 场约 0.35，拟合移到邻近源）。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 2024 年重处理产品（PPSVERS 21.51，SAS 21），DR14 用管线 18/19——0512180101 的图像计数约为星表值的 3.5 倍。第 2 阶段用 m_cat / m_newcat / m_local 三个测量量分开判读，不必用 SAS 18 重新生成。（来源：reviews/T13_001_claude.md。）NXSA's PPS products (4 products plus OBLSLI), EXPID matching and the shared WCS all verified (X_IMA = WCS pixel + 1, to < 0.002 px); the frozen fitter unchanged. On 3 observations S_local/S_cat is 1.05–1.11; ρ(ΔC) is 0.81–0.98 on isolated fields (the crowded 2001 field ~0.35, fits move to a neighbour). The M1–M2 flux-ratio paired difference is +0.0006 [−0.007, +0.015], interval contains 0: no significant difference. 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. Stage 2 uses the three measurements m_cat / m_newcat / m_local to separate the readings, without regenerating with SAS 18. (Source: reviews/T13_001_claude.md.)
20. M1 阶跃经三个独立复分析复核。The M1 step rechecked by three independent re-analyses.三个独立复分析都复现 M1 在 12.9 年处的阶跃：Δχ² = +38.5，阶跃 −0.0157 ± 0.0023；M2/PN 的漂移判读也成立。两点限定（决策日志 2026-10-15 02:55）：按年份做 bootstrap 时，只有 76% 的重抽样里 M1 阶跃超过 Δχ² > 9 的门槛；M1 阶跃之后的漂移不显著（区间 [−0.0014, +0.0003]）。所以"M1 有阶跃"可信，但不是压倒性的。（来源：04_DECISION_LOG.md 2026-10-15 02:55 条目；RT-04 分诊：阶跃成立——日期预登记，24 个扫描日期下 Bonferroni p = 7×10⁻¹¹，平滑漂移移不掉，逐年剔除都为负；措辞收窄为"在 CCD3 损坏的日期有阶跃"，不写"由 CCD3 损坏造成"。）Three independent re-analyses reproduce M1's step at year 12.9: Δχ² = +38.5, step −0.0157 ± 0.0023; the M2/PN drift reading holds too. Two caveats (decision log 2026-10-15 02:55): in year-bootstrap resamples, the M1 step passes the Δχ² > 9 threshold in only 76% of them; M1's post-step drift is not significant (interval [−0.0014, +0.0003]). So "M1 has a step" is credible, but not overwhelming. (Source: 04_DECISION_LOG.md 2026-10-15 02:55 entry; RT-04 triage: the step stands — dates pre-registered, Bonferroni p = 7×10⁻¹¹ over 24 scanned dates, smooth drifts do not remove it, leave-one-year-out all negative; wording narrowed to "a step at the CCD3-loss date", not "caused by".)
21. 排行榜的噪声修正指标经红队检验成立（RT-07）。The leaderboard's noise-corrected metric stands under red-team attack (RT-07).公开 bin 表复现排行榜 A1/A2 到 1e-16；16 个条目的原始与修正后排序完全相同；没有条目级的轴被截到 0；反事实检验（宽度乘 0.8）排名明显变差（A2_nc 0.15 → 0.42）；前 5 个 bin 只占 A1 平方误差的 ≤14%。限定（接受为措辞）：A1 最佳条目只在预先规定的按计数加权下成立（等权时 C-all 的 T3_joint_2D 与 T3_E147mix_2D 互换，三轴前沿不变）；配对区间是原始分数的区间。（来源：reviews/RT-07_triage_claude.md。）The public bin tables reproduce the leaderboard A1/A2 to 1e-16; raw and corrected orders are identical for all 16 entries; no entry-level axis is clipped to 0; a counterfactual (width × 0.8) ranks clearly worse (A2_nc 0.15 → 0.42); the first 5 bins carry ≤14% of the A1 squared error. Qualifiers (accepted as wording): the A1 best entry holds under the pre-registered count weighting only (under equal weights T3_joint_2D and T3_E147mix_2D swap on C-all, the three-axis front is unchanged); the paired intervals are on the raw scores. (Source: reviews/RT-07_triage_claude.md.)
22. T13 第 2 阶段：已评审，但位置文件 bug 使逐行量作废；修正后的临时数字"位置差约 25–38%"。T13 stage 2: reviewed, but a position-file bug voids the row-level numbers; corrected provisional numbers put the position gap at about 25–38%.185 个观测（种子在下载前固定；星表对星表筛选：RATE_2024/RATE_DR14 中位数 1.001，0/185 超出 ±10%）。RT-05 红队发现的 bug：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，约 90% 的行用了别的源的位置——S_local、|r_uv| 筛选、m_local 的 M1 阶跃与漂移、以及 RT-01 的固定位置对照全部作废；第 2 阶段在 T13 第 3 阶段完成后修复重做。修正标签后的临时数字（RT-05 分诊）：m_local ≈ 0.061 / 0.038（M2/M1）；m_cat − m_local = 0.020 [−0.006, +0.041]（M2，占 m_cat 的 25%，未分辨）、0.023 [+0.007, +0.037]（M1，占 38%）——不是"约一半"；RT-01 的"这一半是共享位置的代价"及逐行证据撤回。不受影响的：m_cat ≈ m_newcat（M2 0.086/0.083，M1 0.067/0.065）——亏损不绑定管线版本；流量归一化 (a) 排除为主因（RATE 比 +0.0003/年；S_local/S_cat 1.01–1.03，经核斜率 1.4 最多约 0.035）。位置代价按 ΔC^0.45–0.52 增长，不是乘性 m；偏移是每个观测恒定、两台相机共有的约 0.5 像素矢量。新假设（待检验）：4XMM 的 RA/DEC 带 catcorr 天体测量校正而 PPS 图像 WCS 没有，用未校正 WCS 投影校正坐标正好产生这种偏移。适用范围：DR14–2024 匹配源。（来源：reviews/T13_002_claude.md、reviews/RT-05_triage_claude.md、04_DECISION_LOG.md 2026-10-15 07:30 与 18:06 条目。）185 observations (seed fixed before download; catalogue-to-catalogue screen: RATE_2024/RATE_DR14 median 1.001, 0/185 outside ±10%). The bug the RT-05 red team found: t13_s2_make_pos.sh computed pixel positions in raw row order but paired them with rows sorted by SRC_NUM, so about 90% of rows used another source's position — S_local, the |r_uv| cut, the M1 step and drift in m_local, and RT-01's fixed-position control are all void; stage 2 will be fixed and redone after T13 stage 3. Corrected provisional numbers (RT-05 triage): m_local ≈ 0.061 / 0.038 (M2/M1); m_cat − m_local = 0.020 [−0.006, +0.041] (M2, 25% of m_cat, unresolved) and 0.023 [+0.007, +0.037] (M1, 38%) — not "about half"; RT-01's "that half is the shared-position cost" and its row-level evidence are withdrawn. Unaffected: m_cat ≈ m_newcat (M2 0.086/0.083, M1 0.067/0.065) — the deficit is not tied to a pipeline version; flux normalisation (a) is ruled out as the main cause (RATE ratio +0.0003/yr; S_local/S_cat 1.01–1.03, at most ~0.035 through the kernel slope of 1.4). The position cost grows as ΔC^0.45–0.52, not as a multiplicative m; the offset is a per-observation, camera-common vector of about 0.5 px. New hypothesis (to be tested): 4XMM's RA/DEC carry the catcorr astrometric correction while the PPS image WCS does not, and projecting corrected coordinates through the uncorrected WCS produces exactly this offset. Scope: matched DR14–2024 sources. (Sources: reviews/T13_002_claude.md, reviews/RT-05_triage_claude.md, 04_DECISION_LOG.md 2026-10-15 07:30 and 18:06 entries.)

正在进行（2026-10-24）Running (2026-10-24)

来源：cognition-os/08_RUNNING.md rev 42（2026-10-17T23:51Z）§3、04_DECISION_LOG.md 2026-10-13 至 2026-10-17 20:59 的补记、07_REVISION_LOG.md R001–R046、reviews/RT-43_triage_claude.md、reviews/RT-44_triage_claude.md、reviews/RT-45_triage_claude.md、reviews/T13_006_claude.md、reviews/T13_008_claude.md、reviews/RT-22_triage_claude.md、reviews/RT-23_triage_claude.md、reviews/RT-24_triage_claude.md、reviews/RT-25_triage_claude.md、reviews/RT-26_triage_claude.md、reviews/RT-27_triage_claude.md、reviews/RT-32/33/34/36/37/38_triage_claude.md、.agent/jobs/2026-10-16-detml-m31-ml-convention/REVIEW_claude.md、.agent/jobs/2026-10-17-detml-width-theory/BRIEF.md、.agent/jobs/2026-10-17-detml-width-ladder-L0L1/BRIEF.md、.agent/jobs/2026-10-17-detml-sas22-vs-sas20/BRIEF.md、cognition-os/sessions/2026-10-17.md。Sources: cognition-os/08_RUNNING.md rev 42 (2026-10-17T23:51Z) §3, the 04_DECISION_LOG.md addenda from 2026-10-13 through 2026-10-17 20:59, 07_REVISION_LOG.md R001–R046, reviews/RT-43_triage_claude.md, reviews/RT-44_triage_claude.md, reviews/RT-45_triage_claude.md, reviews/T13_006_claude.md, reviews/T13_008_claude.md, reviews/RT-22_triage_claude.md, reviews/RT-23_triage_claude.md, reviews/RT-24_triage_claude.md, reviews/RT-25_triage_claude.md, reviews/RT-26_triage_claude.md, reviews/RT-27_triage_claude.md, reviews/RT-32/33/34/36/37/38_triage_claude.md, .agent/jobs/2026-10-16-detml-m31-ml-convention/REVIEW_claude.md, .agent/jobs/2026-10-17-detml-width-theory/BRIEF.md, .agent/jobs/2026-10-17-detml-width-ladder-L0L1/BRIEF.md, .agent/jobs/2026-10-17-detml-sas22-vs-sas20/BRIEF.md, cognition-os/sessions/2026-10-17.md.

开放问题（评审口径）Open Problems (review scope)

1. 进度指标已定死，位置项的第一个候选已过地板，但还没有在新位置上检验（主线，2026-10-22）。The progress metric is fixed, the position term's first candidate clears the floor, but it has not yet been tested on fresh positions (the main line, 2026-10-22).判据的演化：“每条条目都在 \(\varepsilon\) 内”被证明噪声受限（R128/R129/R133：完美规律在能段 4 的天花板只有 0.21–0.27），改为每个位置一个合并偏移（R131）；R131 的“规律就压在位置盲地板上”被 R138 推翻（原地板只注入了 0.634 倍的声明 \(\tau^2\)；改正后 \(\varepsilon\) 0.10 处 D 与 H2 的 0.545 高于地板 0.469）。PI 定下容差（04 2026-10-22）：比较规律用 \(\varepsilon\) 0.05、跟踪进度用 \(\varepsilon\) 0.10（\(\varepsilon_w\) 0.10），\(\varepsilon\) 0.20 撤掉（饱和）；\(\varepsilon\) 0.10 \(\equiv\) 0.12 DET_ML（6–10 窗口）。打分手冻结为 score_pooled_v1.py（R147：18 格、位置真值校准的 95 % CI）。位置效应：不是源表（R126：每个位置注入同一份 54 源栅格），是像素级几何且两条能段共享（R134：r = +0.936；R145：去掉拟合输出协变量后 3 因子仍迁移）；导出的相位项被驳回（R140）；R144 的“像素中心臂把偏移放大 2 倍”被 R150 撤回——打分器伪迹（同一格掩膜下比值 1.16，相位在两臂里 11/12 个位置根本没变）。第一个拟合前位置项（拟合盒背景水平 bkg_boxmean_rel + phase_r，\(\leq\) 4 参数）把 D 在 \(\varepsilon\) 0.10 处从 0.545 抬到 0.727（R139；R148：诚实全管线零假设下 p 0.003–0.004，不是相位主导，同一分半被复用、没有新位置）。下一步：POS-FRESH（\(\geq\) 24 个新位置、冻结系数）决定这个项是不是真的；PHASE-DIAG 解释 1.16 的残余膨胀。来源：07 R128–R152、04 2026-10-21/22、reviews/RT-105 至 RT-111。How the criterion evolved: the “every entry within \(\varepsilon\)” rule was shown to be noise-limited (R128/R129/R133: a perfect law's ceiling on band 4 is only 0.21–0.27), so it became one pooled offset per position (R131); R131's “the laws sit at the position-blind floor” is overturned by R138 (the old floor injected only 0.634\(\times\) of the declared \(\tau^2\); after the correction D and H2 at \(\varepsilon\) 0.10, at 0.545, sit above the floor of 0.469). The PI fixed the tolerances (04, 2026-10-22): \(\varepsilon\) 0.05 to compare laws, \(\varepsilon\) 0.10 as the progress tracker (\(\varepsilon_w\) 0.10), \(\varepsilon\) 0.20 dropped as saturated; \(\varepsilon\) 0.10 \(\equiv\) 0.12 DET_ML (the 6–10 window). The scorer is frozen as score_pooled_v1.py (R147: 18 cells, position-true-calibrated 95 % CI). The position effect: not the source list (R126: every position injects the same 54-source grid), but pixel-level geometry shared by both bands (R134: r = +0.936; R145: the 3-factor model still transports without fit-output covariates); the derived phase term is rejected (R140); R144's “the pixel-centre arm amplifies the offsets 2\(\times\)" is withdrawn by R150 — a scorer artefact (ratio 1.16 under the same cell mask, and the phase did not change at all in 11 of 12 positions between the arms). A first pre-fit position term (the fit-box background level bkg_boxmean_rel + phase_r, \(\leq\) 4 parameters) lifts D at \(\varepsilon\) 0.10 from 0.545 to 0.727 (R139; R148: under the honest full-pipeline null p 0.003–0.004, not phase-led, the same split reused, no fresh positions). Next: POS-FRESH (\(\geq\) 24 fresh positions, frozen coefficients) decides whether the term is real; PHASE-DIAG explains the residual 1.16 inflation. Sources: 07 R128–R152, 04 2026-10-21/22, reviews/RT-105 through RT-111.
2. 单图 emldetect 条件律 \(p(\mathrm{DET\_ML} \mid \hat{S}, b, \mathrm{index})\)：在哪个估计对象上、有多大比例的位置通过（主线，PI 2026-10-21 05:39 EDT 定的目标）。The single-image emldetect conditional law \(p(\mathrm{DET\_ML} \mid \hat{S}, b, \mathrm{index})\): under which estimand, and at what fraction of positions does it pass (the main line; the target the PI set on 2026-10-21 05:39 EDT)?四条候选（C、D、DQ、MIX3d）在旧打分手上全部被拒（R103），但旧打分手被专家复核认定混用三个估计对象（R114）：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）进一步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，所以"宽度可以推广"只能读作位置内。新打分手（SCORER-CONTRACT v2，R118）把两个目标分开：未见位置总体（POP-UNOBS）上能段 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）；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（POS-CAL）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——位置效应本身比 0.05 的均值容差大，所以逐位置目标不能靠合并格通过。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，用 72 参数诊断修正），不是位置泛化缺口；专家的先验混合恒等式作为代数成立但其协方差项已被 D 自己的 \(\beta\) 斜率抵消（RT-99）。宽度项的读数被限定（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。PI 目标（04，2026-10-21）：05:39 EDT 定的是不是每个位置，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；工作形式 = 域 \(\theta \leq \theta_{max}\) 且 MASKFRAC \(\geq m_{min}\)，一个位置通过 = 中心/宽度/覆盖率逐位置都通过。11:42 EDT 修订：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，规律不会因为没达到阈值被排除；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，90 % 默认值撤回。第一个点（R120，RT-100 已复算）：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——已解决位置比例约 12 %，这就是进度曲线的起点。R122 限定：一个 \(\leq\) 3 参数的位置项并不能提高已解决比例（能段 4 的 D 4/33 \(\to\) 3/33，配对差 \(-0.031\) [\(-0.091, 0\)]），最好的协变量是 PSF 形状（解释位置间偏移方差的 12–17 %），连神谕位置常数也只到 6/33；偏移在重复两半之间可复现（r = 0.928 / 0.982），但重复两半共用同一份注入源表（Jaccard 1.0），所以"位置的稳定性质"可能是那份固定 \(S_{true}\) 列表的性质——SOURCE-LIST 检验在跑。已分诊：RT-101 \(\to\) R119、RT-100 \(\to\) R120、WIDTH-2ND \(\to\) R121、POS-OFFSET \(\to\) R122、RT-102 \(\to\) R123（\(\delta_1\) 在 emldetect 上高估，只能引用 \(\delta_0\) 的均值结果）、RT-103 \(\to\) R124（POP-UNOBS 计数按分半区间引用）。在跑（不引用数字）：WIDTH-3（高 \(b\) 的条件方差）、SOURCE-LIST、scorer v3。^{[s14][s15][s16][s17]}All four candidates (C, D, DQ, MIX3d) are rejected under the old scorer (R103), but the old scorer is found by an expert review to mix three estimands (R114): 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) further shows 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 — so "the width generalises" can only be read as within position. The new scorer (SCORER-CONTRACT v2, R118) separates two targets: on the population of unseen positions (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); 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 (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), heavy-tailed with df \(\approx\) 3 — the position effect itself is larger than the 0.05 mean tolerance, so the per-position target cannot be met by passing pooled cells. 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 expert's prior-mixing identity holds as algebra but its covariance term is already cancelled by D's own \(\beta\) slope (RT-99). The width term's reading is qualified (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. The PI target (2026-10-21 05:39 EDT): 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; the working form is domain \(\theta \leq \theta_{max}\) and MASKFRAC \(\geq m_{min}\), a position passes if its centre / width / coverage all pass per position. Amended at 11:42 EDT (04): 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 a law is never excluded for missing a threshold; 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, and the 90 % default is withdrawn. The first point (R120, RT-100 replayed): 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 — a solved-position fraction of about 12 %, which is where the progress curve starts. R122 qualifies: a \(\leq\) 3-parameter position term does not raise the solved fraction (band-4 D 4/33 \(\to\) 3/33, paired difference \(-0.031\) [\(-0.091, 0\)]), the best covariate is the PSF shape (12–17 % of the between-position offset variance), and even an oracle position constant reaches only 6/33; the offsets reproduce between the rep-parity halves (r = 0.928 / 0.982), but those halves share the same injected source list (Jaccard 1.0), so "a stable property of the position" may be a property of that position's fixed \(S_{true}\) list — the SOURCE-LIST test is running. Triaged: RT-101 \(\to\) R119, RT-100 \(\to\) R120, WIDTH-2ND \(\to\) R121, POS-OFFSET \(\to\) R122, RT-102 \(\to\) R123 (\(\delta_1\) over-predicts on emldetect, so only the \(\delta_0\) mean result may be quoted), RT-103 \(\to\) R124 (quote the POP-UNOBS counts as split ranges). Running (no numbers quoted): WIDTH-3 (the high-\(b\) conditional variance), SOURCE-LIST, scorer v3.^{[s14][s15][s16][s17]}
3. 零参数插入式能不能描述我们自己的模拟里给定 \(\hat{S}\) 的 \(\Delta C\) 宽度（Q7，当前主攻）。Does the zero-parameter plug-in describe the \(\Delta C\) width given \(\hat{S}\) in our own simulations (Q7, the current main line)?四层都要测：玩具、L0（我们的拟合器）、L1（emldetect）、J（联合拟合）。答案分三层，全部按可观测量分箱、逐行条件化、每箱 \(\geq\) 10,000 行：（一）高计数锚——公式是对的。固定位置时 std z 0.98–1.00（DET_ML \(\gtrsim\) 20）；位置自由、z_off 打分下 0.989–1.003（DET_ML \(\geq\) 50，67 位置、两种先验）。那里零参数宽度不需要任何校准。^[s7]（二）目录区间——系统性偏窄，且没有点修正能闭合。真实 PSF 下合并 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，约 36–60 才回到 1；与先验无关。三条修正路线全部关闭：偏差校正振幅（std z 只到 0.975 / 0.959 / 0.970 / 0.979）、带 Fisher 展布的 \(E[V_{3p} \mid \hat{S}]\)（0.84–0.96）、解析二阶 penalty（Bartlett 符号反、beyond-Fisher 过修 1.3–7.4 倍且低计数出负方差）。实用路线是经验校准。^{[s8][s9][s10]}（三）机制——一个，两条流水线共享（R086）。\(V_{3p}\) = term1 − 4·penalty 是近乎全抵消（term1 是 Var(R) 的 15–31 倍），Fisher penalty 与有效 penalty 只差 0.1–1.4 %，这个约 1 % 的错配被抵消放大成 8–16 % 的方差高估（算例：DET_ML \(\approx\) 3.5 处 term1 = 63.6、Var(R) = 1.90、错配 1.4 % → \(V_{3p}(\hat{S})\)/Var(R) = 1.44）。错配随 PSF 的翼结构（低计数翼像素）与计数上升，所以不能用常数修。同一机制也解释 emldetect（L1）的宽度：用合并估计量，L1 错配在 DET_ML 6–10 是 +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；按行合并 0.940）；错配随计数下降（+0.63 % 于 10–20，100 以上约 0）。R083 的 −0.14 % 是逐格比值的平均，已被 R086 推翻。^{[s11][s12][s13]}仍在跑：任务 0b（同图同打分的配对实验，只让拟合器不同）与 RT-84（按计数修正的 \(V_{3p}\) 能否把 emldetect 宽度拉回 1，不对目标数据拟合）。一句白话：两个几乎相等的大数相减，其中一个差 1 %，结果就差 5–16 %——这就是残差看起来窄的原因，PSF 越集中越明显。All four layers must be tested: the toy, L0 (our fitter), L1 (emldetect) and J (the joint fit). The answer has three layers, all measured in observable bins, conditioned per row, with \(\geq\) 10,000 rows per bin: (i) The high-count anchor — the formula is right. With the position fixed, std z is 0.98–1.00 (DET_ML \(\gtrsim\) 20); with the position free under the z_off scoring it is 0.989–1.003 (DET_ML \(\geq\) 50, 67 positions, both priors). There the parameter-free width needs no calibration at all.^[s7] (ii) The catalogue interval — systematically narrow, and no point correction closes it. Under the real PSF the pooled std z is 0.840 / 0.894 / 0.909 / 0.959 / 0.984 at DET_ML \(\approx\) 4.6 / 7.9 / 11.1 / 22.8 / 36.8, returning to 1 only by about 36–60; it is prior-independent. Three correction routes are all closed: the bias-corrected amplitude (std z only reaches 0.975 / 0.959 / 0.970 / 0.979), \(E[V_{3p} \mid \hat{S}]\) with a Fisher spread (0.84–0.96), and analytic second-order penalties (Bartlett has the wrong sign; beyond-Fisher over-corrects 1.3–7.4× and gives a negative variance at the lowest counts). The practical route is empirical calibration.^{[s8][s9][s10]} (iii) The mechanism — one, shared by both pipelines (R086). \(V_{3p}\) = term1 − 4·penalty is a near-total cancellation (term1 is 15–31× Var(R)); the Fisher penalty matches the effective penalty to 0.1–1.4 %, and this ~1 % mismatch is amplified by the cancellation into an 8–16 % variance over-statement (a worked example: at DET_ML \(\approx\) 3.5, term1 = 63.6, Var(R) = 1.90 and a 1.4 % mismatch give \(V_{3p}(\hat{S})\)/Var(R) = 1.44). The mismatch rises with the PSF's wing structure (low-count wing pixels) and the counts, so it cannot be fixed by a constant. The same mechanism explains emldetect's (L1) width too: with pooled estimators the L1 mismatch at DET_ML 6–10 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; row-pooled 0.940); the mismatch falls with counts (+0.63 % at 10–20, about 0 above 100). R083's −0.14 % was a mean of per-cell ratios and is reversed by R086.^{[s11][s12][s13]} Still running: task 0b (the paired same-images same-scoring run, so that only the fitter differs) and RT-84 (whether a count-corrected \(V_{3p}\) pulls emldetect's width back to 1, not fitted to the target data). In plain words: two nearly equal large numbers are subtracted, one of them is off by 1 %, and the result is off by 5–16 % — that is why the residual looks narrow, and the more concentrated the PSF the worse it gets.
4. L1（emldetect）宽度缺口已被机制解释（R086 推翻 R083，已复算）：penalty 错配机制同样适用于 emldetect——用合并估计量，L1 的 Fisher penalty 错配在 DET_ML 6–10 是 +1.19 % [+1.03, +1.35]（R083 的 −0.14 % 是逐格比值的平均，是被 R086 推翻的读数；“BG_MAP_EML 是原因”也被驳回——它就是图像自己的背景），解释了箱内 L1 宽度缺口的 0.89 [0.84, 0.93]（预测 std z 0.919，实测 0.918；按行合并 0.940）。剩下的 0.11 与按行合并宽度 0.940 [0.922, 0.958] （比预测宽 0.021）归位置/b 混合。错配随计数下降：+0.63 %（10–20）、+0.64 %（20–50）、+0.42 %（50–100）、100 以上约 0。下一步（已派出 RT-84）：检验按计数修正的 \(V_{3p}\) 能否把 emldetect 宽度拉回 1（不对目标数据拟合）。一句白话：同一个相消放大机制同时决定 L0 与 emldetect 的宽度——这把主线（emldetect 条件律）的宽度侧从“失因开放”变成“机制已知、修正待验”。The L1 (emldetect) width deficit is explained by the mechanism (R086 reverses R083, replayed): the penalty-mismatch mechanism applies to emldetect as well — with pooled estimators the L1 Fisher-penalty mismatch at DET_ML 6–10 is +1.19 % [+1.03, +1.35] (R083's −0.14 % was a mean of per-cell ratios, a reading reversed by R086; the “BG_MAP_EML is the cause” reading is rejected too — it is the image's own background), explaining 0.89 [0.84, 0.93] of the within-cell L1 width deficit (predicted std z 0.919, observed 0.918; row-pooled 0.940). The remaining 0.11 and the row-pooled width 0.940 [0.922, 0.958] (0.021 wider than predicted) belong to the position/b mixing. The mismatch falls with counts: +0.63 % (10–20), +0.64 % (20–50), +0.42 % (50–100), about 0 above 100. Next step (RT-84 dispatched): test whether a count-corrected \(V_{3p}\) pulls the emldetect width back to 1 (not fitted to the target data). In plain words: one cancellation-amplification mechanism carries both the L0 and the emldetect widths — the width side of the main line (the emldetect conditional law) moves from “cause open” to “mechanism known, correction to be tested”.
5. 联合拟合损失的正确形式。Correct form of the joint-fit loss.二维形式 L_2d ≈ 2(1−w) 在 6–10 窗口对 MOS 最好，但不是闭合项：窗口不平坦（6–10 残差 +0.004..+0.034，>20 为 −0.009..−0.054）；PN 上 (1−w) 反而闭合（A1 0.01–0.07），L_2d 在 PN 扣多了约 +0.06。T11_021 评审：ℓ_impl ≈ 0.8·(1−w)/μ（接近而非吻合），L_2d/(1−w) 对所有相机恒等于 2，相机分化来自数据。T11_023 评审：MOS 的 ℓ_impl / ((1−w)/μ + m) = 1.02（0.96 用 M2 b4 自己的 m），T11_021 的假说 ΔC ≈ (1−m_cam)(μ+δ) − (1−w) 在 M2 b4 上得到支持。2026-10-16 晚间更新（07 R019，取代 R015 的读法）：按模拟自己的口径目录是 −9.282 % [−9.350, −9.219]（11,401 个观测）——早先引用的 −10.4 % 不可复现，已撤回。15 幅图像的联合模拟试点用 emldetect 的估计背景计分：A 臂 −5.9 %、B 臂 −7.4 %（差 3.39 % / 1.85 %），但与目录不匹配：b 0.028 恒定对 0.02→0.14，两个 6–10 窗口落在不同 \(\hat{S}\)（7.3–8.5 对 9.5–13），匹配 \(\hat{S}\) 后差 4–6 %，试点的亏损到 \(\hat{S}\) > 30 消失而目录停在 −7 %，试点 L2d/μ 5.0 % 对目录 9.8 %。Q1 回到 open（闭合未证）；2-D 损失本身仍 Provisional。下一步：按目录给定的 \(\hat{S}\) 采 b（或对现有行按 b 重加权），按 \(\hat{S}\) 分箱在所有窗口用估计背景计分，对用同一代码重算的目录。The two-dimensional form L_2d ≈ 2(1−w) is best for MOS in 6–10, but is not the closing term: not flat across windows (6–10 residual +0.004..+0.034, >20 is −0.009..−0.054); on PN (1−w) closes instead (A1 0.01–0.07), L_2d over-subtracts by ~+0.06. T11_021 review: ℓ_impl ≈ 0.8·(1−w)/μ (closer to, not matches), L_2d/(1−w) = 2 for every camera, so the camera split lies in the data. T11_023 review: MOS's ℓ_impl / ((1−w)/μ + m) = 1.02 (0.96 with M2 b4's own m), the T11_021 hypothesis ΔC ≈ (1−m_cam)(μ+δ) − (1−w) is supported on M2 b4. 2026-10-16 evening update (07 R019, replacing the R015 reading): under the simulation's own convention the catalogue is −9.282 % [−9.350, −9.219] (11,401 observations) — the earlier −10.4 % is not reproducible and is withdrawn. The 15-image joint-simulation pilot, scored with emldetect's estimated background, reaches arm A −5.9 % and arm B −7.4 % (gap 3.39 % / 1.85 %) but is not matched to the catalogue: b 0.028 flat against 0.02→0.14, the two 6–10 windows at different \(\hat{S}\) (7.3–8.5 against 9.5–13), 4–6 % at matched \(\hat{S}\), the pilot's deficit gone by \(\hat{S}\) > 30 while the catalogue stays at −7 %, and L2d/μ 5.0 % against 9.8 %. Q1 is back to open (closure not shown); the 2-D loss itself stays Provisional. Next: sample the catalogue's b given \(\hat{S}\) (or reweight the existing rows by b), scored per \(\hat{S}\) bin over all windows with the estimated background, against the catalogue recomputed with the same code.
6. C-iso 上的宽度缺口。Width gap on C-iso.零参数条目的 A2 是 0.08–0.11，样本内 V5c 是 0.05。E147 宽度变体更好（0.080）但还没完全闭合。这是当前前沿上最明显的空缺。Zero-parameter entries' A2 is 0.08–0.11, in-sample V5c is 0.05. E147 width variant is better (0.080) but still not closed. This is the most obvious gap on the current front.
7. m 的原因。The cause of m.亏损属于真实数据（T5_009）；不是核函数（网格现已扩展到 S = 5000，T5_010/011）。T11_024–026：三台相机都随年代漂移约 −0.001/年；M1 另有一个只在 2012-12（MOS1 CCD3 损失）处的阶跃 −0.016（T11_025 评审更正：M1 = 阶跃，M2/PN = 漂移）；阶跃不是覆盖效应（加入 MASKFRAC 后仍为 −0.0132 ± 0.0024，T11_026）；星表协变量（曝光、背景、离轴角、MASKFRAC、滤光片）都解释不了漂移（M2/PN 斜率仍偏离 0 达 3–4σ）；Ŝ 构成仅凭星表列无法检验（T11_025）。星表列路线已走完，T11 暂停；剩下的决定性检验是真实图像本地重拟合（按 2012-12 前后分开），用户已批准；T13 第 2 阶段已评审但受位置文件 bug 影响（RT-05），第 4 阶段已修复并重做、与修正后的临时数字一致（m_local 0.061 / 0.038）；位置差约占 m_cat 的 25%（M2，未分辨）到 38%（M1），不是"约一半"；新假设（catcorr 校正的 RA/DEC 对未校正的图像 WCS）待检验；第 3 阶段（SAS 拆分）已评审（T13_003，通过附条件：复现 A 成立，B/C 不是机制，D 待 EXT 拆分确认）；第 4 阶段已评审（T13_004，通过附条件：位置 bug 已修复、重做与 RT-05 临时数字一致；方向更正——开启展源拟合时约五分之一的亮点点源拟合计数偏低 12–20%，这是比 ΔC 亏缺更强的可能结论但尚未定论，第 5 阶段的已知真值注入检验决定是管线欠拟合还是纯点源拟合过拟合）。2026-10-16 更新（07 R014）：第 6 阶段的预注册重注入给出答案——大位移组是近邻环境效应，不是源自身轮廓：在大位移源位置上重新注入点源，95–96 % 出现同样的 A-vs-D 位移（对照组 0–6 %），比例随最近邻距离从 0.75（< 20″）降到 0（> 100″）；最近邻距离 34–36″ 对 55–57″。只能比 big ÷ control（重注入 0.846/0.802，目录 0.932/0.868）。"三位小数吻合"、"D 高估 30 %"、"即 15 % 左尾"三句撤回。它占目录 m 的份额还没测 → 待定 / Pending（第 7 阶段在跑）。The deficit belongs to the real data (T5_009); not the kernel (grids now extended to S = 5000, T5_010/011). T11_024–026: all three cameras drift by ~−0.001/yr; M1 additionally has a step of −0.016 at 2012-12 (the MOS1 CCD3 loss) (T11_025 review correction: M1 = step, M2/PN = drift); the step is not a coverage effect (−0.0132 ± 0.0024 with MASKFRAC included, T11_026); no catalogued covariate (exposure, background, off-axis angle, MASKFRAC, filter) explains the drift (M2/PN slopes stay 3–4σ from 0); Ŝ formation is not testable from the catalogue columns (T11_025). The catalogue-column route is exhausted, T11 paused; the remaining decisive test is the real-image local refit (split by 2012-12), approved by the user; T13 stage 2 is reviewed but affected by the position-file bug (RT-05); stage 4 fixed the bug and redid stage 2, matching the corrected provisional numbers (m_local 0.061 / 0.038); the position gap is about 25% (M2, unresolved) to 38% (M1) of m_cat, not "about half"; the new hypothesis (catcorr-corrected RA/DEC against the uncorrected image WCS) is to be tested; stage 3 (the SAS decomposition) is reviewed (T13_003, PASS with conditions: the replica A validated, B/C not the mechanism, D awaiting the EXT split); stage 4 is reviewed (T13_004, PASS with conditions: the position bug fixed and the redo matching RT-05's provisional numbers; direction corrected — with extent fitting on, about one bright point source in five gets fitted counts 12–20 % low, a potentially stronger conclusion than a ΔC deficit but not yet decided; the known-truth injection test in stage 5 decides whether the pipeline under-fits or the pure point fit over-fits). 2026-10-16 update (07 R014): stage 6's pre-registered re-injection answers it — the big-shift group is a near-neighbour environment effect, not a source-profile effect: point sources re-injected at the big-shift positions show the same A-vs-D shift in 95–96 % of cases (control 0–6 %), the fraction falling from 0.75 (< 20″) to 0 (> 100″) with nearest-neighbour distance; that distance is 34–36″ against 55–57″. Only big ÷ control is comparable (re-injection 0.846/0.802, catalogue 0.932/0.868). "Matches to three decimals", "D over-fits by 30 %" and "the same population as the 15 % left tail" are withdrawn. Its share of the catalogue's m is not yet measured → Pending (stage 7 running).
8. 先验效应只在 6–10 窗口内"弱"（RT-10 分诊限定范围）。The prior effect is "weak" only inside the 6–10 window (RT-10 triage, scope restriction).T6_009 的先验效应是 DET_ML 6–10 窗口内两个重加权均值之差；α = 0 的均值随典型 ML 急剧下降（2–3 时 +0.84，>20 时 +0.21），窗口放宽到 3–10 时符号翻转（α = 2：−0.104 → +0.165）。"α ≤ 1 时 plug-in 足够"只适用于 6–10 窗口，不能用于更暗的源表。三种测法给出 −0.05…−0.10，报范围。（来源：reviews/RT-10_triage_claude.md。）The T6_009 prior effect is a difference of reweighted levels inside DET_ML 6–10; the α = 0 level falls steeply with typical ML (+0.84 at 2–3, +0.21 above 20), and widening the window to 3–10 flips the sign (α = 2: −0.104 → +0.165). "The plug-in suffices at α ≤ 1" holds only inside the 6–10 window and must not be used for fainter lists. Three measurements span −0.05…−0.10, so quote the range. (Source: reviews/RT-10_triage_claude.md.)
9. C-iso60 是"观测量上的拥挤过滤器"，不是严格的孤立源层（RT-14 分诊措辞）。C-iso60 is "an observable crowding filter", not a strict isolation layer (RT-14 triage wording).DIST_NN ≥ 60″ 正好等于拟合支撑半径 15 px，2 个支撑要到 120″ 才完全分开；T4 模拟中 24.5% 的暗邻居落在重叠区，但带来的增量只有 −0.002（6–10 frac −0.0394 对无暗邻居的 −0.0374）。C-iso 的入选与 ML 窗口和离轴环相关，不是随机子集；允许最多 10% 的探测器缺失；公共行上的合并 m 相同。（来源：reviews/RT-14_triage_claude.md。）DIST_NN ≥ 60″ is exactly the fit support radius 15 px; the two supports only become disjoint at 120″; 24.5 % of faint neighbours lie in the overlap zone in the T4 simulation, but the increment is only −0.002 (6–10 frac −0.0394 against −0.0374 without faint neighbours). C-iso membership correlates with the ML window and the off-axis ring, so it is not a random subset; up to 10 % missing detector support is admitted; pooled m on common rows agrees. (Source: reviews/RT-14_triage_claude.md.)
10. 核函数高 S 失效已修复。The kernel high-S failure is fixed.全部 15 个网格扩展到 S = 5000，旧节点差值 0.0，新模拟 |frac| < 0.01 到 S = 4000，6–10 分数逐位复现（T5_010/T5_011 评审）；扩展网格设为高计数报告默认核。剩约 0.6% 的乘性底线（L0 m = 0.004–0.008）。注意：本项目所有星表总体都只到 S ≤ 300（为留在旧核网格内）；T5 第 15–17 阶段把报告总体扩到 S ≤ 4000 并修复了 ML→ΔC 溢出（T5_012–014 评审），高 S 报告已重跑并发布——m 延续到亮源，V5c_eta 在 >300 变差，区间根因已修复，报告页已上线（T5_012–015 评审）；T5 已解除暂停（PI 2026-10-16 批准高 S 总体进层定义）。All 15 grids extended to S = 5000, old nodes reproduce to 0.0, fresh sims |frac| < 0.01 up to S = 4000, 6–10 scores reproduce exactly (T5_010/T5_011 reviews); the extended grids are the default kernel for high-count reports. A ~0.6% multiplicative floor remains (L0 m = 0.004–0.008). Note: every catalogue population in this project stops at S ≤ 300 (to stay inside the old kernel grid); T5 stages 15–17 extended the report populations to S ≤ 4000 and fixed the ML→ΔC overflow (T5_012–014 reviews), and the high-S report is rerun and published — m extends to bright sources, V5c_eta degrades in >300, the interval root cause is fixed and the report page is live (T5_012–015 reviews); T5 is unblocked (the PI approved the high-S population for the layer definitions on 2026-10-16).
11. L0 / L1 的正式计分。Formal scoring of L0 / L1.前缀已经显示理论在噪声底内，但正式登记要等生产全部完成（每个列表 50000 窗口行的目标决定了噪声底）。L1 将量化 emldetect 的工程差异。Prefix already shows theory within noise floor, but formal registration awaits full production (50000 window rows per list target sets the noise floor). L1 will quantify emldetect's engineering gap.
12. 私有位置。Private positions.未开启；等 L0 生产完成和 m 问题定案之后的里程碑。Not opened; awaits milestone after L0 production completes and the m question is settled.
13. 普适性与宽度律的六个开放问题（2026-10-23）。Six open problems in universality and the width law (2026-10-23).（1）M1 失败的原因：M1 能段 1 是已分辨的中心失败（8/21 = 0.381）、M1 能段 4 高计数锚 0/12；三个灾难位置占 M1 能段 4 锚行的 14.9 %，排除后均值从 −0.158 变成 −0.003、格从 0/12 变成 9/12（R169/R179，Provisional；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；"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）。（2）掩膜截断规则：低 b 高 \(\hat{S}\) 的宽度失误是少数边缘拟合位置上部分掩膜拟合盒的截断（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，代价是 9.5–11.9 % 的位置（R180，Provisional；RT-127 在跑）。（3）宽度判据的零假设：\(\leq\) 1-of-8 规则是按 8 个箱标定的，emldetect 窗口有 9–10 个箱、可观测神谕本身也过不了 2/9；\(k^*\) 在 2–3（R171），标签已更正（R180：报告的"primary"其实是 superpop 模式），PI 必须选零假设。（4）隔离的新位置：27 个新位置里 8 个受源拥挤污染，POS-FRESH 因此不能解释成位置项的检验（R165/R173）；POS-ISOLATED 设计 139 个 NN \(\geq\) 45 px 的位置（约 500 virgo CPU-h）。（5）目录偏差的形式：\(\mu_0 \lesssim\) 250 通过（R160/R163）；\((f\cdot\mu)^2\) 在 250 以上恢复 MOS 覆盖率但其 PSF 标定依据已撤回（R174），逐相机 \(c\cdot\mu^p\) 在不相交 \(\mu_0\) 区域之间不迁移（R185，Questioned）。（6）Tier 1 的决定：Tier 0 每个块都失败，Tier 1 压住等 PI；修订方案（5 臂，核心约 574 virgo CPU-h）待 PI 批准（R183）。(1) The cause of the M1 failure: M1 band 1 is a resolved centre failure (8/21 = 0.381) and M1 band 4's high-count anchor is 0/12; three catastrophic positions are 14.9 % of M1 band-4 anchor rows, and excluding them moves the mean from −0.158 to −0.003 and the cells from 0/12 to 9/12 (R169/R179, Provisional; 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; 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). (2) The mask-truncation rule: the low-b high-\(\hat{S}\) width miss is support truncation at a few edge fit positions with partially masked fit boxes (R170); the pre-registered rule "exclude positions with median MASKFRAC_EML < 0.98" removes it on held data at a cost of 9.5–11.9 % of positions (R180, Provisional; RT-127 is running). (3) The width rule's null: the \(\leq\) 1-of-8 rule was calibrated for 8 bins, emldetect windows carry 9–10 bins and the observable oracle itself fails 2/9; \(k^*\) is 2–3 (R171) and the labels are corrected (R180: the report's "primary" is actually the superpop mode), so the PI must pick the null. (4) Isolated fresh positions: 8 of 27 fresh positions are source-crowded, so POS-FRESH cannot be read as a test of the position term (R165/R173); the POS-ISOLATED design has 139 positions with NN \(\geq\) 45 px (about 500 virgo CPU-h). (5) The catalogue bias form: the law passes for \(\mu_0 \lesssim\) 250 (R160/R163); \((f\cdot\mu)^2\) restores MOS coverage above 250 but its PSF-calibration justification is withdrawn (R174), and the per-camera \(c\cdot\mu^p\) form does not transfer across disjoint \(\mu_0\) regions (R185, Questioned). (6) The Tier 1 decision: Tier 0 fails on every block so Tier 1 is held for the PI; the revised plan (5 arms, core about 574 virgo CPU-h) awaits PI approval (R183).
14. 星表律的 2026-10-24 更新。The catalogue law's 2026-10-24 update.星表律（2026-10-24 更新）：DET_ML 有序残差结构确认为标注同义反复（RT-133 的法律精确零假设复现切片内 mean_z 展布，实/零之比约 0.93；固定 \(\mu\) 处 DET_ML 对 \(\Delta C\) 单调，按 DET_ML 开窗就是按残差排序——铁律：永不按 DET_ML 条件化）。断幂律（拟合合同加宽到含 \(\mu_0\) > 600 的拟合半）过 6 个主格中的 5 个（M2 \(\mu_0\) 100–250 失败 +0.133，按 OBS_ID 簇自助通过比例 0.000，稳健失败；PN 0.729，勉强）；M2 低 \(\mu_0\) 补丁"m2low"过 6/6 但候选是对手分数已知之后选的，不是确认性证据，等新的留出分半确认。宽度侧：冻结律的 eta 振幅对干净孤立行太大（RT-135：每相机乘 0.5–0.75 把宽度带回约 1）；逐相机振幅 a = 0.81 PN / 0.86 M1 / 0.67 M2 修好合并留出 sd（1.026/0.995/1.021 对 0.956/0.953/0.942），但没有候选通过预注册的逐格规则（最好 16/87 干净格、冻结 6/87），残差偏斜重尾（PN 中位超额峰度 +2.9）→ 下一步是形状（尾）项；且逐格规则本身要按正确模型零假设重新标定（RT-141：正确模型下规则也只过 37/87，约 57 % 是规则伪影，真实超出约 26 格），没有形状候选通过（最好斜正态 12/87）。来源：[s20] 07 R213、R216、R219、R222、R223、R226、R231。The catalogue law (updated 2026-10-24): the DET_ML-ordered residual structure is confirmed a labelling tautology (RT-133's law-exact null reproduces the within-slice mean_z spread, real/null about 0.93; at fixed \(\mu\) DET_ML is monotone in \(\Delta C\), so opening windows in DET_ML sorts rows by residual by construction — the Iron Rule: never condition on DET_ML). The broken power law (with the fit contract widened to include the FIT half of \(\mu_0\) > 600) passes 5 of 6 primary cells (M2 \(\mu_0\) 100–250 fails at +0.133, per-OBS_ID cluster-bootstrap pass fraction 0.000, a robust fail; PN 0.729, marginal); the M2 low-\(\mu_0\) patch "m2low" passes 6/6 but the candidate was chosen after its opponent's score was already known, so it is not confirmatory and awaits a fresh held split. On width: the frozen law's eta amplitude is too large for clean isolated rows (RT-135: multiplying by 0.5–0.75 per camera brings the width back to about 1); the per-camera amplitudes a = 0.81 PN / 0.86 M1 / 0.67 M2 restore the pooled held sd (1.026/0.995/1.021 against 0.956/0.953/0.942), but no candidate passes the pre-registered per-cell rule (the best 16/87 clean cells, the frozen 6/87) and the residual is skewed and heavy-tailed (PN median excess kurtosis +2.9) → the next candidate is a shape (tail) term; and the per-cell rule itself must be re-calibrated against the correct-model null (RT-141: even under a correct model the rule passes only 37/87, about 57 % a rule artefact, the real excess about 26 cells), with no shape candidate passing (the best skew-normal 12/87). Sources: [s20] 07 R213, R216, R219, R222, R223, R226, R231.

Q7 标记来源：[s7] 07 R076；[s8] 07 R062、R065；[s9] 07 R068、R072；[s10] 07 R080；[s11] 07 R077；[s12] 07 R078；[s13] 07 R086（推翻 R083）、reviews/RT-81_triage_claude.md；[s14] 07 R103、R118（SCORER_CONTRACT v2，.agent/jobs/2026-10-21-task-scorer-contract/REPORT.codexnim.md）；[s15] 07 R114（专家 astra，.agent/jobs/2026-10-21-expert-detml-centre/REPORT.astra.md）；[s16] 07 R115（RT-96b）、R116（CENTRE-DECOMP + RT-99）、R117（RT-98）；[s17] 04_DECISION_LOG.md 2026-10-21 05:39 EDT 与 11:42 EDT（PI 进度指标：无位置阈值，已解决位置比例/面积）；[s18] 07 R128–R152（SOLVED-TOL、POOLED-OFFSET、FLOOR-RECOMPUTE、SCORER-FREEZE、POS-TERM2、PHASE-DIAG）、04_DECISION_LOG.md 2026-10-22 04:22 EDT（\(\varepsilon\) 0.05/0.10）与 10:44 EDT（宽度判据、普适性 Tier 1、次要目标）；[s19] 07 R157–R185（RT-115…RT-125：集群 SE 与种子运气撤回、粗锚重解释、Tier 0 口径错误与其撤回、目录 \(\mu_0\) 区域、宽度判据标定、掩膜截断、M1 三个位置、干净总体、修订 Tier 1 方案）；[s20] 07 R189–R231 与 04_DECISION_LOG.md 2026-10-23 15:34 EDT 至 2026-10-24 05:25 EDT（T1-e 三条独立判定与宽度旗标 AMBIGUOUS、背景通道非普适与候选 C2、代价锚改正与 T1-a 发射、星表同义反复与断幂律、eta 振幅与逐格形状、PSF 加权参考驳回、跨拟合膨胀撤回）。Q7 的历史沿革（R030–R085）见历史页 2026-10-20 段。Q7 marker sources: [s7] 07 R076; [s8] 07 R062, R065; [s9] 07 R068, R072; [s10] 07 R080; [s11] 07 R077; [s12] 07 R078; [s13] 07 R086 (reversing R083), reviews/RT-81_triage_claude.md; [s14] 07 R103, R118 (SCORER_CONTRACT v2, .agent/jobs/2026-10-21-task-scorer-contract/REPORT.codexnim.md); [s15] 07 R114 (the astra expert, .agent/jobs/2026-10-21-expert-detml-centre/REPORT.astra.md); [s16] 07 R115 (RT-96b), R116 (CENTRE-DECOMP + RT-99), R117 (RT-98); [s17] 04_DECISION_LOG.md 2026-10-21 05:39 EDT and 11:42 EDT (the PI progress metric: no position threshold, the solved-position fraction/area); [s18] 07 R128–R152 (SOLVED-TOL, POOLED-OFFSET, FLOOR-RECOMPUTE, SCORER-FREEZE, POS-TERM2, PHASE-DIAG), 04_DECISION_LOG.md 2026-10-22 04:22 EDT (\(\varepsilon\) 0.05/0.10) and 10:44 EDT (the width rule, universality Tier 1, the secondary goals); [s19] 07 R157–R185 (RT-115…RT-125: the cluster SE and the seed-luck withdrawal, the coarse-anchor re-interpretation, the Tier 0 convention error and its withdrawal, the catalogue \(\mu_0\) domains, the width-rule calibration, the mask truncation, the three M1 positions, the clean population, and the revised Tier 1 plan); [s20] 07 R189–R231 and 04_DECISION_LOG.md from 2026-10-23 15:34 EDT through 2026-10-24 05:25 EDT (T1-e's three separate verdicts and the AMBIGUOUS width flag, the background channel's non-universality and candidate C2, the cost-anchor correction and the T1-a launch, the catalogue tautology and the broken power law, the eta amplitude and the per-cell shape, the refuted PSF-weighted reference, and the withdrawn cross-fit inflation). The history of Q7 (R030–R085) is on the History page, 2026-10-20 section.

对每条的把握Confidence Levels