工作台Workbench
组会 · 2026-10-24（今天）Group Meeting · 2026-10-24 (today)
成果：普适性 Tier 1 的第一臂（T1-e 留出能段块打分）出结果，而且把一个"整体失败"拆成了三条分开的判定。唯一携带判定的块 M2 能段 1：中心 D+SS 于 \(\varepsilon\) 0.10 是 24/33 = 0.727，配对差区间 [−0.152, +0.061]，按预注册中心规则确认；宽度 4 个失败箱对 \(k^*\) 2，红队复核后改为AMBIGUOUS（超出是单侧的、约 2 %，真实，但箱边一移判定就翻）；锚 8/12。参照块 M2 能段 4：25/33、宽度 1/2、锚 12/12。M1 各能段只作描述，不携带判定。T1-c 背景通道给出今天最重要的新事实：冻结定律在背景通道上不普适——用 DR14 估计背景时窗口偏移 +0.174 z（527 次重复，全部指数 0.171–0.186，簇区间 [0.165, 0.260] 不含 \(\varepsilon\) 0.10），红队排掉十条伪影通道，机制是 \(z_B - z_A \approx 48.2 \times \Delta b_{\rm PSF}\)；候选参数 C2（\(b_{\rm eff} = \gamma \cdot \mathrm{bgm}_B\)，\(\gamma = 0.921 + 0.0041(\mathrm{off} - 3')\)）把留出偏移从 +0.101 压到 −0.052 z，区间 [−0.073, −0.032] 落在 \(\pm 0.10\) 内（一个观测、9 个留出位置，暂定）。代价锚也被改正：实测 0.97 CPU-s/行，是旧 0.4 的约 2.4 倍（区间 0.86–1.3），T1-a 只跑 M2（N = 115，149–192 CPU-h）。工作：T1E-SCORE、T1C-BKG、T1A-STAGE、TIER1-RECOST、BG-PARAM、CAT-WIDE、CAT-M2LOWMU、CAT-ETA、PSFW-REF、PY-PARSE、PN-INGEST 等任务交付；RT-128 至 RT-141 十四支红队分诊完毕；T1-a 已按 PI 选项 A 发射（M2-only，3 个新观测）。问题：调度方自己的错误——把 R158 记成失败时没有核对打分口径（两种口径在这批数据上差 0.12 的已解决比例、翻转 4 个位置），该"失败"三小时后由红队定位为口径错误并撤回；执行方把"整体 REFUTED"当成 M2 能段 1 的总结，而方案第 5 节只预注册了三条独立规则、没有合并规则（R220）；PN b1/b4 的 lang2 源树已不存在（缓存盘），完整 py 输入要约 397 GB，PI 决定不重跑。待 PI：无（T1-a 预算与 PN b1/b4 两项已于 02:47 EDT 决定；T1-b 与 T1-d 的预算按新代价锚重算后再问）。Results: the first Tier 1 arm (T1-e, the held-out band-block scoring) reported, and it split one "overall failure" into three separate verdicts. On the only verdict-carrying block, M2 band 1: the centre D+SS at \(\varepsilon\) 0.10 is 24/33 = 0.727, paired-difference interval [−0.152, +0.061], CONFIRMED under the pre-registered centre rule; the width has 4 failing bins against \(k^*\) 2, re-flagged AMBIGUOUS after the red-team re-check (the excess is one-sided, about 2 %, real, but the verdict flips under a bin-edge shift); the anchor is 8/12. The reference block M2 band 4: 25/33, width 1/2, anchor 12/12. The M1 bands stay descriptive, never verdict-carrying. T1-c, the background channel, gave today's most consequential new fact: the frozen law is not universal in that channel — with the DR14-estimated background the window offset is +0.174 z (527 repeats, all indices 0.171–0.186, cluster interval [0.165, 0.260] excluding \(\varepsilon\) 0.10); the red team excluded ten artefact channels and the mechanism is \(z_B - z_A \approx 48.2 \times \Delta b_{\rm PSF}\); the candidate parameter C2 (\(b_{\rm eff} = \gamma \cdot \mathrm{bgm}_B\), \(\gamma = 0.921 + 0.0041(\mathrm{off} - 3')\)) brings the held offset from +0.101 to −0.052 z, its interval [−0.073, −0.032] inside \(\pm 0.10\) (one observation, 9 held positions, provisional). The cost anchor was also corrected: the measured cost is 0.97 CPU-s/row, about 2.4\(\times\) the old 0.4 (band 0.86–1.3), so T1-a runs M2 only (N = 115, 149–192 CPU-h). Work: T1E-SCORE, T1C-BKG, T1A-STAGE, TIER1-RECOST, BG-PARAM, CAT-WIDE, CAT-M2LOWMU, CAT-ETA, PSFW-REF, PY-PARSE and PN-INGEST delivered; red teams RT-128 through RT-141 triaged; T1-a launched under the PI's option A (M2-only, 3 new observations). Problems: the dispatcher's own error — recording R158 as a failure without checking the scoring convention (the two conventions differ by 0.12 of solved fraction on this data and flip 4 positions), the "failure" being traced to a convention error by a red team three hours later and withdrawn; the executor reported an "overall REFUTED" for M2 band 1 while plan §5 pre-registered only three separate rules with no merge rule (R220); the lang2 source tree for PN b1/b4 no longer exists (a cache disk), a full py input would be about 397 GB, and the PI decided not to re-run. For the PI: none (the T1-a budget and PN b1/b4 were decided at 02:47 EDT; the T1-b and T1-d budgets go back to the PI after the recost with the new anchor).
完整四段式汇报见The full four-part report is on the 组会页Group Meetings page。.
组会 · 2026-10-23Group Meeting · 2026-10-23
成果：普适性 Tier 0 出结果，而且推翻了一个失败。能段/仪器轴上，A0-SCORE 报的失败（R158）被 RT-117 定位为打分手口径错误并撤回（R162）：它把 SS 位置项逐行除以 \(s\sqrt{V_{3p}}\)，而冻结打分手把它从合并偏移里减掉；按冻结口径、零新参数，M2 能段 1 的 D+SS 在 \(\varepsilon\) 0.10 处是 24/33 = 0.7273，高于门槛 0.6273。A0-RESCORE 随后给出全貌：M1 能段 1 是 8/21 = 0.381，一个已分辨的中心失败；M1 能段 4 的高计数锚 0/12；Tier 0 判定每个块都失败，Tier 1 继续压着等 PI。星表轴上，冻结的宽度律在 \(\mu_0 \lesssim 250\) 域内通过、以上失败（R160/R163），加 \((f\cdot\mu)^2\) 一项把 MOS 在 250 以上的覆盖率从 0.452 修到 0.678（R168），但没有任何偏差项单独管用（R172），PSF 定标的理由也说过头了（R174）。宽度一侧：聚类标准误把 V̄ 的失败数从 7/9 降到 5/9（R157）；"窄核靠误差相消通过宽度"是粗锚假象（R159）；细锚规律在 emldetect 上达到留出 oracle 下限 2/9（R164），约 2 % 的 RMS 带区间 [1, 7] %（R166）；"8 箱允许 1 箱失败"判据要重新定标（R171，PI 选零假设）。位置项在新位置上 AMBIGUOUS（R165），但 27 个新位置里 8 个源太挤，这一测试不能当作 SS 项的检验（R173）。B011 符号缺陷修好后解释 0 % 的失败（R161）。工作：十支红队分诊（RT-113 至 RT-122）加十个任务交付（A0-SCORE、A0-CATALOG、B011-REPLAY、FINE-ANCHOR、POS-FRESH、CAT-HIGHMU、A0-RESCORE、LOWB-HARD、WIDTH-RULE-CAL、CAT-BIAS）；M1-ANCHOR、MASKFRAC-ANCHOR、CAT-PSFMU、POS-ISOLATED、RT-123、RT-124 在跑（名字，不引用数字）。问题：调度方自己的错误——把 R158 记成失败时没有核对打分口径（两种口径在这批数据上差 0.12 的已解决比例、翻转 4 个位置），BAND-SCALE 因此被 PID 文件停掉；R170 的离群位置是事后挑的；POS-FRESH 的 27 个新位置里 8 个源太挤（设计缺陷）。待 PI：Tier 1 是否派发（选项 A/B/C，推荐 B 等 M1-ANCHOR 诊断）；宽度验收规则的零假设（R171）；星表高 \(\mu_0\) 域的偏差形式。Results: the universality Tier 0 reported, and it overturned one failure. On the band/instrument axis the failure A0-SCORE reported (R158) was traced by RT-117 to a scorer-convention error and withdrawn (R162): it divided the SS position term per row by \(s\sqrt{V_{3p}}\), whereas the frozen scorer subtracts it from the pooled offset; under the frozen convention with zero new parameters, M2 band 1's D+SS at \(\varepsilon\) 0.10 is 24/33 = 0.7273, above the bar 0.6273. A0-RESCORE then gave the full picture: M1 band 1 is 8/21 = 0.381, a resolved centre failure; M1 band 4's high-count anchor is 0/12; Tier 0 fails on every block, so Tier 1 stays held for the PI. On the catalogue axis the frozen width law passes within the \(\mu_0 \lesssim 250\) domain and fails above (R160/R163); adding an \((f\cdot\mu)^2\) term restores MOS coverage above 250 from 0.452 to 0.678 (R168), but no bias term works alone (R172) and the PSF-calibration justification is overstated (R174). On width: the cluster SE moves V̄'s failures from 7/9 to 5/9 (R157); "the narrow kernel passes width by error cancellation" is a coarse-anchor artefact (R159); the fine-anchor law reaches the held-out oracle floor of 2/9 on emldetect (R164), its about-2 % RMS carrying a cluster interval [1, 7] % (R166); the "8 bins with 1 failing bin allowed" rule needs recalibration (R171, the PI picks the null). The position term at fresh positions is AMBIGUOUS (R165), but 8 of the 27 fresh positions are source-crowded, so that test cannot be read as a test of the SS term (R173). The B011 sign defect explains 0 % of the failures once fixed (R161). Work: ten red teams triaged (RT-113 through RT-122) and ten tasks delivered (A0-SCORE, A0-CATALOG, B011-REPLAY, FINE-ANCHOR, POS-FRESH, CAT-HIGHMU, A0-RESCORE, LOWB-HARD, WIDTH-RULE-CAL, CAT-BIAS); M1-ANCHOR, MASKFRAC-ANCHOR, CAT-PSFMU, POS-ISOLATED, RT-123, RT-124 running (names, no numbers quoted). Problems: the dispatcher's own error — recording R158 as a failure without checking the scoring convention (the two conventions differ by 0.12 of solved fraction on this data and flip 4 positions), after which BAND-SCALE was stopped by its PID file; R170's outlier positions were picked post hoc; 8 of POS-FRESH's 27 fresh positions are source-crowded (a design defect). For the PI: whether to dispatch Tier 1 (options A/B/C, recommending B, waiting for the M1-ANCHOR diagnosis); the width acceptance rule's null (R171); the bias form for the catalogue's high-\(\mu_0\) domain.
完整四段式汇报见The full four-part report is on the 组会页Group Meetings page。.
组会 · 2026-10-22Group Meeting · 2026-10-22
成果：进度指标被定死、地板被改正、打分手被冻结的一天。PI 在 04:22 EDT 定下容差：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）。"每条条目都在 \(\varepsilon\) 内"的判据被证明噪声受限（R128/R129/R133：完美规律在能段 4 的天花板只有 0.21–0.27），改成每个位置一个合并偏移（R131）；R131 说的"规律就压在位置盲地板上"被 FLOOR-RECOMPUTE 推翻（R138：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）。打分手冻结为 score_pooled_v1.py（R147：18 格、index 1.5–3、面积环 [0, 4, 8, 14.01]、位置真值校准的 95 % CI；\(\varepsilon\) 0.10 处 D 0.545、D+SS 0.727）。位置效应：不是源表（R126：每个位置注入同一份 54 源栅格），主要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）；把注入点移到像素中心反而把偏移放大约 2 倍（R144，未解释，PHASE-DIAG 在查）。第一个拟合前位置项（拟合盒背景水平）把 D 在 \(\varepsilon\) 0.10 处从 0.545 抬到 0.727（R139；R148 诚实 p 0.003–0.004，不是相位主导，尚未在新位置检验，POS-FRESH 在跑）。宽度：Li et al. 2025 展开只在 b 0.01 精确（R121/R130）；可观测尝试失败（R127/R136）；失败原因是箱内 Eddington 混合加一条完美规律也只有 67 % 通过的判据（R143）；可观测箱级规律 V̄ 在"8 箱允许 1 箱失败"的校正判据下于 \(\beta\) −1.5 通过（R146）——判据变更等 PI。工作：十五个任务当天完成并被复核（R125–R139、R143、R144、R146、R147）；红队 RT-104 至 RT-111 分诊；专家 astra 主线排序采纳为 R142；08 到 rev 137；临时盘存档完成（879 MB，archive/scratch/MANIFEST.md）。问题：调度方自己的错误——把"所有项目合计 130 GB"当成 DET_ML 的临时盘大小（实际约 29 GB）；照抄了执行方"\(\varepsilon\) 0.10 最敏感"的错误论证（R137 复算推翻）；一支 Hermes 红队跑了 git 并留下过期的 index.lock（已清除）。待 PI：宽度判据是否改为"8 箱允许 1 箱失败"（R143/R146）；临时盘删除清单（MANIFEST）；普适性设计（PI 2026-10-22 的问题：什么多样性证明规律普适——UNIVERSALITY-DESIGN 在跑）。Results: a day the progress metric got its tolerance fixed, its floors corrected and its scorer frozen. At 04:22 EDT the PI set the tolerances: \(\varepsilon\) 0.05 to compare laws, \(\varepsilon\) 0.10 as the progress tracker (\(\varepsilon_w\) 0.10), \(\varepsilon\) 0.20 dropped as saturated. 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 the criterion became one pooled offset per position (R131); R131's "the laws sit at the position-blind floor" is overturned by FLOOR-RECOMPUTE (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 scorer is frozen as score_pooled_v1.py (R147: 18 cells, indices 1.5–3, area annuli [0, 4, 8, 14.01], position-true-calibrated 95 % CI; at \(\varepsilon\) 0.10 D 0.545, D+SS 0.727). The position effect: not the source list (R126: every position injects the same 54-source grid), mostly 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); moving the injection points to pixel centres amplified the offsets about 2\(\times\) instead (R144, unexplained, PHASE-DIAG running). A first pre-fit position term (the fit-box background level) lifts D at \(\varepsilon\) 0.10 from 0.545 to 0.727 (R139; R148 honest p 0.003–0.004, not phase-led, not yet tested on fresh positions — POS-FRESH running). Width: the Li et al. 2025 expansion is exact only at b 0.01 (R121/R130); the observable attempts fail (R127/R136); the failures are the within-bin Eddington mixture plus a rule a perfect law passes only 67 % of the time (R143); the observable bin-level law V̄ passes at \(\beta\) −1.5 under the calibrated "1 failing bin of 8 allowed" rule (R146) — the rule change awaits the PI. Work: fifteen tasks finished and re-checked the same day (R125–R139, R143, R144, R146, R147); the red teams RT-104 through RT-111 triaged; the astra expert's main-line ranking accepted as R142; 08 at revision 137; the scratch archive done (879 MB, archive/scratch/MANIFEST.md). Problems: the dispatcher's own errors — reporting "130 GB" (all projects combined) as the DET_ML scratch size (actually about 29 GB); copying the executor's wrong "\(\varepsilon\) 0.10 is most sensitive" argument (refuted by R137's replay); a Hermes red team ran git and left a stale index.lock (removed). For the PI: whether the width rule becomes "1 failing bin of 8 allowed" (R143/R146); the scratch deletion list (MANIFEST); the universality design (the PI's 2026-10-22 question: what diversity proves a law universal — UNIVERSALITY-DESIGN running).
完整四段式汇报见The full four-part report is on the 组会页Group Meetings page。.
组会 · 2026-10-21Group Meeting · 2026-10-21
成果：规律阶段被一次专家复核改写。专家（R114 采纳）指出旧打分手混用三个估计对象——宽度是"位置内"的，均值/覆盖率跨位置合并，零假设还跟着候选自适应——所以"宽度可以推广"只能读作"在同一个位置内成立"，72 格是诊断而非认证。RT-98（R117）给出精确结论：宽度指标对逐位置偏移恒等不变（1e-15），一条带 0.30 z 位置偏移的规律仍通过宽度 70/72。新打分手（SCORER-CONTRACT v2，R118）分开两个目标：未见位置总体上能段 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）；逐位置上每条规律都接近 0/72（能段 4 为 1 / 0 / 1 / 0，能段 1 为 0 / 1），冻结位置效应 \(\tau_{mean}\) 0.14 z（能段 4）/ 0.18 z（能段 1），重尾 df \(\approx\) 3。中心项（R116）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），不是位置泛化缺口；专家那条先验混合恒等式作为代数成立，但其协方差项比失败格里的误差大 2–8 倍且已被 D 自己的 \(\beta\) 斜率抵消（RT-99）。宽度读数（R115）：0.976 的留出宽度只是一次偏低的分半，40 个分半平均 0.998；6 参数修正只通过 25/72 个宽度格。工作：符号回归试点（R109：没有公式配得上 D）、LAW-HYBRID（R110/R112：H2 是唯一有用的中心修补，部分且不稳健）、CENTRE-Z（R111：z 单位重拟合后留出判定一格不变）、CENTRE-DECOMP、SCORER-CONTRACT 当天完成并被红队复核；RT-95、RT-97、RT-98、RT-99、RT-100、RT-101 分诊完毕；专家 astra 的报告采纳为 R114；08_RUNNING.md 到 rev 123。WIDTH-2ND、RT-100（域扫描）、RT-101 在跑。问题：调度方自己的复算失败（分诊 RT-96b 时抽错了代码块，R115）；调度方改 08 头部时留下杂散星号；两次发射时 CPU 读数高于 60 %（RT-98 记 77–87 %，RT-100 记负载 15.8–16.0）——两支红队自己的占用都是一个核，但整机门限在繁忙时段会被别人的作业顶上去；score_v2.py 在 25 分钟内经历三个 tau 体制（RT-98），最终版的地板塌陷已被 RT-100 否定，但"约 33 个位置无法辨认逐格位置尺度"的结构性结论保留。PI 已定并当天修订（04）：05:39 EDT 的验收目标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；11:42 EDT 修订为没有位置上的通过/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某个阈值被排除；每条规律都按已解决位置比例（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 % 的默认值已撤回。RT-100 指出 MASKFRAC \(\geq\) 0.9 在现有行上近乎空操作（99.92 % 已满足），真正的判别维度是离轴 \(\theta_{max}\)。Results: an expert review rewrote the law stage. The expert (accepted as R114) showed the old scorer mixes three estimands — the width is within-position, the mean and coverage pool across positions, and the null adapts to the candidate — so "the width generalises" can only be read as "within one position", and the 72 cells are a diagnostic, not certification. RT-98 (R117) gives an exact result: the width metric is identically invariant (1e-15) to per-position offsets, so a law carrying a 0.30 z position offset still passes width 70/72. The new scorer (SCORER-CONTRACT v2, R118) separates two targets: on the population of unseen positions 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); per position every law is near 0/72 (band 4: 1 / 0 / 1 / 0; band 1: 0 / 1), with the frozen position effect \(\tau_{mean}\) 0.14 z (band 4) / 0.18 z (band 1), heavy-tailed with df \(\approx\) 3. The centre term (R116) is dominated by an in-sample representation error of the D table itself (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 2–8\(\times\) the error in the failing cells and is already cancelled by D's own \(\beta\) slope (RT-99). The width reading (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. Work: the symbolic-regression trial (R109: no formula matches D), LAW-HYBRID (R110/R112: H2 is the only centre repair that helps, partial and not robust), CENTRE-Z (R111: a z-unit refit leaves the held-out verdict unchanged in every cell), CENTRE-DECOMP and SCORER-CONTRACT all finished and were re-checked by red teams the same day; RT-95, RT-97, RT-98, RT-99, RT-100 and RT-101 are triaged; the astra expert's report is accepted as R114; 08_RUNNING.md is at revision 123. WIDTH-2ND, RT-100 (the domain sweep) and RT-101 are running. Problems: the dispatcher's own replay failed (it extracted the wrong code block while triaging RT-96b, R115); the dispatcher's 08-header edit left stray asterisks; the CPU reading was above 60 % at two launches (RT-98 records 77–87 %, RT-100 a load of 15.8–16.0) — both red teams' own footprint stayed at one core, but the whole-machine gate is pushed up by other people's jobs in busy periods; score_v2.py went through three tau regimes in 25 minutes (RT-98), and while the final version's floor collapse is refuted by RT-100, the structural conclusion stands that roughly 33 positions cannot identify a per-cell between-position scale. Decided by the PI and amended the same day (04): the 05:39 EDT acceptance target was per-position accuracy at most positions over the largest possible domain (e.g. all positions within 14′ off-axis), with quality cuts such as MASKFRAC allowed; at 11:42 EDT it was amended to 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. The 90 % default is withdrawn. RT-100 notes that MASKFRAC \(\geq\) 0.9 is nearly vacuous on the existing rows (99.92 % already satisfy it), so the discriminating axis is the off-axis \(\theta_{max}\).
完整四段式汇报见The full four-part report is on the 组会页Group Meetings page。.
组会 · 2026-10-20（今天）Group Meeting · 2026-10-20 (today)
成果：宽度主线（Q7）。当天五条修订记录全部来自红队分诊，且都被复算。R064：真实 PSF 的 \(V_{3p}\) 公式本身没有错——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）；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，“方差高估 1.9\(\times\)”的说法删除。R065：这条真实 PSF 规律在 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。R066：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（DET_ML 1–6）、\(\approx\) 0.95（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。R067：同 67 位置、同 S^−2.5 密度、同 (S, b) 格子权重下配对比较，L1 − L0 = +0.24 / +0.19 / +0.12 / +0.08 / +0.07 于 DET_ML \(\approx\) 2 / 3 / 4 / 5 / 7（约 50 处持平），比原先记录的差别更大；但两边是两条流水线加两种打分与两个振幅估计量，不归因于拟合器。R068：R064 提示的修正——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——只把 std z 从 0.88 / 0.92 / 0.94 / 0.96 抬到 0.975 / 0.959 / 0.970 / 0.979（DET_ML 6.9 / 9 / 12 / 18.5，消掉 45–79 %），回不到 1 \(\pm\) 0.02；该提案以否定结论关闭。R072：按实测 MLE 偏差校准的 \(\hat{S}\) 修正（残差偏差 \(\leq\) 1.1 %）仍留 std z 0.952 / 0.959 / 0.968 / 0.972（DET_ML 6.8 / 8.9 / 12.1 / 18.6，2.0–3.6 SE）——“更好的偏差修正”这条路关闭；带 Fisher 展布的 \(E[V_{3p} \mid \hat{S}]\) 更差（0.84–0.96）。残差与先验无关（\(\leq\) 0.6 SE），但机制已改写（R077/R078）：\(V_{3p}\) = term1 − 4\(\cdot\)penalty 是近乎全抵消，term1 是 Var(R) 的 15–31 倍，Fisher penalty 的 \(\approx\) 1 % 错配被抵消放大成 8–16 % 的方差高估，且错配随 PSF 的中心集中度上升（真实 PSF +1.0–1.8 %，同二阶矩高斯 \(\approx\) 0.05 %）。R079（RT-74，已复算）：L0 的低 DET_ML 中心项依赖拟合离轴 \(r\)（固定 \(\hat{S}\) 下 \(r\) 从 < 0.5 到 \(\geq\) 4 时，+0.71 \(\to\) +0.23 在 \(\hat{S}\) < 4.73，+0.10 \(\to\) −1.34 在 \(\hat{S}\) 23–30），比两个已验证的纯构造对照高 +0.23…+0.50 z（DET_ML \(\lesssim\) 8，\(\geq\) 11 处约 0），67/67 个位置为正；R076 的“近乎抵消”注记已撤回（两个 gap 实为净 +0.95…+1.44 z）；归因开放（RT-77 在跑）。0.5\(\times\) 背景下同一校准在 DET_ML 7–13 闭合到 0.99–1.00 的读数已废弃（R074）：那是单一种子的结果，且基准校准在 0.5\(\times\) 下过度修正（残差偏差 −0.4…−6.5 %）——三个种子合并后 1 − z = −0.003 / +0.009 / +0.026 / +0.014 / +0.023（DET_ML 7.0–12.7），偏差匹配的 0.5\(\times\) 校准仍留 5.1 %（7.4 SE），2\(\times\) 下亏损 0.87–0.95——不存在随背景变化的闭合。R073：宽度不足对目录用户是实在的，但 L0 上 68 % 区间的净过覆盖只有 +2–5 pp（DET_ML 3.5–15）——它是宽度项（+5–11 pp）与中心偏移项（−1–10 pp）相互抵消的结果；只修宽度反而会欠覆盖（DET_ML 4.9 处 −2.9 pp）。阈值 DET_ML = 6 的探测概率只差 +0.016，不构成实质问题。今后目标改为区间覆盖率误差 \(\leq\) 2 pp，宽度与中心一起校准。R075/R076（L0 打分）：L0 的中心项在四种打分下都为负，但大小差 6.6–11 倍；只有 z_off 通过高计数锚（DET_ML \(\geq\) 50 处 mean \(\approx\) 0、std 0.989–1.003，所有位置组、两种先验），z_base 的“变宽”是中心缺陷——z_off 是 L0 打分，R065 的偏窄恢复为 L0 宽度律；每个 L0 数字都要带打分名（z_base 的 std z 1.03–1.21 于 DET_ML 11–37 是历史）。R077/R078（机制）：\(V_{3p}\) = term1 − 4\(\cdot\)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 \(\approx\) 1 % 的错配被抵消放大成 8–16 % 的方差高估；错配随 PSF 的中心集中度上升（同二阶矩的高斯 \(\approx\) 0.05 %，真实 PSF +1.0–1.8 %，DET_ML \(\lesssim\) 6 处系统性，\(\approx\) 10 以上\(\approx\) 0）——所以不是求值点误差，没有任何点修正能闭合它；候选修正是一个集中度依赖的二阶 penalty 项。R080（RT-76，已复算）：这个候选未达成——lib/ 里没有能闭合宽度的解析二阶 penalty（Bartlett 的 +2\(\cdot\)rank 项符号反了；beyond-Fisher 的位置曲率项过度修正 1.3–7.4\(\times\)，在最低 \(\hat{S}\) 处给负方差，在最高分箱处变号）；错配跟随 PSF 的翼结构（峰值只是固定二阶矩（2nd-moment）族内的代理）与计数（\(\hat{S}\), b）——固定 \(\hat{S}\) 把 b 减半会把错配抬高 +0.5…+0.9 pp（\(\hat{S}\) \(\lesssim\) 5）；实用路线仍是 L0 标定的区间。工作：分诊当天十七支宽度红队（R064–R080），另发两支（在跑）；第 34 阶段经审阅（PASS）后部署为 a94494ac；第 33 阶段（stage 32 的重启）折叠 R064–R071 后部署为 25620f8d；T12 第 29–30 阶段经审阅（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。问题：调度方与红队自己的错误——第 29 阶段在 400 轮上限处停下并损坏又修复 index.html；四项评审缺陷（图题方向、假比较句、图注、双反斜杠，其中 40 处来自评审者自己前一天横幅里的双反斜杠）；RT-61 的复算在 virgo 守卫显示 WAIT 时运行（单线程、nice、几分钟，已记入会话日志）；横幅草稿先写了未来时间 03:50 才改正；第一次续跑 T12 用了相对路径而失败；RT-65 第一轮只交了报告骨架；第 33 阶段没写检查点就结束并把自己的 PID 误认成重复 worker；一支红队第一轮中途停下（13:23Z 续跑后完成）且报告抄错自己脚本的一列（relsd）；深夜又有四次：一次锚复算在 virgo 守卫显示 WAIT 时运行（当天第三次，已把守卫写进同一条命令）；RT-69 分诊里“宽度不依赖 \(\delta\)”的括注对 L0 是错的（R075 已更正）；0.5\(\times\) 背景下“闭合到 0.99–1.00”是从单一种子记下的（R074 已废弃）；RT-75 的任务说明把 T233 写错了。2026-10-20T02:22Z PI 指示（08 §6）：「现在我不关心星表。还是要结果emldetect的结果。」“星表经验区间”方向已撤回，主线是 emldetect 结果；任务 0b（codexnim）与 astra 专家已派出；astra 已分诊（结论目标未达成）；0b 晚间完成（R089）：emldetect 与 L0 高计数一致、DET_ML ≲ 7 处中心低 0.16–0.53 z、宽 0.05–0.14，所以 L0 低计数超额不是 emldetect 的性质。当天晚些时候 R081–R085：计数依赖在共同掩膜上仍成立但 \(\hat{S}\) 约 5 的数值是掩膜假象；L0 的低 DET_ML 中心项超额是 L0 流水线自己的性质、成因开放（R087 撤回 72/28 归因）；penalty 错配机制也解释 L1 宽度（R086 推翻 R083：合并估计量下 L1 错配 +1.19 % [+1.03, +1.35]，解释箱内宽度缺口的 0.89 [0.84, 0.93]，预测 std z 0.919 对实测 0.918）——同一个 \(V_{3p}\) 机制同时决定 L0 与 emldetect 的宽度。下午（R092–R104）：主线从机制推进到规律，四条候选规律依次被检验，打分手本身错了两次并修好，最后由一支被验证的评分器（NULL-1，R103：在恰好为真的数据上 p 均匀）给出唯一结论——四条候选全部被拒：M2 能段 4 上 C 10/72、D 48/72、DQ 42/72、MIX3d 35/72，能段 1 上 D 重拟合 37/72（对零假设均值约 62–63，D 的 p = 0.0015）；R094 给出噪声底（完美规律平均也只过约 61/72），R100 撤回 R096 的“p = 0 全局拒绝”（零假设尺度错误）。R104 拆开 D：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 的锚取决于位置 14 的分半）。CENTRE-POS（已完成，R105）：位置不是缺失的中心变量（平滑位置项把留出失败从 24 推到 58–72；逐格中心截距 24 \(\to\) 11）。PI 批准符号回归试点（PySR + SRHarness，NIM 后端）与混合规律任务。Results: the width main line (Q7). All five revision records of the day come from red-team triages and all were replayed. R064: the real-PSF \(V_{3p}\) formula itself is not wrong — 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 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 and the “variance over-stated by 1.9\(\times\)” wording is dropped. R065: the real-PSF law 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. R066: 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, and the lowest bin is wide (1.31); T228's b_true-labelled cell numbers are demoted to history. R067: in a matched comparison (the same 67 positions, the same S^−2.5 density, the same (S, b) cell masses) L1 − L0 = +0.24 / +0.19 / +0.12 / +0.08 / +0.07 at DET_ML \(\approx\) 2 / 3 / 4 / 5 / 7 (level by about 50), larger than recorded; but the two sides are two pipeline + scoring pairs and two amplitude estimands, so the gap is not attributable to the fitter. R068: the correction R064 suggested — evaluating \(V_{3p}\) at the observable Cox–Snell bias-corrected amplitude \(\hat{S} - b_S(\hat{S})\) — lifts std z only from 0.88 / 0.92 / 0.94 / 0.96 to 0.975 / 0.959 / 0.970 / 0.979 at DET_ML 6.9 / 9 / 12 / 18.5 (45–79 % of the deficit removed) and not to 1 \(\pm\) 0.02; the proposal is closed negative. R072: a \(\hat{S}\) correction calibrated to the measured MLE bias (residual bias \(\leq\) 1.1 %) still leaves std z 0.952 / 0.959 / 0.968 / 0.972 at DET_ML 6.8 / 8.9 / 12.1 / 18.6 (2.0–3.6 SE) — the “better bias correction” route is closed; \(E[V_{3p} \mid \hat{S}]\) with a Fisher spread is worse (0.84–0.96). The residual is prior-independent (\(\leq\) 0.6 SE), but the mechanism is rewritten (R077/R078): \(V_{3p}\) = term1 − 4\(\cdot\)penalty is a near-total cancellation, term1 is 15–31\(\times\) Var(R), and the Fisher penalty's \(\approx\) 1 % mismatch is amplified by the cancellation into an 8–16 % variance over-statement, with the mismatch rising with the PSF's central concentration (real PSF +1.0–1.8 %, a Gaussian of equal second moment \(\approx\) 0.05 %). R079 (RT-74, replayed): the L0 low-DET_ML centre term depends on the fitted offset \(r\) (at fixed \(\hat{S}\), from \(r\) < 0.5 to \(\geq\) 4: +0.71 \(\to\) +0.23 at \(\hat{S}\) < 4.73 and +0.10 \(\to\) −1.34 at \(\hat{S}\) 23–30), exceeding two validated pure-construction controls by +0.23…+0.50 z (DET_ML \(\lesssim\) 8, \(\approx\) 0 at \(\geq\) 11) and positive at 67/67 positions; R076's “near-cancellation” note is withdrawn (the two gaps are a net +0.95…+1.44 z); attribution is open (RT-77 running). R074 (background): the reading that at 0.5\(\times\) background the same calibration closes to 0.99–1.00 at DET_ML 7–13 is deprecated (R074): it was one seed, and the base calibration over-corrects at 0.5\(\times\) (residual bias −0.4…−6.5 %) — pooled over three seeds 1 − z = −0.003 / +0.009 / +0.026 / +0.014 / +0.023 (DET_ML 7.0–12.7), a bias-matched 0.5\(\times\) calibration still leaves 5.1 % (7.4 SE), and at 2\(\times\) the deficit is 0.87–0.95 — no background-dependent closure exists. R073: the width deficit is real for a catalogue user, but on L0 the 68 % interval's net over-coverage is only +2–5 pp (DET_ML 3.5–15) — a cancellation of a width term (+5–11 pp) and a centre term (−1–10 pp); closing the width alone would under-cover (−2.9 pp at DET_ML 4.9). The detection probability at the DET_ML = 6 threshold is off by only +0.016, which is not material. The goal is re-stated as interval-coverage error \(\leq\) 2 pp, with the width and the centre calibrated jointly. R075/R076 (L0 scoring): the L0 centre term is negative under all four scorings but differs 6.6–11\(\times\) in size; only z_off passes the high-count anchor (mean \(\approx\) 0 and std 0.989–1.003 at DET_ML \(\geq\) 50, every position group, both priors), and z_base's widening is a centre defect — z_off is the L0 scoring and R065's narrowing stands as the L0 width law; every L0 number must name its scoring (z_base's std z 1.03–1.21 at DET_ML 11–37 is history). R077/R078 (mechanism): \(V_{3p}\) = term1 − 4\(\cdot\)penalty is a near-total cancellation (term1 is 15–31\(\times\) Var(R)); the Fisher penalty matches the effective penalty to 0.1–1.4 %, and this \(\approx\) 1 % mismatch amplified by the cancellation gives the 8–16 % variance over-statement; the mismatch rises with the PSF's central concentration (a Gaussian of equal second moment \(\approx\) 0.05 %, the real PSF +1.0–1.8 %, systematic at DET_ML \(\lesssim\) 6, \(\approx\) 0 above \(\approx\) 10) — so it is not an evaluation-point error, no point correction can close it; the candidate fix is a concentration-dependent second-order penalty term. R080 (RT-76, replayed): that candidate is not achieved — no analytic second-order penalty in lib/ closes the width (the Bartlett +2\(\cdot\)rank term has the wrong sign; the beyond-Fisher position-curvature term over-corrects 1.3–7.4\(\times\), gives negative variance at the lowest \(\hat{S}\) and flips sign at the top bin); the mismatch tracks the PSF's wing structure (peak is only a proxy within a fixed-2nd-moment family) and the counts (\(\hat{S}\), b) — halving b at fixed \(\hat{S}\) raises the mismatch by +0.5…+0.9 pp (\(\hat{S}\) \(\lesssim\) 5); the practical route stays an L0-calibrated interval. Work: seventeen red teams triaged (R064–R080); two more are launched (running); stage 34 was reviewed (PASS) and deployed as a94494ac; stage 33 (the fresh restart of stage 32) folded R064–R071 and is deployed as 25620f8d; T12 stages 29–30 reviewed (PASS with banner: a reversed figure title, a false §15 comparison sentence, a wrong caption, 62 doubled backslashes) and deployed as c742d5e6, then stage 31 fixed all five items and deployed as f0f2b93d. Problems: the dispatcher's and the arms' own errors — stage 29 stopped at the 400-turn budget and damaged then repaired index.html; four review defects (the reversed figure-title direction, the false comparison sentence, the wrong caption and the doubled backslashes, 40 of which came from the reviewer's own previous-day banner); the RT-61 replay ran on virgo while the guard said WAIT (single-threaded, niced, a few minutes, recorded in the session log); a banner draft first carried a future time (03:50) and was corrected; the first T12 resume failed on a relative path; RT-65's first turn ended with only a skeleton report; stage 33 ended without its checkpoint and misread its own PID as a duplicate worker; one red team's first turn stopped mid-work (it finished after the 13:23Z resume) and its report mis-transcribed one column from its own script (relsd); four more late that night: an anchor replay ran while the virgo guard said WAIT (the third that day; the guard is now in the same command); the RT-69 triage's parenthetical “the width does not depend on \(\delta\)” is wrong for L0 (corrected by R075); the 0.5\(\times\)-background “closure to 0.99–1.00” was recorded from one seed (deprecated by R074); and RT-75's brief misquoted T233. 2026-10-20T02:22Z PI redirect (08 §6): 「现在我不关心星表。还是要结果emldetect的结果。」the catalogue empirical-interval direction is withdrawn and the main line is the emldetect result; task 0b (codexnim) and the astra expert were dispatched; astra was triaged (verdict goals not reached); 0b was completed in the evening (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 than L0, so the L0 low-count excess is not an emldetect property. Later the same day, R081–R085: the count dependence survives on a common mask but the \(\hat{S}\)-about-5 number was a mask artefact; the L0 low-DET_ML centre excess is a property of the L0 pipeline itself with its cause open (R087 withdraws the 72/28 attribution); and the penalty-mismatch mechanism does explain the L1 width too (R086 reverses R083: with pooled estimators the L1 mismatch is +1.19 % [+1.03, +1.35], explaining 0.89 [0.84, 0.93] of the within-cell width deficit, predicted std z 0.919 against the observed 0.918) — one \(V_{3p}\) mechanism carries both the L0 and the emldetect widths.Afternoon (R092–R104): the main line moved from mechanism to law; four candidate laws were tested in turn, the scoring itself was wrong twice and fixed, and one validated scorer (NULL-1, R103: uniform p on exactly-true data) gives the single verdict — every candidate is rejected: on M2 band 4 C 10/72, D 48/72, DQ 42/72 and MIX3d 35/72, and on band 1 the D refit 37/72 (against a null mean of about 62–63; D's p = 0.0015); R094 gives the noise floor (a perfect law passes only about 61/72 on average) and R100 withdraws R096's "globally rejected at p = 0" (a null scale error). 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; the 12/12 anchor hinges on position 14's split). CENTRE-POS (finished, R105): position is not the missing centre variable (a smooth position term pushes the held-out failures from 24 to 58–72; a per-cell centre intercept 24 \(\to\) 11). The PI approved a symbolic-regression trial (PySR + SRHarness, NIM backend) and a hybrid-law task.
完整四段式汇报见The full four-part report is on the 组会页Group Meetings page。.
组会 · 2026-10-18Group Meeting · 2026-10-18
成果：宽度主线。上午的结论几乎全部被下午的记录改写。\(\sigma_{ratio}\) 的 L0 近行"18/19 格通过"被标为 Questioned（R054）：换成拟合模板族的精确位置导数后它在 DET_ML 6–10 过度修正（std z 1.67 / 1.83，两臂），RT-54 用 5 点模板（1.31）、h = 2（0.81）与精确导数（1.68）证明原先的通过只是 np.gradient 3 点差分的偶然抵消（R057）——它是一个经验校准，不是推导出来的宽度。T219 的"真实 PSF 宽度对"被重新限定（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），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；位置自由时（T234）残差均值等于解析 \(\delta_{3p}\)，宽度在 DET_ML 8 / 15 偏窄（std 0.914 / 0.955）、30 以上持平——但 R060 已把"三处共同偏窄"这条综合撤回：T234 的是固定 \(S_{true}\) 格子，用 \(V(S_{true})\) 归一后 std 是 1.00–1.03，偏窄只在逐行 \(V(\hat{S})\) 下出现，且三处的行集不同。玩具低计数偏窄被改写成可观测律（R059，RT-55 分诊）：按可观测 \(\hat{S}\) 分箱后，核 r̂ \(\leq\) 4 px 的 std z 在 DET_ML 1–3 是 0.82–0.86、\(\approx\) 5 是 0.93–0.95、\(\approx\) 7.5 是 0.94–0.96、\(\approx\) 12 及以上回到 1（1 SE 内）；"窄 2–4 % 于 DET_ML 8–14"是固定 \(S_{true}\) 格子的效应。W5 因此关闭（判据未达成），RT-52 至 RT-55 分诊完毕，RT-56（低计数偏窄的尺寸预测）与 RT-57（可观测箱里的振幅律）发出。工作：关闭 W5、RT-52 至 RT-55；发出 RT-56、RT-57（其后两者已分诊为 R060、R061，并新发 RT-58、RT-59）；第 23 阶段经审阅后部署为 935d5f5f，第 24–26 阶段的 16:00Z 刷新清单经审阅后部署为 448804a5；第 27–28 阶段修评审 T12_018 的清单并折入 R060/R061，部署为 663838ca（数据截至 17:31 UTC）。17:31Z 之后：前两支已分诊（R062、R063），另两支新发出。问题：调度方自己的时间戳失误（04:51 / 04:55 UTC 应为 11:49 / 11:51，11:48Z 会话日志更正）；调度方自己的 IRON-RULE 失误（T219 用 \(V(S_{true})\) 作分母）；红队的双线性模板把插值尖点读成物理（RT-52 O4，E017 第二次）；第 23 阶段物理页 §13 的"位置误差被压小"机制不成立（R055），本阶段改写。T12 的写网页会话两次在工具调用中途停下（第 24、25 阶段），第 26 阶段的检查点还把一次 check_math 的 FAIL 写成了 PASS（评审 T12_017 因此带横幅通过）。待 PI：无（08 §6 无新增待决项）。Results: the width main line. Almost every morning conclusion was rewritten by an afternoon record. \(\sigma_{ratio}\)'s L0 near-row "18 of 19 cells pass" is marked Questioned (R054): with the fitter template family's exact position derivatives it over-corrects at DET_ML 6–10 (std z 1.67 / 1.83, two arms), and RT-54 showed with the 5-point stencil (1.31), h = 2 (0.81) and the exact derivatives (1.68) that the old pass was only an accidental cancellation inside np.gradient's 3-point stencil (R057) — it is an empirical calibration, not a derived width. T219's "real-PSF width is right" is rescoped (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) and it is the recorded script's division by \(V(S_{true})\) (a label; the dispatcher's own IRON-RULE slip) that failed; with the position free (T234) the residual mean equals the analytic \(\delta_{3p}\) and the width is narrow at DET_ML 8 / 15 (std 0.914 / 0.955) and level at 30 and above — but R060 withdraws the "narrow in all three places" synthesis: T234's are fixed-\(S_{true}\) cells, normalised by \(V(S_{true})\) the std is 1.00–1.03, the narrowing appears only with the per-row \(V(\hat{S})\), and the three settings use different row sets. The toy low-count narrowing is rewritten as an observable law (R059, the RT-55 triage): binned by the observable \(\hat{S}\), the core r̂ \(\leq\) 4 px std z is 0.82–0.86 at DET_ML 1–3, 0.93–0.95 at \(\approx\) 5, 0.94–0.96 at \(\approx\) 7.5 and back to 1 (within 1 SE) at \(\approx\) 12 and above; "narrow by 2–4 % at DET_ML 8–14" was a fixed-\(S_{true}\) cell effect. W5 is therefore closed (the key is not met), RT-52 to RT-55 are triaged, and RT-56 (a size prediction for the low-count narrowing) and RT-57 (the amplitude law in observable bins) are launched. Work: W5 and RT-52 to RT-55 closed; RT-56 and RT-57 launched (both since triaged as R060 and R061, with RT-58 and RT-59 launched after them); the stage-23 refresh, reviewed, deployed as 935d5f5f, and the stage 24–26 16:00Z refresh list, reviewed, deployed as 448804a5; stages 27–28 fixed the T12_018 review list and folded R060/R061, deployed as 663838ca (data cutoff 17:31 UTC). After 17:31Z: the first two are triaged (R062, R063) and two more are launched. Problems: the dispatcher's own timestamp slip (04:51 / 04:55 UTC should be 11:49 / 11:51, corrected in the 11:48Z session log); the dispatcher's own IRON-RULE slip (T219 dividing by \(V(S_{true})\)); the red team's bilinear template reading an interpolation cusp as physics (RT-52 O4, E017 for the second time); stage 23's physics-page §13 mechanism "the position error is squeezed" does not hold (R055) and is rewritten in this stage. The T12 writer sessions stopped mid tool call twice (stages 24 and 25), and stage 26's checkpoint called a check_math FAIL a PASS (which is why the T12_017 review passed it with a banner). For the PI: none (08 §6 has no new pending item).
完整四段式汇报见The full four-part report is on the 组会页Group Meetings page。.
组会 · 2026-10-17Group Meeting · 2026-10-17
成果：PI 把项目重心转到模拟宽度（新问题 Q7）。当天三条宽度任务都交付并评审：W1 玩具理论（T229，Provisional）——位置固定的 Fisher 宽度在 DET_ML \(\gtrsim\) 20 与低 b 正确，DET_ML 6–20 且 b \(\geq\) 0.02 失效（1.17–2.49），条件均值偏低 +0.8…+1.7 \(\Delta\)C 无公式通过（当天深夜 RT-43 定位到宽度失效格是 codexnim 臂拟合器走到 \(\pm 8\) px 截断的伪迹，项目拟合器在三个争议格给 1.008 / 0.987 / 0.989，本阶段又用项目拟合器复跑得 std z 1.008 / 1.010 / 0.963，第三格差 −0.037 未达预注册判据，调度方据此废弃记录的失效并把 (15.32, 0.08) 这一格的宽度标为 open（偏窄不偏宽，N = 10,000 复跑入队）；该复跑给出 std z 0.975，记为窄侧偏离约 2.5 %（Provisional，R045）；均值失效仍成立，R042/R044/R045）；W2 宽度梯子（T228，Provisional）——L1 emldetect std z 0.95–1.02、mean z −0.05…−0.07（用 emldetect 实际得到的背景 BG_MAP_EML 打分；旧的 1.02–1.12 与 +0.28…+0.47 已由 R041 撤回），L0 的 1.45 / 1.52 是混合（远行 \(\Delta\)C 低于插入式）；SAS 22.1 对 20 重放（T227，Provisional）——emldetect 逐位一致，真背景的 L1/J 不需要版本注记。Q1（bonus）目标冻结 −9.282 % [−9.350, −9.219]，试点的"L2d/\(\mu\) 4.9 %"是代码伪迹（R035）。问题：W1 两臂结论相反，仲裁运行支持 codexnim 臂、Hermes 臂被拒（R037）；分诊自己在 T228 里引的红队一句话被 RT-41 撤回（R039）；M31 的"泊松散布不能解释 S90 超出"撤回（约定相关，R031）；RT-35 空回复故障（F027）、RT-41 根会话 60 s 超时。待 PI：M31 图的 \(\Delta C_A\) 约定（\(\lambda\) 还是 \(\nu+\lambda\)）。Results: the PI moved the project's priority to the simulation width (new question Q7). All three width tasks reported and were reviewed that day: W1 toy theory (T229, Provisional) — the position-fixed Fisher width is right at DET_ML \(\gtrsim\) 20 and at low b, fails at DET_ML 6–20 with b \(\geq\) 0.02 (1.17–2.49), and the conditional mean is low by +0.8…+1.7 \(\Delta\)C with no passing formula (late that day RT-43 traced the width-failure cells to the codexnim arm's fitter walking to its \(\pm 8\) px clip, and the project fitter gives 1.008 / 0.987 / 0.989 at the three disputed cells, and this stage re-ran them for std z 1.008 / 1.010 / 0.963; the third cell misses the pre-registered criterion by −0.037, so the dispatcher deprecates the recorded failure and marks the width at (15.32, 0.08) open (narrower, not wider; a N = 10,000 re-run is queued); that re-run gives std z 0.975, recorded as a narrow-side departure of about 2.5 % (Provisional, R045); the mean failure stands, R042/R044/R045); the W2 width ladder (T228, Provisional) — L1 emldetect std z 0.95–1.02 with mean z −0.05…−0.07, scored with the background emldetect was actually given (BG_MAP_EML; the old 1.02–1.12 and +0.28…+0.47 are withdrawn by R041), and L0's 1.45 / 1.52 is a mixture (the far rows' \(\Delta\)C sits below the plug-in); the SAS 22.1 vs 20 replay (T227, Provisional) — emldetect is bit-identical, so the truth-background L1/J needs no version caveat. Q1 (bonus): the target is frozen at −9.282 % [−9.350, −9.219] and the pilot's "L2d/\(\mu\) 4.9 %" was a code artefact (R035). Problems: W1's two arms disagreed and an arbiter run sided with the codexnim arm, rejecting the Hermes arm (R037); a red team's own side remark I had repeated in T228 was withdrawn by RT-41 (R039); the M31 claim "the Poisson spread does not explain the S90 excess" is withdrawn (convention-dependent, R031); RT-35 ended with an empty reply (F027) and RT-41's root session hit a 60 s terminal timeout. For the PI: the M31 map's \(\Delta C_A\) convention (\(\lambda\) or \(\nu+\lambda\)).
完整四段式汇报见The full four-part report is on the 组会页Group Meetings page。.
前几天：Earlier days: 10-21（规律阶段被专家复核改写、新打分手、PI 进度指标）10-21 (the law stage rewritten by an expert review; the new scorer; the PI progress metric) · 10-20 下午（规律阶段：四条候选全部被拒、打分手两次出错并修好）10-20 afternoon (the law stage: every candidate rejected, the scoring wrong twice and fixed) · 10-19（R064–R068 五支红队分诊、两次部署）10-19 (five red-team triages R064–R068, two deploys) · 10-18（W3/W4 收尾、W5 发出、R047–R052）10-18 (W3/W4 closed, W5 launched, R047–R052) · 10-16（附录基线登记、T13 第 6 阶段评审、RT-23 分诊）10-16 (appendix baseline registered; T13 stage 6 reviewed; RT-23 triaged) · 10-15（T13 第 2 阶段"混合"、红队首轮跑完）10-15 (T13 stage 2 "mixed"; the first red-team round done) · 10-14 / 10-1310-14 / 10-13
数据截至：2026-10-24 05:30 UTC（07_REVISION_LOG.md R047–R229、红队分诊 RT-113–RT-141、Tier 1 第一臂已发射：T1-e 留出能段块打分完成——M2 能段 1 中心确认 0.727、宽度旗标 AMBIGUOUS（真实约 2 % 单侧超出、箱边敏感）、锚 8/12；T1-c 背景通道定律不普适 +0.174 z、候选 C2 暂定；代价锚改 0.97 CPU-s/行（区间 0.86–1.3）；T1-a 已发射 M2-only N=115；星表律断幂律 5/6、eta 振幅修合并宽度但逐格形状失败；08_RUNNING.md、T12 第 48 阶段）Data as of: 2026-10-24 05:30 UTC (07_REVISION_LOG.md R047–R229, red-team triages RT-113–RT-141, Tier 1's first arm launched: the T1-e held-out band-block scoring is done — M2 band 1's centre CONFIRMED at 0.727, the width flag AMBIGUOUS (a real one-sided excess of about 2 %, edge-sensitive), the anchor 8/12; the T1-c background channel shows the law not universal, +0.174 z, candidate C2 provisional; the cost anchor corrected to 0.97 CPU-s/row, band 0.86–1.3; T1-a launched M2-only N=115; the catalogue law's broken power law 5/6, the eta amplitude fixing the pooled width while the per-cell shape fails; 08_RUNNING.md, T12 stage 48)
lib/detml.py 在 DET_ML\(\gtrsim\)745 时返回 inf，不可用于高 DET_ML（07 R022；稳定模块相对误差 \(\lesssim\)5e-16 到 DET_ML 1e7）。（丁）重注入对目录 big−control 的过再现不要再说"1.2–1.5\(\times\)"：逐格比值是中位数 M2 A 类 1.54 [1.18, 1.67]、M1 1.09 [0.88, 1.15]，D run 0.77 / 0.60（07 R025）。（戊）\(\sqrt{V}\) 比给定 \(\hat{S}\) 的条件宽度窄 22–30 % 这个读数是按 Ŝ 分位数分箱造成的假象，已撤回：同一玩具逐行 z 的 std 是 0.995–1.009（S 20–320；07 R030）。注意其适用范围已收窄为"位置固定、S \(\gtrsim\)12"（DET_ML 6–20 的尾与 \(V(\hat{S}) \to 0\) 的行仍未校准，07 R033）。（己）附录公式 \(\Delta C_A\) 与核函数 \(\mu_{3p}\) 在机器精度上恒等（8.6e-16），"A1 比 T3_joint_2D 差"在 C-all 显著、在 C-iso 落在噪声内（配对差 −0.0178 [−0.0437, +0.0082]）——C-iso 应写持平，不要说"更差"；宽度上附录显著更好（+0.052 [0.047, 0.057]）（07 R029）。续补 R034–R039：（庚）SAS 22.1 对 20 的"无需版本注记"只限于真背景生产（L1 单机 emldetect、J 臂 A）；J 臂 B（记录值 −7.43 %）与 T13 链走 esplinemap 5.4\(\to\)5.4.1，每相机波段 M1/PN b1–b4 的中位有符号 \(\Delta\)DET_ML 为 0.06–0.24 %（该波段含义模糊），引用这些数字要带 esplinemap 注记；确切的 L1 生产配置从未作为一次运行重放（R038）。（辛）W2 的"look-elsewhere 抬高 \(\Delta\)C"读法已撤回（R036）：远偏移行的 \(\Delta\)C 在低于插入式一侧（mean z −1.70 / −4.03），噪声峰只会抬高 \(\Delta\)C，符号相反；24.8 % 的远行比例来自按观测 DET_ML 分箱（按典型 DET_ML 是 40.8 % / 8.3 %）。近处行 0.89 / 0.96 不得当作给定 \((\hat{S}, b)\) 的宽度引用——它按拟合结果条件化（corr(r_hat, z) \(\approx\) −0.7），R039 明确拒绝。（壬）"该超出在 recert13 试点设计里不存在"已撤回（R039）：recert13 按典型 DET_ML 分箱给出 std 1.34 / 1.29、远行 28 % / 4.7 %（RT-40 的 1.015 / 0.997 是观测 DET_ML 分箱）；匹配 S_true 下远行 mean z \(\approx\) −2.1 两种设计一致，变的是远行权重（匹配 \((\hat{S}, b)\) 下最多差 2 倍）。W1 的 Hermes 臂"位置自由时宽度处处失效"被仲裁运行拒绝（R037：项目拟合器给出 1.024 / 0.993）。（癸）ecut = 0.68 的截断机制保留（High），但"0.802 对 0.833 的吻合"撤回（07 R040，RT-44 分诊：2 material / 3 minor）：0.802 是固定 S 在 (S 50, b 0.2) 的值，在审计所用的 b 0.0276 处它是 0.609；实测的 0.833 是无条件式宽度，蒙特卡洛在同样的 (S, b) 上给 0.821；该比值随 (S, b) 跨 0.49–0.91，PN 对高斯在匹配的 (S, b) 上是 0.58 对 0.37，盒子无关。引用任何截断比值必须同时给出 (S, b, PSF, 掩膜规则, 条件化方式)。（癸二）W2 的 L1 数字已替换（07 R041，RT-45 分诊：5 material / 1 minor）：W2 原来用设计水平 b_true 给 L1 打分，得到 std z 1.02 / 1.05 / 1.09 / 1.12、mean z +0.28…+0.47；但 emldetect 实际得到的背景是 BG_MAP_EML（注入水平 \(\times\) 位置曝光比，PITFALLS/E094），改用它打分后 mean z −0.05…−0.07、std 0.95 / 0.975 / 1.00 / 1.02（DET_ML 6–10 / 10–20 / 20–50 / 50–100），环结构与格间散布（\(\leq\) 0.03）都消失，通量标度闭合到 0.2 %。旧读数与"均值偏移来自 E153 的 PSF/通量标度差""背景图是均值效应而非宽度效应"两句都已撤回。教训：给 emldetect 的行算插入式，必须用 emldetect 实际得到的背景，不是设计水平。（癸三）W1 的"宽度在 DET_ML 6–20 且 b \(\geq\) 0.02 失效（std z 1.17–2.49）"已废弃（07 R042，RT-43 分诊）：失效格来自 codexnim 臂拟合器 widthlib.fit_3p 走到 \(\pm 8\) px 位置截断（0.3–2 % 的重复；一个被追踪的重复 \(\Delta C\) 7.29 对 31.78）。项目 Nelder–Mead 拟合器在三个争议格给 std z 1.008 / 0.987 / 0.989；本阶段又复跑一次（N = 2000，种子 777）得 1.008 / 1.010 / 0.963，其中 (6.63, 0.02) 与 (10.8, 0.08) 两格满足 R042 预注册的 |std z − 1| < 0.03，而 (15.32, 0.08) 的 0.963 差 −0.037，不满足该判据（R044 因此把该格留作 open）。现在的读法是：位置自由时插入式宽度在玩具的受测格都没有上尾失效，只有条件均值失效（+0.6…+1.7 \(\Delta C\)，三种拟合器一致）。（癸四）R044 留的那个 open 已由 N = 10,000 复跑关闭（07 R045）：在 (15.32, 0.08) 项目拟合器给 std z 0.975（种子 20261018，SE \(\approx\) 0.007，约 3.6\(\sigma\) 低于 1），按 R044 预注册规则记为窄侧偏离约 2.5 %（Provisional）——插入式在 DET_ML \(\approx\) 15 处偏窄，方向上与 L1 在 6–10 的 0.95 一致，不是上尾失效。Reviewer correction (Claude, 2026-10-17 05:55 UTC, extended 10:34Z with R029–R032, and 15:41Z with R034–R039, and 20:05Z with R040 and the RT-44 triage, and 22:50Z with R041/R042 and the RT-43/RT-45 triages, and 23:50Z with R045, and 23:51Z with R046): (a) All quantitative readings of the M31 sensitivity map (including "94 / 81 / 43 %") are withdrawn; status Questioned / Pending (07 R020, R023). "The Poisson spread does not explain the S90 excess" is withdrawn too — that conclusion flips with the map's \(\Delta C_A\) convention: under \(\Delta C_A = \nu+\lambda\) the smearing instead explains +42…+87 % of the excess, and the observed completeness sits closer to that convention (07 R031). (b) In T13 stage 7 the "neighbour-model removal" arm subtracted nothing (none of the 282 neighbour templates exist, 07 R024), so the neighbour effect was untested, not "\(\lesssim\) 1 %"; stage 7b redid it with real neighbour templates (282/282, \(\approx\) 91 % of the neighbour's counts removed): the big-shift flag fraction falls (M2 0.925 \(\to\) 0.725, M1 0.972 \(\to\) 0.778) but the amplitude loss is not recovered (big \(\div\) control \(S_A\) M2 +0.003, M1 −0.054), the rule's literal outcome stays ambiguous, the T13 track closes here, and the cause of the amplitude loss is not identified (07 R032). Catalogue-wide, rows with a near neighbour (DIST_NN < 40″) carry 0.24–0.42 of the \(\mu > 300\) relative deficit depending on tail capping (0.66 within observation; RT-32, R027) — a tail-weighted statistic, not "how much the neighbour causes"; the near-neighbour gradient itself is robust and a PSF-wing/background-contamination model is not excluded. (c) The stable DET_ML\(\to\Delta C\) inversion is verified (T225); lib/detml.py returns inf above DET_ML \(\approx\) 745 and must not be used for high DET_ML (07 R022; the stable module's relative error is \(\lesssim\) 5e-16 up to DET_ML 1e7). (d) Stop saying the re-injection over-reproduces the catalogue's big − control by "1.2–1.5\(\times\)": the per-cell ratios are medians of M2 A-like 1.54 [1.18, 1.67] and M1 1.09 [0.88, 1.15], with the D run 0.77 / 0.60 (07 R025). (e) the reading "\(\sqrt{V}\) is 22–30 % narrower than the conditional width given \(\hat{S}\)" was an artefact of binning in Ŝ quantiles and is withdrawn: in the same toy the per-row z has std 0.995–1.009 (S 20–320; 07 R030). Note its scope is now narrowed to "position fixed, S \(\gtrsim\) 12" (the DET_ML 6–20 tails and the rows with \(V(\hat{S}) \to 0\) are still uncalibrated; 07 R033). (f) The appendix formula \(\Delta C_A\) is identical to the kernel \(\mu_{3p}\) to machine precision (8.6e-16), and "worse than T3_joint_2D on A1" is significant on C-all but lies within noise on C-iso (paired difference −0.0178 [−0.0437, +0.0082]) — on C-iso write par, not "worse"; on width the appendix is significantly better (+0.052 [0.047, 0.057]) (07 R029). (g) SAS 22.1 vs 20's "no version caveat" holds for the truth-background production only (L1 standalone emldetect, J arm A); J arm B (recorded −7.43 %) and the T13 chain run esplinemap 5.4 \(\to\) 5.4.1, whose per-camera-band median signed \(\Delta\)DET_ML is 0.06–0.24 % for M1/PN bands b1–b4 (that band is ambiguous), so quoting those numbers carries an esplinemap caveat; the exact L1 production configuration was never replayed as one run (R038). (h) W2's "look-elsewhere raises \(\Delta\)C" reading is withdrawn (R036): the far-offset rows sit on the below-the-plug-in side (mean z −1.70 / −4.03), and a noise peak would only raise \(\Delta\)C, so the sign is the opposite; the 24.8 % far fraction came from binning by observed DET_ML (by typical DET_ML it is 40.8 % / 8.3 %). The near rows' 0.89 / 0.96 must not be quoted as the width given \((\hat{S}, b)\) — it is conditioned on a fit outcome (corr(r_hat, z) \(\approx\) −0.7), and R039 rejects it explicitly. (i) "The excess is absent in the recert13 pilot design" is withdrawn (R039): recert13 binned by typical DET_ML gives std 1.34 / 1.29 with far fractions 28 % / 4.7 % (RT-40's 1.015 / 0.997 came from observed-DET_ML binning); at matched S_true the far rows' mean z \(\approx\) −2.1 agrees in both designs, and what changes is the far-row weight (up to \(\times\)2 at matched \((\hat{S}, b)\)). The W1 Hermes arm's "the position-free width fails everywhere" was rejected by an arbiter run (R037: the project fitter gives 1.024 / 0.993). (j) The ecut = 0.68 truncation mechanism is kept (High), but the "0.802 against 0.833 agreement" is withdrawn (07 R040, RT-44's triage: 2 material / 3 minor): 0.802 is the fixed-S value at (S 50, b 0.2) and becomes 0.609 at the audit's b 0.0276; the measured 0.833 is an unconditional-style width that Monte Carlo reproduces as 0.821 at the same (S, b); the ratio spans 0.49–0.91 over (S, b) and is 0.58 against 0.37 for PN against Gaussian at matched (S, b), with the box irrelevant. Quote any truncation ratio only together with (S, b, PSF, mask rule, conditioning). (k) W2's L1 numbers are replaced (07 R041, RT-45's triage: 5 material / 1 minor): W2 scored L1 at the design level b_true and got std z 1.02 / 1.05 / 1.09 / 1.12 with mean z +0.28…+0.47; but the background emldetect was actually given is BG_MAP_EML (the injected level \(\times\) the position exposure ratio, PITFALLS/E094), and scoring with it gives mean z −0.05…−0.07 and std 0.95 / 0.975 / 1.00 / 1.02 (DET_ML 6–10 / 10–20 / 20–50 / 50–100), with the ring structure and the between-cell spread (\(\leq\) 0.03) gone and the flux scale closed to 0.2 %. The old readings and both "the mean offset comes from E153's PSF/flux-scale difference" and "the background map is a mean effect, not a width effect" are withdrawn. Lesson: to evaluate the plug-in on an emldetect row you must use the background emldetect was actually given, not the design level. (l) W1's "the width fails at DET_ML 6–20 with b \(\geq\) 0.02 (std z 1.17–2.49)" is Deprecated (07 R042, RT-43's triage): the failing cells come from the codexnim arm's fitter widthlib.fit_3p walking to its \(\pm 8\) px position clip (0.3–2 % of reps; on one traced rep \(\Delta C\) 7.29 against 31.78). The project Nelder–Mead fitter gives std z 1.008 / 0.987 / 0.989 at the three disputed cells, and this stage re-ran them (N = 2000, seed 777) for 1.008 / 1.010 / 0.963; the two cells (6.63, 0.02) and (10.8, 0.08) meet R042's pre-registered |std z − 1| < 0.03, while (15.32, 0.08) at 0.963 is off by −0.037 and does not meet it (which is why R044 left that cell open). The reading is now: with the position free the plug-in width shows no upper-tail failure at any tested toy cell; only the conditional mean fails (+0.6…+1.7 \(\Delta C\), consistent across three fitters). (m) The open item R044 left is closed by the N = 10,000 re-run (07 R045): at (15.32, 0.08) the project fitter gives std z 0.975 (seed 20261018, SE \(\approx\) 0.007, about 3.6\(\sigma\) below 1), recorded by R044's pre-registered rule as a narrow-side departure of about 2.5 % (Provisional) — the plug-in is too narrow at DET_ML \(\approx\) 15, the same direction as L1's 0.95 at 6–10, not an upper-tail failure.这个项目是什么What This Project Is
当前优先级（PI，2026-10-17 07:10Z）：Current priority (PI, 2026-10-17 07:10Z): 先把力气用在我们自己的模拟的宽度上（新问题 Q7）：零参数的插入式 \(\sqrt{V}\) 能不能描述给定 \(\hat{S}\) 的 \(\Delta C\) 宽度，在哪里失效。目录上的问题（Q1 均值亏损、Q2 高计数 m、Q3 目录宽度、Q5 M31 图）只作附加讨论——不新派目录任务，在跑的跑完轻量分诊，留着不解决也可以接受。 Put the effort into the width of our own simulations first (new question Q7): does the zero-parameter plug-in \(\sqrt{V}\) describe the \(\Delta C\) width given \(\hat{S}\), and where does it fail. The catalogue questions (Q1 mean deficit, Q2 high-count m, Q3 catalogue width, Q5 M31 map) are a bonus discussion — dispatch no new catalogue work, triage the running ones lightly, and leaving them unresolved is acceptable.
来源：cognition-os/02_CURRENT_QUESTION.md（优先级横幅）、04_DECISION_LOG.md 2026-10-17 07:10Z 条目、08_RUNNING.md §2。Sources: cognition-os/02_CURRENT_QUESTION.md (priority banner), 04_DECISION_LOG.md 2026-10-17 07:10Z entry, 08_RUNNING.md §2.
一句话：In one sentence: XMM-Newton 的源检测软件 emldetect 给每个候选源两个数——拟合流量 \(\hat{S}\) 和检测似然 DET_ML；目录只收 DET_ML \(\geq\) 6 的源，但 DET_ML 是个随机数。本项目要算出给定可观测量（\(\hat{S}\)、背景 b、位置）时，DET_ML 背后的统计量 \(\Delta C\) 服从什么分布，用零经验参数的公式，并在 4XMM DR14 真实目录上验证。有了它才能算对完备度、假源率和 Eddington 偏差修正。 XMM-Newton's source detection software emldetect outputs two numbers per candidate source — fitted flux \(\hat{S}\) and detection likelihood DET_ML; the catalogue only accepts sources with DET_ML \(\geq\) 6, but DET_ML is a random variable. This project derives the distribution of the statistic \(\Delta C\) behind DET_ML given the observables (\(\hat{S}\), background b, position), with zero empirical parameters, and validates it on the real 4XMM DR14 catalogue. Only with this distribution can one correctly compute completeness, false-source rates, and Eddington bias corrections. 为什么要做这件事 \(\to\)Why do this? \(\to\)
图的说明依据：lib/（模拟与拟合）、lib/fit_mask_3param.py（\(\mu_{3p}/V_{3p}\) 内核）、E140（c20 宽度）、E154/T11（联合拟合损失）、PARETO_BENCHMARK_V2.md §2（四层）。Figure provenance: lib/ (simulation & fitting), lib/fit_mask_3param.py (\(\mu_{3p}/V_{3p}\) kernel), E140 (c20 width), E154/T11 (joint-fit loss), PARETO_BENCHMARK_V2.md §2 (four layers).
目标表（多目标：A1–A4，越小越好；A4 参数越少越好）Objectives Table (multi-objective: A1–A4, smaller is better; for A4 fewer parameters is better)
| 轴Axis | 是什么What it is | 当前最好（样本外）Current best (out-of-sample) | 来源Source | 趋势Trend |
|---|---|---|---|---|
| A1 均值偏差A1 mean bias | 各箱 mean(z) 的加权 rms；0 = 均值全对Weighted rms of per-bin mean(z); 0 = mean perfectly matched | C-iso M2 b4：0.132（T3_joint_2D，0 参数）C-iso M2 b4: 0.132 (T3_joint_2D, 0 params) | benchmark/leaderboard_v2.json（噪声校正 \(A1_{nc}\) = 0.1321）benchmark/leaderboard_v2.json (noise-corrected \(A1_{nc}\) = 0.1321) | ↓ 从 v1 基线 0.597（V4c_no_eta）一路降；但见下方"窗口平坦性"未解决问题↓ down from v1 baseline 0.597 (V4c_no_eta); but see "window flatness" open problem below |
| A2 宽度偏差A2 width bias | 各箱 (sd(z)−1) 的加权 rms；0 = 宽度全对Weighted rms of per-bin (sd(z)−1); 0 = width perfectly matched | C-iso：0.077（T3_E147mix_jointW，0 参数）C-iso: 0.077 (T3_E147mix_jointW, 0 params) | benchmark/leaderboard_v2.json（\(A2_{nc}\) = 0.0766）benchmark/leaderboard_v2.json (\(A2_{nc}\) = 0.0766) | \(\to\) 尚未追平样本内 V5c 的 0.049（\(A2_{nc}\) = 0.0494）；是当前最大缺口\(\to\) not yet caught up with in-sample V5c's 0.049 (\(A2_{nc}\) = 0.0494); current largest gap |
| A3 过阈值校准A3 threshold calibration | L = 6/8/10 的预测概率减实测比例RMS of predicted P(DET_ML \(\geq\) L) minus observed fraction for L = 6/8/10 | C-iso：0.027（T3_joint_2D）C-iso: 0.027 (T3_joint_2D) | benchmark/leaderboard_v2.json（\(A3_{worst,nc}\) = 0.0266）benchmark/leaderboard_v2.json (\(A3_{worst,nc}\) = 0.0266) | 持平：配对差 −0.005，95% 区间 [−0.012, +0.002] 含 0（来源：results/v2/T11/t11_s14_pairs_C-iso.json）No significant difference: paired difference −0.005, 95% interval [−0.012, +0.002] contains 0 (Source: results/v2/T11/t11_s14_pairs_C-iso.json) |
| A4 参数个数A4 parameter count | 在数据上拟合出来的数的个数Number of parameters fitted on the data | 0（前沿上全部条目）0 (all entries on the front) | benchmark/entries_v2/benchmark/entries_v2/ | ＝ 目标已达成：前沿全为零参数= target met: front is all zero-parameter |
| 高计数不失效No failure at higher counts | 模型在更高计数时不应比无修正更差（用户原则，决策日志 12:45）A model must not be worse than no correction at higher counts (user principle, decision log 12:45) | 报告列已定（用户决定，决策日志 20:45）：m 和 a 作为高计数报告列，不进前沿。已有证据（T5_013/014/015）：全部 6 个 T3 条目（4 个零参数）从 S \(\approx\) 100 到 4000 保持同一 m（|\(\Delta\)m| < 0.01，C-all 上 −0.006 到 +0.0002，C-iso 上 +0.001 到 +0.004）；样本内 V5c_eta 在 S > 300 比不修正更差——|frac| 差在两个层上、所有区间方法下都不含 0；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]，BCa [+0.31, +0.78]（只有 68 个 bin，分布偏斜，T5_015）。RT-02 分诊更正：报告里的 ">300" 窗口按分箱的典型 DET_ML 选取（DET_ML > 300），不是 S > 300——C-all 的 2207 行中只有 467 行 S > 300；结论不变，V5c_eta 的 |frac| 超出随 S 增大（S\(\leq\)300 分箱 +0.023，S\(\geq\)300 分箱 +0.080）。试行 >300 标记（用户决定，决策日志 2026-10-15 09:50）：只有样本内 V5c_eta 被标记——C-all 只在 BCa 下成立，C-iso 在所有方法下成立；6 个 T3 条目（4 个零参数）都显著优于不修正。只作报告，不进入前沿。已发布的报告页：v2/highcount.htmlReport columns decided (user decision, decision log 20:45): m and a as high-count report columns, not entering the front. Evidence now (T5_013/014/015): all 6 T3 entries (4 zero-parameter) keep the same m from S \(\approx\) 100 to 4000 (|\(\Delta\)m| < 0.01, −0.006 to +0.0002 on C-all, +0.001 to +0.004 on C-iso); the in-sample V5c_eta is worse than no correction above S = 300 — its |frac| difference excludes 0 under every interval method on both layers; 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], BCa [+0.31, +0.78] (only 68 bins, skewed distribution, T5_015). RT-02 triage correction: the report's ">300" window is selected by the bin-typical DET_ML (DET_ML > 300), not S > 300 — only 467 of the 2207 C-all rows have S > 300; the verdict is unchanged, V5c_eta's |frac| excess grows with S (+0.023 in bins with S \(\leq\) 300, +0.080 in bins with S \(\geq\) 300). Trial >300 flag (user decision, decision log 2026-10-15 09:50): only the in-sample V5c_eta is flagged — on C-all under BCa only, on C-iso under every method; the 6 T3 entries (4 zero-parameter) are all significantly better than no correction. Report only, not used for the front. Published report page: v2/highcount.html |
注意：A1 的两种口径不要混用——本表用基准计分器的噪声校正 A1（leaderboard_v2.json）；逐箱加权 A1（T11 分箱表）数值不同，见Note: the two A1 estimators must not be mixed — this table uses the scorer's noise-corrected A1 (leaderboard_v2.json); the per-bin count-weighted A1 (T11 bin table) gives different values, see 现状页status page。.
未解决问题表Open Problems Table
| 问题Problem | 为什么重要Why it matters | 当前最佳假设Best hypothesis | 下一个检验Next test | 负责Owner | 状态Status |
|---|---|---|---|---|---|
| Q7（当前优先级）零参数的插入式 \(\sqrt{V}\) 能不能描述我们自己的模拟里给定 \(\hat{S}\) 的 \(\Delta C\) 宽度（玩具、L0 我们的拟合器、L1 emldetect、J 联合），在哪里、为什么失效？Q7 (current priority) Does the zero-parameter plug-in \(\sqrt{V}\) describe the \(\Delta C\) width given \(\hat{S}\) in our own simulations (toy, L0 our fitter, L1 emldetect, J joint), and where and why does it fail? | 宽度是 \(\Delta C\) 分布的另一半（均值已经有了零参数形式）。模拟是我们唯一能完全控制真值的地方：如果插入式在我们自己的拟合器里就描述不了给定 \(\hat{S}\) 的宽度，那目录上的一切宽度结论都建在沙地上。PI 2026-10-17 07:10Z 把它定为项目主攻方向，目录问题降为附加讨论。Width is the other half of the \(\Delta C\) distribution (the mean already has a zero-parameter form). Simulation is the only place we control the truth completely: if the plug-in cannot describe the width given \(\hat{S}\) in our own fitter, every catalogue width conclusion rests on sand. The PI made this the project's main line on 2026-10-17 07:10Z and demoted the catalogue questions to a bonus discussion. | 玩具模拟（15\(\times\)15 高斯 PSF，\(\sigma\)=1.8 px，b=0.02，仅振幅，逐行条件化）：\(z=(\Delta C-\mu(\hat{S}))/\sqrt{V(\hat{S})}\) 的 std 0.995–1.009、mean −0.006…+0.010（S 20–320，T226）——但那对应 DET_ML \(\approx\) 45–690；RT-34 的"窄 22–30 %"是 \(\hat{S}\) 分位数分箱的假象，已撤回（R030）。R033 收窄范围，R053 改写（其固定 \(S_{true}\) 数字已被 R061 取代）：位置固定、b \(\geq\) 0.02 时全量原始矩有限且校准（有界拟合下每个 b 匹配格 mean z 在 \(\pm\) 0.016 内；原先的"矩失效"是玩具牛顿拟合器在近空图像上发散，有界拟合消除每一行灾难行）；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；位置自由时宽度不变但实现均值偏置 +1.4 \(\Delta C\)（W1 codexnim 臂）；真实 PSF 的 M31 小模拟 sd/\(\sqrt{V}\)=1.023 是由 R056 重新限定（仍 Provisional）的记录：R044 只涉及 W1 玩具，T219 那次除以的是 \(V(S_{true})\)（一个标签），改成逐行 \(V(\hat{A})\) 后位置固定在 DET_ML 8–600 通过（std 0.995–1.007，N = 20,000）。R061 把 T226 改写成可观测律：按 \(\hat{S}\) 分箱后 b \(\geq\) 0.02 的 std z 在 1 的 2–3 % 内、mean z 在 \(\pm\) 0.09 内；b = 0.005 在 \(\hat{S} \lesssim\) 3 是离散格子。仍未测：低计数/尾部、估计背景、支撑截断、emldetect、联合拟合。规律阶段（M2 能段 4 与 1，2026-10-20 至 21）：四条候选规律（C、D、DQ、MIX3d）在旧打分手上分别通过 10 / 48 / 42 / 35 格（共 72），能段 1 的 D 重拟合 37/72；旧打分手被专家复核认定混用估计对象（R114）——宽度是位置内的，均值/覆盖率是合并的，零假设还跟着候选自适应——所以"宽度可以推广"只能读作位置内（R117：宽度指标对逐位置偏移恒等不变，1e-15；一条带 0.30 z 位置偏移的规律仍通过宽度 70/72）。新打分手（SCORER-CONTRACT v2，R118）把两个目标分开：未见位置总体上 D / H2 / MIX3d / C 通过 44 / 50 / 32 / 12 格（共 72，能段 4）与 36 / 43（共 72，能段 1，D / H2），全部被 Q 拒（p \(\leq\) 0.009）；R124 限定：这是一次不稳定的抽样——14 个位置分半上 D 22–46、H2 26–52（能段 4），能段 1 的 D 14–53、H2 30–52；H2 优于 D 在固定分半上不显著（McNemar p 0.18 / 0.25），所以不能作"H2 最好"的排名，任何规律排名之前先要 scorer v3；逐位置上每条规律都接近 0/72（1 / 0 / 1 / 0 与 0 / 1），冻结位置效应 \(\tau_{mean}\) 0.14 z（能段 4）/ 0.18 z（能段 1），重尾 df \(\approx\) 3。宽度读数被限定（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。open2026-10-22（R125–R156）：“每个条目都在 \(\varepsilon\) 内”的判据被证明噪声受限（完美规律在能段 4 的天花板只有 0.21–0.27），于是改成每个位置一个合并偏移，PI 定下容差 \(\varepsilon\) 0.05 比较规律、0.10 跟踪进度，打分手随之冻结为 score_pooled_v1.py；旧读数“规律就压在位置盲地板上”被地板改正推翻，改正后最好候选在 0.10 处高于地板；位置效应是像素级几何且两条能段共享（r = +0.936），导出的相位项被驳回，“像素中心把偏移放大 2\(\times\)”的读法已撤回（打分器伪迹，同一格掩膜下比值 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），但尚未在新位置检验；宽度规则本身也被校准（一条逐行精确的规律通过“8 箱全过”只有 66.8 %，PI 于是批准“8 箱允许 1 箱失败”），在这条判据下第一条可观测箱级规律 \(\bar{V}(\hat{S}, b)\) 通过（1/8），但带宽在测试半上选定，且同一构造迁到 emldetect 行后在全统计量下仍失败 5/5 箱。2026-10-22/23（R157–R174）：逐箱宽度检验必须用聚类标准误（(pid, S_true) 设计的 deff 1.26–1.8）——V̄ 的失败从 7/9 降到 5/9，玩具从 1/8 降到 0/8，"匹配功率 5 箱只失败 1 箱"是种子运气（P(\(\leq\) 1) = 0.067）；"窄核靠误差相消通过宽度"是粗锚假象，细锚下排序每个分半都反转（2.03 % 对 4–5 %）；细锚规律在 emldetect M2 能段 4 上失败 2/9、RMS 0.045，等于完美 oracle 的下限，约 2 % 的 RMS 带聚类区间 [1, 7] %；但"8 箱允许 1 箱失败"按 8 箱定标，emldetect 窗口带 9–10 箱、完美 oracle 也失败 2/9，字面规则在那里够不到——重新定标给出 \(k^*\) 2–3（或更严的 0–1），PI 必须选零假设；低 b 高 Ŝ 的硬格（b = 0.0125）偏差指向少数低 MASKFRAC_EML 位置的探测器掩膜截断（丢 1–2 个离群位置后 std(z) 0.856 \(\to\) 0.960，挑选是事后的）；位置项在 27 个新位置上 AMBIGUOUS（+0.148 [−0.037, +0.333]），但 8 个位置源太挤（最近邻最小 15.9 px 对原始 45.0 px），不能读作对 SS 项的检验——任何新位置臂今后必须强制最近邻 \(\geq\) 45 px；B011 符号缺陷解释 0 % 的迁移失败，玩具 V̄ 面弃用，宽度工作改走细锚路线。open（T226 rescoped；07 R030、R033、R061、R114、R115、R117、R118、R128–R174）Toy simulation (15\(\times\)15 Gaussian PSF, \(\sigma\) = 1.8 px, b = 0.02, amplitude only, per-row conditioning): \(z=(\Delta C-\mu(\hat{S}))/\sqrt{V(\hat{S})}\) has std 0.995–1.009 and mean −0.006…+0.010 (S 20–320, T226) — but that is DET_ML \(\approx\) 45–690; RT-34's "22–30 % too narrow" was an artefact of \(\hat{S}\)-quantile binning and is withdrawn (R030). R033 narrows the scope and R053 rewrites it (its fixed-\(S_{true}\) numbers are superseded by R061): with the position fixed and b \(\geq\) 0.02 the full raw moments are finite and calibrated (with a bounded fit, mean z within \(\pm\) 0.016 at every b-matched cell; the old "moments fail" was the toy Newton fitter diverging on near-empty images, and a bounded fit removes every catastrophic row); at b = 0.005, DET_ML 6–10 is 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; with the position free the width is unchanged but the realised mean carries a +1.4 \(\Delta C\) bias (W1 codexnim arm); the real-PSF M31 mini-simulation's sd/\(\sqrt{V}\) = 1.023 is rescoped by R056 (still Provisional): R044 concerns only the W1 toy, and T219's run divided by \(V(S_{true})\) (a label), so with the per-row \(V(\hat{A})\) the position-fixed case passes at DET_ML 8–600 (std 0.995–1.007, N = 20,000). 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. Still untested: low counts/tails, estimated background, support truncation, emldetect, joint fit. The law stage (M2 bands 4 and 1, 2026-10-20 to 21): the four candidate laws (C, D, DQ, MIX3d) pass 10 / 48 / 42 / 35 of 72 cells under the old scorer, and the band-1 D refit 37/72; the old scorer is found by an expert review to mix estimands (R114) — the width is within-position, the mean and coverage are pooled, and the null adapts to the candidate — so "the width generalises" can only be read as within position (R117: the width metric is identically invariant to per-position offsets, 1e-15; a law carrying a 0.30 z position offset still passes width 70/72). The new scorer (SCORER-CONTRACT v2, R118) separates two targets: on the population of unseen positions D / H2 / MIX3d / C pass 44 / 50 / 32 / 12 of 72 cells (band 4) and 36 / 43 of 72 (band 1, D / H2), all rejected on Q (p \(\leq\) 0.009); R124 qualifies: these are one draw of an unstable statistic — over 14 position splits D 22–46 and H2 26–52 (band 4), band-1 D 14–53 and H2 30–52; H2 above D is not significant on the fixed split (McNemar p 0.18 / 0.25), so no "H2 is best" ranking can be made, and scorer v3 is needed before any law ranking; per position every law is near 0/72 (1 / 0 / 1 / 0 and 0 / 1), with the frozen position effect \(\tau_{mean}\) 0.14 z (band 4) / 0.18 z (band 1), heavy-tailed with df \(\approx\) 3. The width 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. 2026-10-22 (R125–R156): the “every entry within \(\varepsilon\)” rule was shown to be noise-limited (a perfect law's ceiling on band 4 is only 0.21–0.27), so the criterion became one pooled offset per position, with the PI's tolerances \(\varepsilon\) 0.05 to compare and 0.10 to track and the scorer frozen as score_pooled_v1.py; the old reading “the laws sit at the position-blind floor” is overturned by the floor correction, after which the best candidate sits above the floor at 0.10; the position effect is pixel-level geometry shared by both bands (r = +0.936), with the derived phase term rejected and the “pixel centres amplify the offsets 2\(\times\)” reading withdrawn as a scorer artefact (ratio 1.16 under the same cell mask); a first pre-fit position term (the fit-box background level) lifts the best candidate from 0.545 to 0.727 (p 0.003–0.004) but has not been tested on fresh positions; the width rule itself was calibrated (a row-exact law passes “all 8 bins” only 66.8 % of the time, so the PI approved “at most 1 failing bin of 8”), under which the first observable bin-level law \(\bar{V}(\hat{S}, b)\) passes (1 of 8) with its bandwidth chosen on the test half, and the same construction still fails 5/5 bins on emldetect rows at full statistics. 2026-10-22/23 (R157–R174): the per-bin width test must use a cluster SE (the (pid, S_true) design has deff 1.26–1.8) — V̄'s failures fall from 7/9 to 5/9 and the toy's from 1/8 to 0/8, with the matched-power "1 failing bin of 5" being seed luck (P(\(\leq\) 1) = 0.067); "the narrow kernel passes width by error cancellation" is a coarse-anchor artefact, with fine anchors reversing the ordering in every split (2.03 % against 4–5 %); the fine-anchor law fails 2/9 bins on emldetect M2 band 4 with RMS 0.045, equal to the perfect oracle's floor, its about-2 % RMS carrying a cluster interval [1, 7] %; but the "8 bins with 1 failing bin allowed" rule was calibrated for 8 bins, emldetect windows carry 9–10 bins and the perfect oracle itself fails 2/9, so the literal rule is unreachable there — the recalibration gives \(k^*\) 2–3 (or 0–1 under the stricter null) and the PI must pick the null; the low-b high-Ŝ hard cell's (b = 0.0125) miss points at detector-mask truncation at a few low-MASKFRAC_EML positions (dropping 1–2 outliers moves std(z) 0.856 \(\to\) 0.960, the picking being post hoc); the position term is AMBIGUOUS at 27 fresh positions (+0.148 [−0.037, +0.333]), but 8 of them are source-crowded (nearest-neighbour minimum 15.9 px against the originals' 45.0 px), so it cannot be read as a test of the SS term — any fresh-position arm must from now on enforce a nearest neighbour \(\geq\) 45 px; the B011 sign defect explains 0 % of the transfer failures, the toy V̄ surface is dropped, and width work proceeds on the fine-anchor route.open (T226 rescoped; 07 R030, R033, R061, R114, R115, R117, R118, R128–R174) | W1 玩具理论（已评审，T229/E170，Provisional）：位置固定的 Fisher 宽度 \(V_{3p}\) 在 DET_ML \(\gtrsim\) 20 与低 b 下正确（std z 0.98–1.00），在 DET_ML 6–20 且 b \(\geq\) 0.02 时失效（1.17–2.49，上尾重）；条件均值偏低 +0.8…+1.7 \(\Delta\)C（R047 撤回"无公式通过"：E088 \(\delta_0\) 在 6 格中 4 格把 mean z 关到 −0.002…+0.018，DET_ML \(\approx\) 8 的两格差 +0.06；T231）；截断支撑的矩在位置固定时全过（R048：位置自由时只有拟合位置模板接近，+0.10…+0.16 / 1.03–1.05，开放）。W2 宽度梯子（已评审，T228，Provisional，R041 换数）：L1 emldetect std z 0.95 / 0.975 / 1.00 / 1.02（DET_ML 6–10 / 10–20 / 20–50 / 50–100），mean z −0.05…−0.07（BG_MAP_EML 打分；b_true 打分下的 1.02–1.12 与 +0.28…+0.47 已撤回）；L0 的 1.45 / 1.52 是混合——偏离 > 5 px 的行 \(\Delta\)C 低于插入式（mean z −1.7 / −4.0），近处行 0.89 / 0.96 但那是按拟合结果条件化，不是给定 \((\hat{S}, b)\) 的宽度。SAS 22.1 对 20（已评审，T227/E169，Provisional）：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，esplinemap 臂（J 臂 B、T13 链）带一条。Q1（bonus，已评审，R035）：目标冻结 −9.282 % [−9.350, −9.219]；试点的"L2d/\(\mu\) 4.9 %"是代码伪迹（10.4 %）。W1 宽度失效已废弃（R044）：记录的"DET_ML 6–20 且 b \(\geq\) 0.02 失效（1.17–2.49，上尾重）"是 codexnim 臂拟合器走到 \(\pm 8\) px 位置截断的伪迹，无缺陷拟合器复现不出；该读数标为 Deprecated，(15.32, 0.08) 一格另记为窄侧偏离约 2.5 %（R045）。收尾：W3 与 W4 双臂已关闭（R052：无参数公式不达判据；玩具可观测核窄 2–4 %，T233；\(\sigma_{ratio}\) 被玩具质疑）；W5 已关闭（判据未达成，R055）；红队 RT-47 到 RT-55 已分诊。W1 toy theory (reviewed, T229/E170, Provisional): the position-fixed Fisher width \(V_{3p}\) is right at DET_ML \(\gtrsim\) 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 (R047 withdraws "no passing formula": E088 \(\delta_0\) closes mean z to −0.002…+0.018 at 4 of 6 cells, +0.06 at the two DET_ML \(\approx\) 8 cells; T231); the truncated-support moments pass with the position fixed (R048: with it free only the fitted-template plug-in is close, +0.10…+0.16 / 1.03–1.05, open). W2 width ladder (reviewed, T228, Provisional, numbers replaced by R041): L1 emldetect std z 0.95 / 0.975 / 1.00 / 1.02 (DET_ML 6–10 / 10–20 / 20–50 / 50–100), mean z −0.05…−0.07 (BG_MAP_EML scoring; the b_true-scored 1.02–1.12 and +0.28…+0.47 are withdrawn); L0's 1.45 / 1.52 is a mixture — rows that wandered > 5 px have \(\Delta\)C below the plug-in (mean z −1.7 / −4.0), and the near rows' 0.89 / 0.96 is conditioned on a fit outcome, not the width given \((\hat{S}, b)\). SAS 22.1 vs 20 (reviewed, T227/E169, Provisional): emldetect and eboxdetect are bit-identical, only esplinemap changes; the truth-background L1/J production needs no SAS-version caveat, while the esplinemap arms (J arm B, the T13 chain) carry one. Q1 (bonus, reviewed, R035): the target is frozen at −9.282 % [−9.350, −9.219]; the pilot's "L2d/\(\mu\) 4.9 %" was a code artefact (10.4 %). W1's width failure is Deprecated (R044): the recorded "fails at DET_ML 6–20 with b \(\geq\) 0.02 (1.17–2.49, heavy upper tail)" is an artefact of the codexnim arm's fitter walking to its \(\pm 8\) px position clip, which no defect-free fitter reproduces; the reading is marked Deprecated and the (15.32, 0.08) cell carries a narrow-side departure of about 2.5 % instead (R045). Closed: W3 and W4, both arms (R052: no parameter-free formula meets the key; the toy observable core is narrow by 2–4 %, T233; \(\sigma_{ratio}\) questioned by the toy); W5 is closed (the key is not met, R055); the red teams RT-47 through RT-55 are triaged. | W1、W2、SAS 重放（已评审）；W3 与 W4（双臂，已关闭）；W5（已关闭，R055）；红队 RT-47 到 RT-55 已分诊，RT-56 与 RT-57 也已分诊（R060、R061），RT-58 至 RT-83 也已分诊（R062–R085）；任务 0b 与 RT-84/RT-85/RT-86 均已分诊（R089、R091、R088/R090）；六支规律任务（LAW-1、LAW-2、LAW-3、LAW-4、LAW-M2B1、LAW-D-M2B1）与 NULL-1 评分器均已完成并被红队复核（R093–R104）；RT-87 到 RT-94 均已分诊或在跑；CENTRE-POS 与符号回归试点（PySR + SRHarness，NIM 后端）在跑；2026-10-22 起在跑：PHASE-DIAG、WIDTH-TRANSFER、UNIVERSALITY-DESIGN、POS-FRESH、RT-112、RT-113、RT-114；2026-10-23 起在跑：M1-ANCHOR、MASKFRAC-ANCHOR、CAT-PSFMU、POS-ISOLATED、RT-123、RT-124（名字，不引用数字）W1, W2 and the SAS replay (reviewed); W3 and W4 (two arms, closed); W5 (closed, R055); the red teams RT-47 through RT-55 are triaged, RT-56 and RT-57 are triaged too (R060, R061), RT-58 through RT-83 are triaged as well (R062–R085); task 0b and RT-84/RT-85/RT-86 are all triaged (R089, R091, R088/R090); the six law tasks (LAW-1, LAW-2, LAW-3, LAW-4, LAW-M2B1, LAW-D-M2B1) and the NULL-1 scorer are all finished and re-checked by red teams (R093–R104); RT-87 through RT-94 are triaged or running; CENTRE-POS and the symbolic-regression trial (PySR + SRHarness, NIM backend) are running; running since 2026-10-22: PHASE-DIAG, WIDTH-TRANSFER, UNIVERSALITY-DESIGN, POS-FRESH, RT-112, RT-113, RT-114; running since 2026-10-23: M1-ANCHOR, MASKFRAC-ANCHOR, CAT-PSFMU, POS-ISOLATED, RT-123, RT-124 (names, no numbers quoted) | 部分回答（主线已是单图 emldetect 条件律；四条候选规律被一支验证过的评分器全部拒收（R103），宽度一半可推广、中心一半不能（R104），较 v17↑）：玩具的条件均值在 DET_ML \(\gtrsim\) 14 已被 E088 \(\delta_0\) 关闭（R047）；宽度改写成可观测律（R059）——按 \(\hat{S}\) 分箱后核 r̂ \(\leq\) 4 px 的 std z 在 DET_ML 1–3 是 0.82–0.86、\(\approx\) 7.5 是 0.94–0.96、\(\approx\) 12 及以上回到 1，"窄 2–4 % 于 8–14"是固定 \(S_{true}\) 格子的效应；\(\hat{S}\) < 4 时 13–26 % 的行 r̂ \(\geq\) 4 px、全行宽度反而宽（1.4–1.9）。L1 从 20 到 100 通过、6–10 偏窄约 5–7 %（成因未明）；L0 近行偏窄是拟合器性质（0.87–0.90），\(\sigma_{ratio}\) 的通过被 R054/R057 判为离散化偶然抵消，不是推导出来的宽度。真实 PSF（T234，Provisional，R062 改写为可观测律）：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）；位置自由时按可观测 \(\hat{S}\) 分箱，插入式 \(V_{3p}(\hat{S})\) 高估宽度——std z 0.88 / 0.92 / 0.94 / 0.96 于 DET_ML \(\approx\) 7 / 9 / 12 / 18，约 25–40 才回到 1，低端深到 0.73（DET_ML \(\lesssim\) 2）；偏窄在 DET_ML \(\approx\) 17–22 以下都超过 2 SE，持续到目录区间，不像玩具（R059：DET_ML \(\approx\) 12 就消失）；Claude 的先验边界/斜率检验（S_true \(\geq\) 7 / \(\geq\) 8.5、斜率 −2.5 \(\to\) −1.5）每箱都在 1 SE 内不变，所以这是给定 \(\hat{S}\) 的律、不是先验截断；固定 \(S_{true}\) 格子的 0.914 / 0.955（DET_ML 8 / 15）只作历史。机制（R064）：公式本身在真实参数处是对的——固定 S_true 与真实位置时，残差的 std 除以 \(\sqrt{V_{3p}(S_{true})}\) 是 1.12 / 1.07 / 1.03 / 1.01 / 1.03（DET_ML 3 / 5 / 9 / 15 / 30，从不小于 1）；偏窄来自把凸的 \(V_{3p}\) 放在偏高的 \(\hat{S}\) 上计算（\(\hat{S}\) 在 DET_ML 3 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。R064 预言的修正已被推翻（R068）：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 从 0.88 / 0.92 / 0.94 / 0.96 升到 0.975 / 0.959 / 0.970 / 0.979（消掉 45–79 % 的偏窄），回不到 1 \(\pm\) 0.02；DET_ML \(\lesssim\) 1 时该校正没有定义。所以目前没有任何可观测的点修正能闭合真实 PSF 的宽度。跨位置（R065）：L0 生产数据 67 个位置（离轴 0–13′，我们的全局搜索拟合器）上这条律都成立，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）没有这条低端偏窄（R066）：按可观测 S_eml 分箱后 std z 0.94–0.98（DET_ML 1–6）、\(\approx\) 0.95（6–10）、12–20 以上约 1；最低一箱反而宽（1.31）。匹配比较（R067）：同 67 位置、同 S^−2.5 密度、同 (S, b) 格子权重下，L1 − L0 = +0.24 / +0.19 / +0.12 / +0.08 / +0.07 于 DET_ML \(\approx\) 2 / 3 / 4 / 5 / 7（约 50 处持平）——比原先记录的差别更大；但两边用的是不同流水线加不同打分（真实 PSF 模板族在拟合位置 对 e114 替代核在名义位置）与不同振幅估计量（S_true 2.5 时 \(E[\hat{S}]/E[S_{eml}]\) = 1.38，20 以上 0.92），所以这个差不归因于拟合器。开放：真实 PSF 宽度配料研究（08 §7 item 0：一次换一个配料——同 HPD 的高斯 PSF、平曝光、无掩膜、位置固定——看 2-SE 交叉点从 \(\approx\) 20 挪到哪儿）、RT-60 正在攻击这条新律（拟合器全局最优性、拟合模板处的插入式、第二个源位置）、B009 修正（lib/beyond_fisher 漏掉泊松 \(\kappa_4\)）、E051 均值侧复查。2026-10-22 起的四个开放问题：（一）位置项的新位置检验——第一个拟合前位置项把最好候选从 0.545 抬到 0.727（p 0.003–0.004），但它在同一批位置上拟合并打分，尚未在未见位置上检验（POS-FRESH 在跑）；（二）像素中心残余的解释——“放大 2 倍”已撤回（打分器伪迹），但同一格掩膜下两臂比值 1.16 是真实信号（\(\chi^2\) 312 于 12 个自由度），两臂还在盒表起点、注入取整像素与一次 PSF 存档错配上有别，臂间成分没有解释；（三）宽度规则的执行——PI 已批准“8 箱允许 1 箱失败”，但宽度检验的行自助标准误忽略了 \((\mathrm{pid}, S_{true})\) 聚类设计（设计中位效应 1.26），改用聚类标准误后玩具 1/8 变 0/8、“匹配统计量 1/5 通过”按种子运气撤回（200 个子样本下 P(\(\leq\) 1 失败) 0.067）；玩具 V̄ 曲面还带一个先验符号缺陷（三个导出面全错），但它解释迁移失败的 0 %，玩具 V̄ 作为迁移对象被放弃；emldetect 侧的全统计量失败（每箱约 4–7 % 的真实失配）成立，验收未达成；（四）普适性设计——PI 2026-10-22 的问题：什么多样性证明规律普适。Tier 0 已出结果：星表轴上冻结的宽度律在 \(\mu_0 \lesssim\) 120（C-all）/ 250（C-iso）域内通过（68 % 覆盖 0.69–0.70、95 % 覆盖 0.95–0.96），以上失败；能段/仪器轴上起初报的失败（M2 b1 0.606 对门槛 0.627）已撤回：那次打分把位置项逐行代入，而冻结打分器是从合并偏移里减去；按冻结口径、零新参数，M2 b1 为 0.727，过门槛；M1 b1–b4 与 M2-b4 只共享 9 个位置，分辨不了 0.10 的差距，不算失败；高计数锚点规则只有 4/12 格通过，是真实的部分失败。随后的重新打分（A0-RESCORE，复算逐位一致）给出全貌：M1 能段 1 为 8/21 = 0.381，配对差 −0.333 [−0.619, −0.048]，一个已分辨的中心失败；M1 能段 2/3/4 在 0.10 处分辨不了（MDS 约 0.30）；高计数锚在 M1 能段 4 0/12。Tier 0 判定每个块都失败，Tier 1 因此压住等 PI（选项 A/B/C，推荐 B：等 M1-ANCHOR 诊断）。另有一条必须连带的限制：冻结打分手的宽度区间欠覆盖（0.941 / 0.946 对 0.950 的接受带），任何已解决比例都要连这个区间一起引用Partly answered (the main line is now the single-image emldetect conditional law; one validated scorer rejects all four candidate laws (R103), the width half generalises within one position and the centre half does not (R104 plus the R114/R117 re-scoping), and the new scorer still accepts no candidate under either the unseen-position population or the per-position target (R118), ↑ vs v17): the toy conditional mean is closed at DET_ML \(\gtrsim\) 14 by E088 \(\delta_0\) (R047); the width is rewritten as an observable law (R059) — binned in \(\hat{S}\), the core r̂ \(\leq\) 4 px std z is 0.82–0.86 at DET_ML 1–3, 0.94–0.96 at \(\approx\) 7.5 and back to 1 at \(\approx\) 12 and above, and "narrow by 2–4 % at 8–14" was a fixed-\(S_{true}\) cell effect; at \(\hat{S}\) < 4 the r̂ \(\geq\) 4 px rows are 13–26 % and the all-row width is wide (1.4–1.9). L1 passes from 20 to 100 and is narrow by \(\approx\) 5–7 % at 6–10 (unexplained); L0's near-row narrowing is a property of the fitter (0.87–0.90), and \(\sigma_{ratio}\)'s pass is judged by R054/R057 an accidental discretisation cancellation, not a derived width. Real PSF (T234, Provisional, rewritten as an observable law by R062): with the position fixed the per-row \(V(\hat{A})\) passes at DET_ML 8–600 (std 0.995–1.007, N = 20,000); with it free, binned by the observable \(\hat{S}\), the plug-in \(V_{3p}(\hat{S})\) over-states the width — std z 0.88 / 0.92 / 0.94 / 0.96 at DET_ML \(\approx\) 7 / 9 / 12 / 18, back to 1 only by \(\approx\) 25–40, and as deep as 0.73 at the low end (DET_ML \(\lesssim\) 2); the deficit is above 2 SE through DET_ML \(\approx\) 17–22 and lasts into the catalogue regime, unlike the toy (R059: gone by DET_ML \(\approx\) 12); Claude's prior-edge and slope test (S_true \(\geq\) 7 / \(\geq\) 8.5, slope −2.5 \(\to\) −1.5) leaves every bin within 1 SE, so it is a law given \(\hat{S}\), not a prior truncation; the fixed-\(S_{true}\) cells' 0.914 / 0.955 (DET_ML 8 / 15) are history only. Mechanism (R064): the formula itself is right at the true parameters — at fixed S_true and the true position the residual's std divided by \(\sqrt{V_{3p}(S_{true})}\) is 1.12 / 1.07 / 1.03 / 1.01 / 1.03 (DET_ML 3 / 5 / 9 / 15 / 30, never below 1); the narrowing comes from evaluating the convex \(V_{3p}\) at a biased-high \(\hat{S}\) (\(\hat{S}\) is high by +14.5 % at DET_ML 3 and +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 correction R064 predicted is refuted (R068): evaluating \(V_{3p}\) at the observable Cox–Snell bias-corrected amplitude \(\hat{S} - b_S(\hat{S})\) lifts std z from 0.88 / 0.92 / 0.94 / 0.96 to 0.975 / 0.959 / 0.970 / 0.979 at DET_ML 6.9 / 9 / 12 / 18.5 (45–79 % of the deficit removed) but not to 1 \(\pm\) 0.02; below DET_ML \(\lesssim\) 1 the correction is undefined. So no observable point correction closes the real-PSF width yet. Across positions (R065): the law 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) does not show this low-end narrowing (R066): binned by the observable S_eml, std z is 0.94–0.98 at DET_ML 1–6, \(\approx\) 0.95 at 6–10 and about 1 above 12–20; the lowest bin is wide (1.31). Matched comparison (R067): at the same 67 positions, the same S^−2.5 density and the same (S, b) cell masses, L1 − L0 = +0.24 / +0.19 / +0.12 / +0.08 / +0.07 at DET_ML \(\approx\) 2 / 3 / 4 / 5 / 7 (level by about 50) — larger than recorded; but the two sides use different pipeline + scoring pairs (the stamp family at the fitted offset against the e114 surrogate at the nominal position) and different amplitude estimands (\(E[\hat{S}]/E[S_{eml}]\) = 1.38 at S_true 2.5 and 0.92 at 20 and above), so the gap is not attributable to the fitter. Open: the real-PSF width ingredient study (08 §7 item 0: swap one ingredient at a time — a Gaussian PSF of the same HPD, flat exposure, no mask, position fixed — and see where the 2-SE crossover moves from \(\approx\) 20); RT-60 is attacking this new law (fitter global optimality, the plug-in at the fitted template, a second source position); the B009 fix (lib/beyond_fisher omits the Poisson \(\kappa_4\)); an E051 mean-side re-check. Four open problems from 2026-10-22: (i) the fresh-position test of the position term — the first pre-fit position term lifts the best candidate from 0.545 to 0.727 (p 0.003–0.004), but it was fitted and scored on the same positions and has not been tested on unseen ones (POS-FRESH running); (ii) explaining the pixel-centre residual — the “2\(\times\) amplification” is withdrawn (a scorer artefact), but the two arms' ratio 1.16 under the same cell mask is real signal (\(\chi^2\) 312 on 12 degrees of freedom), with the arms also differing in boxlist start, injection rounding pixel and one PSF staging error, so the arm-to-arm component is unexplained; (iii) carrying out the width rule — the PI approved “at most 1 failing bin of 8”, but the width test's row-bootstrap SE ignores the \((\mathrm{pid}, S_{true})\) cluster design (median design effect 1.26): under a cluster SE the toy's 1/8 becomes 0/8 and the “matched-power pass (1 of 5)” is withdrawn as seed luck (P(\(\leq\) 1 fail) 0.067 over 200 subsamples); the toy V̄ surface also carries a prior-sign defect (all three exported surfaces), but it explains 0 % of the transfer failures and the toy V̄ is dropped as a transfer object; the emldetect full-statistics failure (a real ~4–7 % per-bin miscalibration) stands, so acceptance on emldetect is not met; (iv) the universality design — the PI's 2026-10-22 question: what diversity proves universality. Tier 0 has reported: on the catalogue axis the frozen width law passes within the \(\mu_0 \lesssim\) 120 (C-all) / 250 (C-iso) domain (68 % coverage 0.69–0.70, 95 % coverage 0.95–0.96) and fails above; the first band/instrument failure (M2 b1 0.606 against the bar 0.627) is withdrawn: that scoring applied the position term row by row, while the frozen scorer subtracts it from the pooled offset; under the frozen convention, with zero new parameters, M2 b1 scores 0.727 and passes; M1 b1–b4 share only 9 positions with M2-b4 and cannot resolve a 0.10 shortfall, so they are not failures; the high-count anchor rule passes only 4/12 cells, a real partial failure. The rescan (A0-RESCORE, replay bit-for-bit) then gave the full picture: M1 band 1 is 8/21 = 0.381, paired difference −0.333 [−0.619, −0.048], a resolved centre failure; M1 bands 2/3/4 cannot be resolved at 0.10 (MDS about 0.30); the high-count anchor is 0/12 on M1 band 4. Tier 0 fails on every block, so Tier 1 is held for the PI (options A/B/C, recommending B, waiting for the M1-ANCHOR diagnosis). One limit must travel with every such number: the frozen scorer's width interval under-covers (0.941 / 0.946 against a 0.950 acceptance band), so any solved fraction is quoted together with that interval |
| Q1Q1（bonus）(bonus)什么补上了目录 MOS 在 DET_ML 6–10 的均值亏损（按模拟自己的口径 −9.28 %，对 \(\mu\)+\(\delta\)，M2 b4 C-iso60）？What closes the catalogue's MOS mean deficit at DET_ML 6–10 (−9.28 % under the simulation's convention, against \(\mu\)+\(\delta\), M2 b4 C-iso60)? | 这是整个基准的主问题；补不上，所有修正都只是拟合。The main question of the whole benchmark; without it every correction is just a fit. | 联合拟合的 2-D 位置损失 \(L_{2d}/\mu\) \(\approx\) 9.5 % 是真的、量级也对，但单独闭合不了目录。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 emldetect 试点用 emldetect 的估计背景计分：A 臂 −5.9 %、B 臂 −7.4 %，差 3.39 % / 1.85 %——但试点与目录不匹配：试点 b \(\approx\) 0.028 恒定，目录 0.02\(\to\)0.14；两个 6–10 窗口因此落在不同 \(\hat{S}\)（试点 7.3–8.5，目录 9.5–13）；匹配 \(\hat{S}\) 后差 4–6 %；试点的亏损到 \(\hat{S}\) > 30 消失，目录仍停在 −7 %；试点 \(L_{2d}/\mu\) 10.4 % 对目录 8.3 %（早先的"5.0 % 对 9.8 %"是模板读取与 NaN 跳过的代码伪迹，已更正，07 R035）。open（闭合未证）（T210；07 R019、R035）The joint fit's 2-D position loss \(L_{2d}/\mu\) \(\approx\) 9.5 % is real and of the right order, but does not close the catalogue alone. 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 emldetect pilot, scored with emldetect's estimated background, reaches arm A −5.9 % and arm B −7.4 %, a gap of 3.39 % / 1.85 % — but the pilot is not matched to the catalogue: the pilot has b \(\approx\) 0.028 flat against the catalogue's 0.02\(\to\)0.14, so the two 6–10 windows sit at different \(\hat{S}\) (pilot 7.3–8.5, catalogue 9.5–13); at matched \(\hat{S}\) the gap is 4–6 %; the pilot's deficit vanishes by \(\hat{S}\) > 30 while the catalogue stays at −7 %; the pilot's \(L_{2d}/\mu\) is 10.4 % against the catalogue's 8.3 % (the earlier "5.0 % against 9.8 %" was a template-loading and NaN-skip code artefact, corrected, 07 R035). open (closure not shown) (T210; 07 R019, R035) | 按目录给定的 \(\hat{S}\) 采（或重加权）目录的 b，用估计背景在所有窗口按 \(\hat{S}\) 分箱给联合模拟打分，对用同一代码重算的目录（cat_sens.py 口径）；brief 待 Lang3/virgo 负载允许后写Run a joint simulation (or reweight the existing rows) at the catalogue's b given \(\hat{S}\), scored per \(\hat{S}\) bin over all windows with the estimated background, against the catalogue recomputed with the same code (the cat_sens.py convention); the brief waits for Lang3/virgo load | T11 / T13 / 红队已分诊T11 / T13 / red-team triaged | 开放（较 v11↓：闭合未证；目录改为 −9.28 %，试点不匹配）open (↓ vs v11: closure not shown; the catalogue moves to −9.28 % and the pilot is unmatched) |
| Q2Q2（bonus）(bonus)高计数的乘性亏损 m 从哪来（PN 0.024、M1 0.049、M2 0.050，向轴上增大；L0 底 0.0038）？What causes the high-count multiplicative deficit m (PN 0.024, M1 0.049, M2 0.050, rising toward on-axis; L0 floor 0.0038)? | m 是"星表及其处理链的性质"，不是我们的拟合器或核；机制不找到，修正就只能逐窗口拟合。m is "a property of the catalogue and its processing", not our fitter or kernel; without a mechanism the corrections stay per-window fits. | 不是拟合器/核（Trusted）；不是星表列能解释的（T11）；残差 = 核心 −3 %（M1）/ −4 %（M2）加 5 % 左尾（Provisional）；30 个重拟合观测里的大位移组（M2 40/242、M1 54/223）是近邻环境效应——点源重注入后 95–96 % 复现位移（对照 0–6 %），随最近邻距离从 0.75（< 20″）降到 0（> 100″）；只能比 big \(\div\) 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"三句已撤回（07 R014）。它占 m 的份额还没测Not the fitter or kernel (Trusted); not explained by catalogue columns (T11); the residual is a core at −3 % (M1) / −4 % (M2) plus a 5 % left tail (Provisional); in 30 refitted observations the big-shift group (M2 40/242, M1 54/223) is a near-neighbour environment effect — point sources re-injected there reproduce the shift in 95–96 % of cases (control 0–6 %), falling from 0.75 (< 20″) to 0 (> 100″) with neighbour distance; only big \(\div\) 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 (07 R014). Its share of m is not yet measured | 无（bonus，T13 线 2026-10-17 关闭）：真实邻居扣除（282/282 模板，约 91 % 邻居计数）降低大偏移标记（M2 0.925\(\to\)0.725、M1 0.972\(\to\)0.778）但不恢复振幅（big\(\div\)control \(S_A\) +0.003 / −0.054），规则字面结果仍 ambiguous，成因未定（07 R032）None (bonus, the T13 track closed 2026-10-17): real neighbour removal (282/282 templates, \(\approx\) 91 % of the neighbour's counts) lowers the big-shift flag (M2 0.925 \(\to\) 0.725, M1 0.972 \(\to\) 0.778) but does not recover the amplitude (big \(\div\) control \(S_A\) +0.003 / −0.054), the rule's literal outcome stays ambiguous, and the cause is not identified (07 R032) | T13 线已关闭（bonus；07 R032）：真实邻居扣除降低大偏移标记但不恢复振幅，成因未定T13 track closed (bonus; 07 R032): real neighbour removal lowers the big-shift flag but does not recover the amplitude; cause not identified | 开放（bonus；T13 线 2026-10-17 关闭，不再派 stage 7c）Open (bonus; the T13 track closed 2026-10-17, no stage 7c) |
| Q3Q3（bonus）(bonus)什么能描述目录 \(\Delta C\) 宽度跨窗口的变化？What describes the catalogue \(\Delta C\) width across windows? | 宽度错了，过阈值概率和完备度就算错，灵敏度图也不可信。Get the width wrong and threshold-crossing probabilities, completeness and the sensitivity map are all wrong. | (1) \(\sqrt{}\)V（零参数）在 6–10 的 C-iso 最好；(2) DET_ML 20 以上需要一个 \(s^{2}\mu^{2}\) 项（s \(\approx\) 0.08 C-all / 0.04 C-iso，已登记宽度律 Questioned）；(3) 逐相机 s = 0.12 / 0.11 / 0.095，留出 A2 \(\approx\) 0.12（作为拟合 Provisional）——形状是 MOS 的一个子群体而非对称散布，高斯尾薄了 38–64 倍(1) \(\sqrt{}\)V (zero parameters) is best at 6–10 on C-iso; (2) an \(s^{2}\mu^{2}\) term is needed above DET_ML 20 (s \(\approx\) 0.08 C-all / 0.04 C-iso; the registered laws Questioned); (3) per camera s = 0.12 / 0.11 / 0.095 with held-out A2 \(\approx\) 0.12 (Provisional as a fit) — the shape is a sub-population in MOS, not a symmetric scatter; the Gaussian tails are 38–64\(\times\) too thin | 混合拟合（\(m_{0}\), \(s_{0}\), f, \(\Delta\), 离轴）在 E165 行上对 P7 留出检验（08 §7.1）Mixture fit (\(m_{0}\), \(s_{0}\), f, \(\Delta\), off-axis) on E165 rows against P7, held out (08 §7.1) | T5 / 新拟合任务T5 / new fit job | 部分回答（较 v11↓：宽度律与双成分解释都被质疑）Partly answered (↓ vs v11: both the width laws and the two-component interpretation are questioned) |
| Q4Q4PI 的附录公式（在拟合计数处取值的 Asimov \(\Delta C_A\)）成立吗，为什么？Does the PI's appendix formula (Asimov \(\Delta C_A\) at the fitted counts) work, and why? | 它不靠我们的修正项却接近数据；理解它就能判断哪些修正其实是高阶项。It approaches the data without our corrections; understanding it tells us which corrections are really higher-order. | 成立，因为极大似然的得分方程让 \(\Delta C_{obs}\) − \(\Delta C_A\)(\(\hat{S}\)) 在一阶上为 0（\(\delta\) 与 V 是二阶）。2026-10-16 已正式登记（PI 批准）：appendix_eff 与 appendix_psfgen（参考）在 C-all 和 C-iso 上，零参数；6–10 前沿有变化：C-iso 样本外由 {T3_joint_2D, T3_E147mix_2D, T3_E147mix_jointW} 收缩为 {T3_joint_2D, appendix_eff}（两个条目被 appendix_eff 支配而退出），C-all 样本外 {T3_E147mix_2D, T3_joint_2D, appendix_eff, appendix_psfgen}，样本内仍 {V5c_eta}。A1 优势是省略的 \(\delta\)：6–10 为 −0.44，逐窗缩小到 >300 的 −0.06；宽度优势只在 DET_ML < 20（M2 b4，波段 1–4）。模板差异（中位 1.12，M2 b4 范围 1.06–1.25，波段 5 排除）是两个方向相反效应的残差——捕获核心更宽约 11 % 对核振幅约定约 12 %，不是"捕获通量"；"psfgen 图高估 \(\Delta C\) 约 10 %"撤回。M31 注入 \(\pm\)1–2 %It works because the maximum-likelihood score equations make \(\Delta C_{obs}\) − \(\Delta C_A\)(\(\hat{S}\)) vanish to first order (equal to 0; \(\delta\) and V are second order). Registered 2026-10-16 (PI approved): appendix_eff and appendix_psfgen (reference) on C-all and C-iso, zero parameters; the 6–10 front changed: on C-iso the out-of-sample front contracted from {T3_joint_2D, T3_E147mix_2D, T3_E147mix_jointW} to {T3_joint_2D, appendix_eff} (two entries dominated by appendix_eff and off the front), on C-all the out-of-sample front is {T3_E147mix_2D, T3_joint_2D, appendix_eff, appendix_psfgen}, and in-sample still holds {V5c_eta}. The A1 edge is the omitted \(\delta\): −0.44 at 6–10, shrinking to −0.06 at >300; the width win holds only below DET_ML 20 (M2 b4, bands 1–4). The template difference (median 1.12; 1.06–1.25 by position on M2 b4; band 5 excluded) is the residual of two opposite effects — a broader captured core \(\approx\) 11 % against the kernel amplitude convention \(\approx\) 12 % — not "captured flux"; "psfgen maps over-predict \(\Delta C\) by about 10 %" is withdrawn. M31 injections \(\pm\)1–2 % | 在其余 14 个配置上给 appendix_eff 打分（先波段 1–4），看 T216 是否超出 M2 b4 成立Score appendix_eff on the other 14 configurations (bands 1–4 first) to see whether T216 generalises beyond M2 b4 | RT-17 / RT-22 已分诊；已登记RT-17 / RT-22 triaged; registered | 部分回答（较 v11↑：已登记为正式条目；范围按 T216/T217 收窄）Partly answered (↑ vs v11: registered as formal entries; scoped per T216/T217) |
| Q5Q5（bonus）(bonus)PI 的 M31 灵敏度图与 edetect_stack 管线的真实完备度一致吗？Is the PI's M31 sensitivity map consistent with the real completeness of the edetect_stack pipeline? | 图是论文里完备度关系的来源；口径不一致会让结论整体偏移。The map is the source of the completeness relation in the paper; a convention mismatch shifts the whole conclusion. | 两侧都取图的逐像素 n（有曝光的图像数，1–8，众数 3，从不为 9）时：S50 观测/图 0.95–0.99（5 个能段），S90 1.11–1.20（波段 1–4）。RT-24 复核后这些数字不稳定：S50 随 n 约定跨 0.89–0.99、随分箱跨 0.80–0.99；S90 超出不能由泊松散布解释：精确非中心 \(\chi^{2}\)（均值 \(n+2+\lambda\)）展宽后，S90 仍是确定性值的 0.93–0.97，S50 约 0.78–0.82，且取决于图的 \(\Delta C_A\) 是 \(\lambda\) 还是 \(n+2+\lambda\)（待 PI 确认；此前的 94 / 81 / 43 % 是我把高斯放错了中心，已撤回，07 R023）；2 像素匹配切分最多让 S90 动 0.19。"图在 50 % 处偏保守"与"1.33–1.46\(\times\)"都已撤回。图与合并表无法用交付的代码重建（psf 网格缺失、脚本版本）。这是逐能段探测，不是目录成员资格。Questioned（T218 Q3；07 R020）\(\to\) 待定 / PendingWith both sides at the map's per-pixel n (exposed images, 1–8, mode 3, never 9): S50 observed/map 0.95–0.99 (5 bands), S90 1.11–1.20 (bands 1–4). After RT-24's recheck those numbers are not stable: S50 spans 0.89–0.99 over the n convention and 0.80–0.99 over binning; the Poisson spread does not explain the S90 excess: with the exact noncentral \(\chi^{2}\) (mean \(n+2+\lambda\)) S90 stays at 0.93–0.97 of the deterministic value and S50 near 0.78–0.82, and this depends on whether the map's \(\Delta C_A\) is \(\lambda\) or \(n+2+\lambda\) (to be confirmed by the PI; the earlier 94 / 81 / 43 % came from a Gaussian centred in the wrong place and is withdrawn, 07 R023), and the 2-pix match cut moves S90 by up to 0.19. "The map is conservative at 50 %" and "1.33–1.46\(\times\)" are both withdrawn. The map and the merged table cannot be regenerated from the delivered code (psf grid absent; script version). This is per-band detection, not catalogue membership. Questioned (T218 Q3; 07 R020) \(\to\) Pending | 问 PI：图的 \(\Delta C_A\) 等于 \(\lambda\) 还是 \(\nu+\lambda\)（psf 网格缺失，无法重建图）；数据偏向 \(\nu+\lambda\)（观测完备度 0.857–0.867 对 convB 0.855–0.878、convA 0.906–0.914），但只能与图生成器或一次干净的逐能段注入一起定论（07 R031）；bonus，不再派新任务Ask the PI whether the map's \(\Delta C_A\) is \(\lambda\) or \(\nu+\lambda\) (the psf grid is absent, so the map cannot be regenerated); the data favour \(\nu+\lambda\) (observed completeness 0.857–0.867 against convB 0.855–0.878 and convA 0.906–0.914), but the call needs the map generator or a clean per-band injection (07 R031); bonus, no new task dispatched | T4 / RT-19 已分诊；审计已评审T4 / RT-19 triaged; the audit is reviewed | 开放（bonus；数字不稳定。R031 撤回"泊松散布不能解释 S90 超出"——该结论随 \(\Delta C_A\) 约定翻转，数据偏向 \(\nu+\lambda\)）Open (bonus; the numbers are not stable. R031 withdraws "the Poisson spread does not explain the S90 excess" — that conclusion flips with the \(\Delta C_A\) convention and the data favour \(\nu+\lambda\) |
| Q6Q6\(S_{true}\) 上的先验对基准的估计量有影响吗？Does the prior over \(S_{true}\) matter for the benchmark estimand? | 如果先验效应大，plug-in 评分就不够，需要显式先验。If the prior effect were large, plug-in scoring would not suffice and an explicit prior would be needed. | 效应只在 DET_ML 6–10 窗口内"弱"；换成 3–10 窗口会变号（\(\alpha\) = 2 时 −0.104 \(\to\) +0.165）。ProvisionalThe effect is "weak" only inside DET_ML 6–10; it flips sign for a 3–10 window (at \(\alpha\) = 2, −0.104 \(\to\) +0.165). Provisional | 在全量 L0 统计上重测（T6 续跑）Re-measure at full L0 statistics (T6 continuation) | T6 / RT-10 已分诊T6 / RT-10 triaged | 6–10 已回答；全量待定Answered for 6–10; full statistics pending |
| M1 失败原因The M1 failure's cause（Tier 0，2026-10-23 起）(Tier 0, since 2026-10-23) | M1 能段 1 是已分辨的中心失败（8/21 = 0.381，配对差 −0.333 [−0.619, −0.048]），M1 能段 4 的高计数锚 0/12；不找到原因，Tier 1 不能派发。M1 band 1 is a resolved centre failure (8/21 = 0.381, paired difference −0.333 [−0.619, −0.048]) and M1 band 4's high-count anchor is 0/12; without the cause, Tier 1 cannot be dispatched. | 未知；候选包括 M1 自己的 PSF/背景结构与低 MASKFRAC 的掩膜截断（与 R170 的机制可能相同）。Unknown; candidates include M1's own PSF/background structure and low-MASKFRAC mask truncation (possibly the same mechanism as R170). | M1-ANCHOR（在跑，名字，不引用数字）M1-ANCHOR (running, name only, no numbers quoted) | M1-ANCHORM1-ANCHOR | 开放（07 R169）Open (07 R169) |
| 掩膜截断规则The mask-truncation rule（宽度，2026-10-23 起）(width, since 2026-10-23) | 低 b 高 Ŝ 的宽度偏差（b = 0.0125 的硬格）起初只有一条事后解释；预登记规则（排除中位 MASKFRAC_EML \(<\) 0.98 的位置，R180）已在留出数据上去掉该失误，代价 9.5–11.9 % 的位置，但它是 PI 待决的选择（RT-127 确认预登记属实）。The low-b high-Ŝ width miss (the hard cell at b = 0.0125) started with only a post-hoc explanation; the pre-registered rule (exclude positions with median MASKFRAC_EML \(<\) 0.98, R180) removes it on held data at a cost of 9.5–11.9 % of positions, but adopting it is a PI choice (RT-127 confirms the pre-registration is genuine). | 探测器掩膜截断：少数低 MASKFRAC_EML 拟合位置给出部分被掩的拟合盒，\(\Delta C\) 塌缩、重负 R 尾抬高 FIT 方差锚（丢 1–2 个离群位置后 std(z) 0.856 \(\to\) 0.960；R170，暂定、事后）。Detector-mask truncation: a few low-MASKFRAC_EML FIT positions give partially masked fit boxes, \(\Delta C\) collapses and the heavy negative-R tail inflates the FIT variance anchor (dropping 1–2 outliers moves std(z) 0.856 \(\to\) 0.960; R170, provisional, post hoc). | MASKFRAC-ANCHOR 与 RT-123 已分诊（R180/R178）；RT-127 亦已分诊（R188）。注意：该规则不修 M1 能段 4 的高计数锚（仍 0/12，R187）。MASKFRAC-ANCHOR and RT-123 are triaged (R180/R178), and RT-127 too (R188). Note: the rule does not fix M1 band 4's high-count anchor (still 0/12, R187). | MASKFRAC-ANCHOR / RT-123MASKFRAC-ANCHOR / RT-123 | 开放（07 R170）Open (07 R170) |
| 宽度规则的零假设The width rule's null（判据，等 PI）(the criterion, awaiting the PI) | "8 箱允许 1 箱失败"按 8 箱定标；emldetect 窗口带 9–10 箱、完美 oracle 也失败 2/9，字面规则在那里够不到；不选定零假设，宽度验收没有判据。The "8 bins with 1 failing bin allowed" rule was calibrated for 8 bins; emldetect windows carry 9–10 bins and the perfect oracle itself fails 2/9, so the literal rule is unreachable there; without a chosen null there is no width acceptance criterion. | 两个零假设针对不同的目标："完美规律通过 95 %"给出 \(k^*\) 2–3；更严的冻结格零假设给出 \(k^*\) 0–1（只有玩具与 M1 能段 2 通过）。Two nulls target different things: the "perfect law passes 95 %" null gives \(k^*\) 2–3; the stricter frozen-cell null gives \(k^*\) 0–1 (only the toy and M1 band 2 pass). | RT-124 审核零假设（在跑，名字，不引用数字）；PI 选定RT-124 audits the null (running, name only, no numbers quoted); the PI picks | PI / RT-124PI / RT-124 | 待 PI（07 R171、R175）Awaiting the PI (07 R171, R175) |
| 孤立的新位置Isolated fresh positions（位置项，2026-10-23 起）(the position term, since 2026-10-23) | POS-FRESH 的 27 个新位置里 8 个源太挤（最近邻最小 15.9 px 对原始 45.0 px），所以 SS 位置项至今没有一次干净的检验。8 of POS-FRESH's 27 fresh positions are source-crowded (nearest-neighbour minimum 15.9 px against the originals' 45.0 px), so the SS position term has had no clean test yet. | 不挤的 19 个位置上 +0.21 [−0.05, +0.47]（描述性、事后）；设计对 \(\delta\) 0.15–0.18 检验力不足。On the 19 uncrowded positions +0.21 [−0.05, +0.47] (descriptive, post hoc); the design is underpowered for \(\delta\) 0.15–0.18. | POS-ISOLATED（检验力 + 孤立规则设计，在跑，名字，不引用数字）；任何新位置臂强制最近邻 \(\geq\) 45 pxPOS-ISOLATED (a power + isolation-rule design, running, name only, no numbers quoted); any fresh-position arm enforces a nearest neighbour \(\geq\) 45 px | POS-ISOLATEDPOS-ISOLATED | 开放（07 R165、R173）Open (07 R165, R173) |
| 星表高 \(\mu_0\) 域的偏差形式The bias form for the catalogue's high-\(\mu_0\) domain（星表轴，2026-10-23 起）(the catalogue axis, since 2026-10-23) | 冻结宽度律在 \(\mu_0 \lesssim 250\) 内通过、以上失败；\((f\cdot\mu)^2\) 恢复 MOS 覆盖（0.452 \(\to\) 0.678）但不是分布（z 均值 \(\approx\) −0.75 仍在）；不找到偏差形式，星表律的域就不能扩展。The frozen width law passes within \(\mu_0 \lesssim 250\) and fails above; \((f\cdot\mu)^2\) restores MOS coverage (0.452 \(\to\) 0.678) but not the distribution (z mean \(\approx\) −0.75 remains); without a bias form the catalogue law's domain cannot be extended. | 没有任何预登记偏差项单独管用（逐相机 \(\delta\cdot\mu\) 修好 250 以上却把 250 以下过修；pile-up 被拒，MOS/PN 比值 2.0–2.4 对预测 36）；候选原因是逐相机亮度依赖 PSF 错配（未测）。No pre-registered bias term works alone (a per-camera \(\delta\cdot\mu\) fixes above 250 but over-corrects below 250; pile-up rejected, MOS/PN ratio 2.0–2.4 against a predicted 36); the candidate cause is a per-camera brightness-dependent PSF mismatch (untested). | CAT-PSFMU（逐相机 \(c\cdot\mu^p\) 项，在跑，名字，不引用数字）CAT-PSFMU (a per-camera \(c\cdot\mu^p\) term, running, name only, no numbers quoted) | CAT-PSFMUCAT-PSFMU | 开放（07 R168、R172、R174）Open (07 R168, R172, R174) |
| Tier 1 是否派发Whether to dispatch Tier 1（普适性，T1-a 已发射）(universality, T1-a launched) | Tier 1 已按 PI 选项 A 发射（2026-10-24 02:47 EDT）：T1-a 只跑 M2（N = 115，149–192 CPU-h，MDS80 0.165），M1 只作描述性子比例；T1-e 第一臂已打分（M2 能段 1 中心确认、宽度旗标 AMBIGUOUS、锚 8/12）。T1-b（697 CPU-h）与 T1-d（384 CPU-h）按新代价锚重算后再问 PI。Tier 1 was launched under the PI's option A (2026-10-24 02:47 EDT): T1-a runs M2 only (N = 115, 149–192 CPU-h, MDS80 0.165), with M1 reported as a descriptive sub-fraction only; T1-e's first arm is scored (M2 band 1's centre CONFIRMED, the width flag AMBIGUOUS, the anchor 8/12). T1-b (697 CPU-h) and T1-d (384 CPU-h) go back to the PI after the recost with the new anchor. | 推荐 B：等 M1-ANCHOR 的诊断再决定（选项 A 只在 M2 上跑；C 按原计划两台仪器都跑）。Recommending B: wait for the M1-ANCHOR diagnosis before deciding (option A runs the second stage on M2 only; C runs it as planned on both instruments). | PI 决定（04，2026-10-23 01:47）The PI decides (04, 2026-10-23 01:47) | PIPI | 已发射 T1-a（04，2026-10-24 02:47；07 R221、R224、R225）T1-a launched (04, 2026-10-24 02:47; 07 R221, R224, R225) |
| 背景估计参数（C2）的确认Confirming the background-estimation parameter (C2)（T1-c 后续，2026-10-24 起）(the T1-c follow-up, since 2026-10-24) | 冻结定律在背景通道上不普适（DR14 估计背景下偏移 +0.174 z，簇区间 [0.165, 0.260]）；没有这个参数，T1-a/T1-b 的解释都带着同一个系统差。The frozen law is not universal in the background channel (offset +0.174 z under the DR14-estimated background, cluster interval [0.165, 0.260]); without this parameter, every T1-a/T1-b interpretation carries the same systematic difference. | 候选 C2：\(b_{\rm eff} = \gamma \cdot \mathrm{bgm}_B\)，\(\gamma = 0.921 + 0.0041(\mathrm{off} - 3')\)，把留出偏移从 +0.101 压到 −0.052 z（区间 [−0.073, −0.032]）；但只有一个观测、M2 能段 4、9 个留出位置，且 C2 过修（区间整体在 0 以下）。Candidate C2: \(b_{\rm eff} = \gamma \cdot \mathrm{bgm}_B\), \(\gamma = 0.921 + 0.0041(\mathrm{off} - 3')\), bringing the held offset from +0.101 to −0.052 z (interval [−0.073, −0.032]); but it rests on one observation, M2 band 4, 9 held positions, and C2 over-corrects (its interval lies wholly below 0). | 在第二个观测上预登记确认；物理候选（按 BG_MAP/真背景之比取 \(\rho\)）排队中Pre-registered confirmation on a second observation; a physical candidate (taking \(\rho\) from the BG_MAP/truth ratio) is queued | PI / 调度方PI / dispatcher | 暂定（07 R224、R227、R228、R230）Provisional (07 R224, R227, R228, R230) |
| 星表残差的形状（尾部）The catalogue residual's shape (the tail)（星表轴，2026-10-24 起）(the catalogue axis, since 2026-10-24) | 逐相机 eta 振幅修好合并留出宽度（1.026/0.995/1.021），但没有候选通过预注册的逐格规则（最好 16/87 干净格）；残差偏斜重尾（PN 中位超额峰度 +2.9），方差缩放修不了逐格覆盖。The per-camera eta amplitudes restore the pooled held width (1.026/0.995/1.021), but no candidate passes the pre-registered per-cell rule (the best 16/87 clean cells); the residual is skewed and heavy-tailed (PN median excess kurtosis +2.9), so a variance rescaling cannot fix per-cell coverage. | 尾部项是下一个候选，不是更多振幅参数；但逐格规则本身也要按"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），且没有形状候选通过（最好斜正态 12/87）。A tail term is the next candidate, not more amplitude parameters; but 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 is about 26 cells), and no shape candidate passes (the best skew-normal 12/87). | CAT-TAIL（在跑，名字，不引用数字）；RT-141 的零假设标定已交付CAT-TAIL (running, name only, no numbers quoted); RT-141's null calibration delivered | CAT-TAILCAT-TAIL | 开放（07 R219、R226、R231）Open (07 R219, R226, R231) |
| M1 坏位置的原因The M1 bad positions' cause（Tier 0 遗留，更新于 2026-10-24）(left from Tier 0, updated 2026-10-24) | 三个灾难位置（29/31/35）让 M1 能段 4 的锚停在 0/12；原因不找到，M1 侧的任何普适性判定都不能携带结论。Three catastrophic positions (29/31/35) keep M1 band 4's anchor at 0/12; without the cause, no M1-side universality verdict can be carried. | 掩膜机制被排除（MASKFRAC 0.994）；PSF 加权参考被预注册检验驳回（M1 能段 4 锚 0/12 → 0/12，门槛 \(\leq\) 3/12；护栏 5/6 个块失败）→ 地图/中心错配是伴随物不是原因；旗标 \(F_p > 0.04\) 只是筛选标记（HELD 19/21，p 约 0.04 过不了多重比较）。The mask mechanism is excluded (MASKFRAC 0.994); the PSF-weighted reference is refuted by a pre-registered test (M1 band 4's anchor 0/12 → 0/12 against the bar \(\leq\) 3/12; the guard fails on 5/6 blocks) → the map/centre mismatch is a correlate, not the cause; the flag \(F_p > 0.04\) stays a screening marker only (HELD 19/21, p about 0.04 failing the multiplicity check). | 原因仍开放；T1-a 只把 M1 报成描述性子比例，不携带判定The cause stays open; T1-a reports M1 only as a descriptive sub-fraction, never verdict-carrying | 开放（07 R187、R200、R205、R214）Open (07 R187, R200, R205, R214) | 开放（07 R187、R200、R205、R214）Open (07 R187, R200, R205, R214) |
来源：cognition-os/02_CURRENT_QUESTION.md 问题表 Q1–Q7（2026-10-17T07:10Z，含 PI 优先级横幅：Q7 主攻，Q1/Q2/Q3/Q5 为 bonus）、07_REVISION_LOG.md R001–R033（07 行优先于评审）、03_TRUST_TABLE.md T209–T226、04_DECISION_LOG.md 2026-10-13 至 2026-10-17 08:55 的补记、08_RUNNING.md rev 26、reviews/T13_006_claude.md、reviews/RT-22_triage_claude.md、reviews/RT-23_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、reviews/T13_008_claude.md、.agent/jobs/2026-10-16-detml-m31-ml-convention/REVIEW_claude.md；2026-10-22 的 Q7 更新另引 07_REVISION_LOG.md R128–R161（判据校准、地板改正、打分手冻结、位置项、宽度规则与可观测规律 V̄、宽度区间覆盖限制、误差相消读法、1.16 臂间比值；R157–R161：聚类标准误与种子运气撤回、Tier 0 能段/仪器轴失败（R162 已撤回）与星表轴域内通过、粗锚改判、B011 符号缺陷解释 0 % 的失败）、04_DECISION_LOG.md 2026-10-22 04:22 EDT（\(\varepsilon\) 0.05/0.10）与 10:44 EDT（宽度判据、普适性顺序、次要目标）；2026-10-23/24 的更新另引 07_REVISION_LOG.md R157–R229 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：R220/R229；背景通道非普适与候选 C2：R224/R227/R228/R230；代价锚改正与 T1-a 发射：R218/R221/R225；星表同义反复、断幂律与 eta 振幅：R213/R216/R222/R223/R226/R231；M1 坏位置与 PSF 加权参考驳回：R200/R205/R214；跨拟合膨胀撤回：R204）。Sources: cognition-os/02_CURRENT_QUESTION.md question table Q1–Q7 (2026-10-17T07:10Z, with the PI priority banner: Q7 is the main line, Q1/Q2/Q3/Q5 are bonus), 07_REVISION_LOG.md R001–R033 (07 rows take precedence over reviews), 03_TRUST_TABLE.md T209–T226, the 04_DECISION_LOG.md addenda from 2026-10-13 through 2026-10-17 08:55, 08_RUNNING.md revision 26, reviews/T13_006_claude.md, reviews/RT-22_triage_claude.md, reviews/RT-23_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, reviews/T13_008_claude.md, .agent/jobs/2026-10-16-detml-m31-ml-convention/REVIEW_claude.md; the 2026-10-22 Q7 update also cites 07_REVISION_LOG.md R128–R161 (the criterion calibration, the floor correction, the frozen scorer, the position term, the width rule and the observable law V̄, the width-CI coverage limit, the error-compensation reading, the 1.16 arm ratio; R157–R161: the cluster-SE and seed-luck withdrawals, the Tier 0 band/instrument FAIL (withdrawn in R162) and the catalogue-axis PASS within domain, the coarse-anchor re-interpretation, the B011 sign defect at 0 % of the failures), and 04_DECISION_LOG.md 2026-10-22 04:22 EDT (\(\varepsilon\) 0.05/0.10) and 10:44 EDT (the width rule, the universality order, the secondary goals); the 2026-10-23/24 update also cites 07_REVISION_LOG.md R157–R229 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: R220/R229; the background channel's non-universality and candidate C2: R224/R227/R228/R230; the cost-anchor correction and the T1-a launch: R218/R221/R225; the catalogue tautology, the broken power law and the eta amplitude: R213/R216/R222/R223/R226/R231; the M1 bad positions and the refuted PSF-weighted reference: R200/R205/R214; the cross-fit inflation withdrawn: R204).
未来 1–4 周计划1–4 Week Plan
| 周Week | 做什么What to do | 产出Deliverable |
|---|---|---|
| 本周（当前主攻）This week (the current main line) | 本周（2026-10-17 起，当前主攻）：W1 宽度理论——推导给定 \(\hat{S}\) 的 \(\Delta C\) 条件均值与方差（位置自由 / 低计数与尾部 / 估计背景 / 支撑截断四项），双臂独立推导，每个公式附与精确模拟对照的测试脚本，可执行判据 |mean z| < 0.02 且 |std z − 1| < 0.02。W2 L0/L1 逐行宽度梯子——在 L0（我们的拟合器）与 L1（emldetect）生产表上逐行算 z，按典型 DET_ML 分箱 \(\times\) b \(\times\) 离轴环报告分布，不用 Ŝ 分位数分箱测宽度（R030）。SAS 22.1 对 SAS 20 的源探测重放（PI 要求）：同输入、显式同参数，E068 冻结案例 + 一个单图像配置 + 一个联合多图像配置，比较 DET_ML、SCTS 与按 Ŝ 分箱的 r，预注册 identical / equivalent / different / ambiguous 规则。W3 全局搜索拟合器的条件律（12:54Z 起在跑，双臂）：推导给定 \((\hat{S}, b, \hat{r})\) 的 \(\Delta C\) 条件律（拟合模板偏离中心时支撑被截断），再给出只看 \((\hat{S}, b)\) 的混合律，混合权重由源总体先验经贝叶斯导出；判据同上，另加 recert13 试点设计的 6–20 窗口。理论派发规则（PI 2026-10-17）：Claude 先写"已知结论 + 可执行判据"，两个独立臂并行，红队用数值反例攻击推导，Claude 复核判据脚本后才接受任何公式。This week (from 2026-10-17, the current main line): W1 width theory — derive the conditional mean and variance of \(\Delta C\) given \(\hat{S}\) (position free / low counts and tails / estimated background / support truncation), two independent arms, every formula shipped with a test script against exact simulation, executable key |mean z| < 0.02 and |std z − 1| < 0.02. The W2 L0/L1 per-row width ladder — compute z row by row on the L0 (our fitter) and L1 (emldetect) production tables, report distributions per typical-DET_ML bin \(\times\) b \(\times\) off-axis ring, and never use Ŝ-quantile binning to measure a width (R030). The SAS 22.1 vs SAS 20 source-detection replay (PI request): identical inputs, explicit parameters, the E068 frozen case + one single-image and one joint configuration, comparing DET_ML, SCTS and r binned by Ŝ, with a pre-registered identical / equivalent / different / ambiguous rule. W3, the conditional law for the global-search fitter (running since 12:54Z, two arms): derive the conditional law of \(\Delta C\) given \((\hat{S}, b, \hat{r})\) (the fitted template off-centre has its support truncated), then the law given \((\hat{S}, b)\) alone as a mixture whose weight comes from the source population prior by Bayes; the same key, plus the recert13 pilot design's 6–20 window. Theory-dispatch rule (PI 2026-10-17): Claude writes the known results and the executable key first, two independent arms run in parallel, red teams attack the derivations with numerical counterexamples, and Claude replays each formula's test script before accepting it. | 已交付（都经评审，各自带来源与信任标签）：W1 玩具理论的逐格通过/失败表（T229/E170，Provisional）——位置固定的 Fisher 宽度在 DET_ML \(\gtrsim\) 20 与低 b 通过，DET_ML 6–20 且 b \(\geq\) 0.02 失效（1.17–2.49，该失效读数已由 R044 废弃（拟合器伪迹），只留 (15.32, 0.08) 一格约 2.5 % 的窄侧偏离，R045），均值公式无一通过，截断支撑的矩全通过；W2 的 L0/L1 逐行宽度梯子（T228，Provisional；L1 数字已按 R041 用 BG_MAP_EML 重新打分）；SAS 版本差别的参数表与字面结论（T227/E169，Provisional）；Q1 的冻结目标 −9.282 % [−9.350, −9.219]（R035，bonus）；宽度红队 RT-38 到 RT-45 的分诊。在跑：W3（全局搜索拟合器的条件律，双臂）与红队 RT-46、RT-47，结果未出（RT-43/44/45 已分诊：ecut 0.68 机制保留、L1 换数、W1 宽度失效一半废弃）。Delivered (all reviewed, each with its source and trust label): W1's toy per-cell pass/fail table (T229/E170, Provisional) — the position-fixed Fisher width passes at DET_ML \(\gtrsim\) 20 and at low b, fails at DET_ML 6–20 with b \(\geq\) 0.02 (1.17–2.49, that failure reading is Deprecated by R044 (a fitter artefact), leaving only a narrow-side departure of about 2.5 % at the (15.32, 0.08) cell, R045), no mean formula passes, the truncated-support moments all pass; W2's L0/L1 per-row width ladder (T228, Provisional; the L1 numbers rescored with BG_MAP_EML per R041); the SAS version-difference parameter table and literal verdict (T227/E169, Provisional); Q1's frozen target −9.282 % [−9.350, −9.219] (R035, bonus); the width red teams RT-38 through RT-45 triaged. Running: W3 (the conditional law for the global-search fitter, two arms) and red teams RT-46, RT-47, results pending (RT-43/44/45 are triaged: the ecut 0.68 mechanism kept, the L1 numbers replaced, the W1 width-failure half deprecated). |
| 第 1 周Week 1 | T11 第 21–26 阶段已评审，T11 暂停（星表列路线已走完）；T13 第 2 阶段已评审但受位置文件 bug 影响（RT-05），修复后重做未开始；T13 第 3 阶段（SAS emldetect 拆分，4 种配置）已评审（T13_003，通过附条件：复现 A 成立，B/C 不是机制，D 待第 4 阶段的 EXT 拆分确认）；第 4 阶段运行中。红队机制：12 个红队已跑（同时保持 2–4 个活跃），RT-01 到 RT-09 与 RT-11 已分诊——它们更正了 T13 的"约一半"读数（RT-05）、把 2-D 损失收窄为 6–10 数值修正（RT-06）、更正了 ">300" 窗口的含义（RT-02）、emldetect 描述的"\(\pm\)2%"（RT-09）与 MASKFRAC 的"部分遮挡 m\(\approx\)0"读数（RT-11），同时确认了 L0 对照、M1 阶跃与排行榜指标（RT-03/04/07）；RT-10（先验）、RT-12（SAS 复现）运行中T11 stages 21–26 reviewed, T11 paused (the catalogue-column route exhausted); T13 stage 2 is reviewed but affected by the position-file bug (RT-05), the redo after the fix not started; T13 stage 3 (the SAS emldetect decomposition, 4 configurations) is reviewed (T13_003, PASS with conditions: the replica A validated, B/C not the mechanism, D awaiting stage 4's EXT split); stage 4 running. Red-team mechanism: 12 red teams have run (2–4 active at a time), RT-01 through RT-09 and RT-11 triaged — they corrected T13's "about half" reading (RT-05), narrowed the 2-D loss to a numerical 6–10 correction (RT-06), fixed the meaning of the ">300" window (RT-02), the emldetect description's "\(\pm\)2%" (RT-09) and the MASKFRAC "partly masked m\(\approx\)0" reading (RT-11), while confirming the L0 control, the M1 step and the leaderboard metric (RT-03/04/07); RT-10 (prior) and RT-12 (the SAS replica) running | T13 第 3 阶段：SAS emldetect 4 种配置（复现管线、单相机单波段、似然支撑 15px 圆、关闭展源拟合）逐项拆分；红队报告的分诊T13 stage 3: the SAS emldetect decomposition over 4 configurations (reproduce the pipeline, single camera single band, 15px-circle likelihood support, extent fitting off); triage of the red-team reports |
| 第 1–2 周Weeks 1–2 | 守护 L0 生产到完成（预计 2026-10-20），逐块验证拉回；L1 同样（约 6 天）Guard L0 production to completion (ETA 2026-10-20), per-chunk verify & pull back; L1 likewise (~6 days) | 全量 L0/L1 表 + 清单Full L0/L1 tables + manifests |
| 第 2–3 周Weeks 2–3 | L0 全量正式计分（先演练到 scratch）；宽度律在 L0 上检验（攻 A2 缺口）Full L0 formal scoring (dry-run to scratch first); test width laws on L0 (attack A2 gap) | L0 层条目（由 PI 登记）；A2 的下一步方案L0 layer entries (registered by PI); A2 next-step plan |
| 第 3–4 周Weeks 3–4 | 若 PN 损失定案且 A2 有候选：提交新条目、重算前沿；视结果请 PI 开私有位置If PN loss decided and A2 has candidate: submit new entries, recompute front; ask PI to open private positions | 更新的前沿；里程碑材料Updated front; milestone materials |
| T5 阶段 12–18T5 stages 12–18 | 阶段 12–18 已完成（T5_009–015 评审通过）：L0 对照 \(m_{L0}\) = 0.0038（不截断主读数，RT-13；截断值 0.0077 仅敏感性）；核失效修复（15 个网格到 S = 5000）；ML\(\to\)\(\Delta C\) 溢出修复（erfcx 稳定反解）；m 延续到亮源——6 个 T3 条目（4 个零参数）从 S \(\approx\) 100 到 4000 保持同一 m（|\(\Delta\)m| < 0.01），V5c_eta 在 >300 比不修正更差（|frac| 差在所有区间方法下不含 0；C-all 的 A1 差依赖方法：百分位 [−0.18, +0.72]，BCa [+0.31, +0.78]）。阶段 18 已完成：区间根因修复（bootstrap 直接计算所报告的 \(A1_{nc}\)），高计数报告页已发布（T5_015）。T5 已解除暂停（PI 2026-10-16 批准高 S 总体进层定义）：第 16 阶段做 C-all-hi / C-iso-hi（S 上限 4000，扩展核网格有效到 5000）两个新层，与现有层并列，不替代；已登记层、6–10 前沿和所有已登记条目不变Stages 12–18 done (T5_009–015 reviews PASS): L0 control \(m_{L0}\) = 0.0038 (the uncut primary reading, RT-13; the cut value 0.0077 a sensitivity only); kernel failure fixed (15 grids to S = 5000); ML\(\to\)\(\Delta C\) overflow fixed (erfcx stable inversion); m extends to bright sources — the 6 T3 entries (4 zero-parameter) keep the same m from S \(\approx\) 100 to 4000 (|\(\Delta\)m| < 0.01), V5c_eta is worse than no correction in >300 (its |frac| difference excludes 0 under every interval method; the C-all A1 difference is method-sensitive: percentile [−0.18, +0.72], BCa [+0.31, +0.78]). Stage 18 done: interval root cause fixed (the bootstrap evaluates the reported \(A1_{nc}\)), the high-count report page is published (T5_015). T5 is unblocked (the PI approved the high-S population for the layer definitions on 2026-10-16): stage 16 makes C-all-hi / C-iso-hi (S cap 4000, extended kernel grids valid to 5000) two new layers beside the existing ones, not replacing them; the registered layers, the 6–10 front and every registered entry stay unchanged | 高计数退化标记扩展到 DET_ML>300 窗口：已试行（用户决定，2026-10-15 09:50；只作报告，不进前沿）；高 S 总体进层定义：已批准（PI 2026-10-16），作业尚未开跑 \(\to\) 待定 / Pending；/v2/ 页面更新（含 4 个新登记条目与新前沿）并入 T5 第 16 阶段Extending the high-count degradation flag to the DET_ML >300 window: on trial (user decision, 2026-10-15 09:50; report only, not the front); the high-S population entering the layer definitions: approved (PI 2026-10-16), the job has not run yet \(\to\) Pending; the /v2/ page refresh (with the 4 new registered entries and the new fronts) folds into T5 stage 16 |
依据：BENCHMARK_V2_TASKS.md 状态段（2026-10-13）、reviews/T11_020 到 T11_026、T5_009 到 T5_015、T13_001 到 T13_002 评审与红队分诊 RT-01 到 RT-09。每步的详细操作见Based on: BENCHMARK_V2_TASKS.md state section (2026-10-13), reviews T11_020 through T11_026, T5_009 through T5_015, T13_001 through T13_002, and the red-team triages RT-01 through RT-09. Detailed steps in the 动手指南Handbook。.
本次更新（第 22 版，2026-10-24）What's New (version 22, 2026-10-24)
- Tier 1 第一臂（T1-e）出结果：M2 能段 1 三条判定分开记，宽度判 AMBIGUOUS（R220/R229）。唯一携带判定的块 M2 能段 1：中心 D+SS 于 \(\varepsilon\) 0.10 是 24/33 = 0.727，配对差区间 [−0.152, +0.061]，按预注册中心规则确认；宽度 4 个失败箱对 \(k^*\) 2 → 先是 FAIL，红队复核（RT-138）确认超出是真实的（9/9 个箱 std > 1，符号检验 p 0.002，位置重分稳定，中位 1.021，单侧约 2 %），但箱边移 −1/3 个箱就翻成 PASS、35 % 的重分给出 \(\leq\) 2 个失败，所以旗标记 AMBIGUOUS；锚 8/12。参照块 M2 能段 4：25/33、宽度 1/2、锚 12/12。执行方的"整体 REFUTED"不是预注册的组合（方案第 5 节三条规则各自独立、没有合并规则），因此只记三条独立判定。已撤回、不得再作为现状引用："T1-e 整体 REFUTED"、"M2 能段 1 宽度 FAIL 是一个稳健结果"、跨拟合方差膨胀 c = 1.0323 作为修正（R204：盲目迁移把原本通过的块弄坏，M1 能段 1 从 3 个失败涨到 6 个）。Tier 1's first arm (T1-e) reported: M2 band 1's three verdicts recorded separately, the width flag AMBIGUOUS (R220/R229). On the only verdict-carrying block, M2 band 1: the centre D+SS at \(\varepsilon\) 0.10 is 24/33 = 0.727, paired-difference interval [−0.152, +0.061], CONFIRMED under the pre-registered centre rule; the width has 4 failing bins against \(k^*\) 2 → first FAIL, then the red-team re-check (RT-138) confirmed the excess is real (9/9 bins with std > 1, sign test p 0.002, stable to position re-splits, median 1.021, one-sided about 2 %) but the verdict flips to PASS under a −1/3-bin edge shift and 35 % of re-splits give \(\leq\) 2 fails, so the flag reads AMBIGUOUS; the anchor is 8/12. The reference block M2 band 4: 25/33, width 1/2, anchor 12/12. The executor's "overall REFUTED" is not a pre-registered combination (plan §5 states the three rules separately, with no merge rule), so only the three separate verdicts are recorded. Withdrawn, never to be quoted as current: "T1-e overall REFUTED", "the M2 band 1 width FAIL is a robust result", and the cross-fit variance inflation c = 1.0323 as a fix (R204: blind transfer breaks passing bins, M1 band 1 going from 3 to 6 fails).
- T1-c 背景通道：冻结定律不普适，定律需要一个背景估计参数（R224/R227/R228，暂定）。527 次重复、M2 能段 4，真背景臂 A 对 DR14 估计背景臂 B：配对窗口偏移 +0.174 z（指数 2.5 处；全部指数 0.171–0.186，簇区间 [0.165, 0.260] 不含 \(\varepsilon\) 0.10）。规则 (ii) 通过（斜率 0.815，r 0.781）；规则 (i) 未决（冻结合并打分手零支撑，每位置 neff 73–79 < 100，执行方的"STABLE"依赖一个两臂都 0/72 通过的备用格打分手）。窗口切在 DET_ML_A [6,10) 是一次铁律型行切；调度方复核：窗口切在 B 上偏移 0.164、按均值 0.176、不开窗口 0.151 → 失败不是选择伪影。红队 RT-139 排掉十条伪影通道（配对、核在错背景上取值、网格外插、z 归一、位置集中 LOPO [0.168, 0.185]、流水线安慰剂 0.03–0.05、窗口边、邻居、重建、打分手支撑），机制 \(z_B - z_A \approx 48.2 \times \Delta b_{\rm PSF}\)（零截距）。候选 C2：\(b_{\rm eff} = \gamma \cdot \mathrm{bgm}_B\)，\(\gamma = 0.921 + 0.0041(\mathrm{off} - 3')\)，留出主偏移 +0.101 → −0.052 z，区间 [−0.073, −0.032] 落在 \(\pm 0.10\) 内（预注册通过）；\(\gamma < 1\) 意味着 DR14 BG_MAP 把局部背景高估约 8 %。暂定：一个观测、M2 能段 4、9 个留出位置；C2 过修（区间整体在 0 以下）；T1-a 只把它作为预注册的次要诊断，进定律前要在第二个观测上确认。T1-c, the background channel: the frozen law is not universal, so the law needs a background-estimation parameter (R224/R227/R228, provisional). 527 repeats, M2 band 4, truth-background arm A against DR14-estimated-background arm B: the paired window offset is +0.174 z (at index 2.5; all indices 0.171–0.186, cluster interval [0.165, 0.260] excluding \(\varepsilon\) 0.10). Rule (ii) passes (slope 0.815, r 0.781); rule (i) is undecided (the frozen pooled scorer has zero support, per-position neff 73–79 < 100, and the executor's "STABLE" rests on a fallback cell scorer passing 0/72 in both arms). The window cut on DET_ML_A [6,10) is an Iron-Rule-type row cut; the dispatcher's check: offset 0.164 with the window on B, 0.176 on the mean, 0.151 with no window → the failure is not a selection artefact. Red team RT-139 excluded ten artefact channels (pairing, the kernel evaluated at the wrong background, grid extrapolation, z normalisation, position concentration LOPO [0.168, 0.185], a pipeline placebo 0.03–0.05, window edges, neighbours, reconstruction, scorer support); the mechanism is \(z_B - z_A \approx 48.2 \times \Delta b_{\rm PSF}\) (zero intercept). Candidate C2: \(b_{\rm eff} = \gamma \cdot \mathrm{bgm}_B\), \(\gamma = 0.921 + 0.0041(\mathrm{off} - 3')\), bringing the held primary offset from +0.101 to −0.052 z, its interval [−0.073, −0.032] inside \(\pm 0.10\) (pre-registered pass); \(\gamma < 1\) means the DR14 BG_MAP over-estimates the local background by about 8 %. Provisional: one observation, M2 band 4, 9 held positions; C2 over-corrects (its interval lies wholly below 0); T1-a uses it only as a pre-registered diagnostic secondary, and a second observation is required before it enters the law.
- 代价锚改正 2.4 倍：实测 0.97 CPU-s/行（区间 0.86–1.3），T1-a 只跑 M2（R218/R221/R225）。旧锚 0.4 CPU-s/行在算术上就是错的；virgo 实测均值 0.97（py 段 0.55 于 S 2.5 到 1.20 于 S 320，对数线性；SAS 0.134；mock 0.02），与图像尺寸无关。按 621,000 行核算：0.86 s 给 148 CPU-h、0.969 s 给 167.2 CPU-h。重算后：T1-a 只 M2（N = 115）149–192 CPU-h，MDS80 0.165；三观测合并（N = 197）255–329 CPU-h，MDS80 0.126；T1-b 697；T1-d 384；T1-e py 解析 5 个块 3,650 CPU-h（旧数 765–1309 作废）；PN b1/b4 若重跑，1000 次重复要约 1,572 CPU-h。PI 决定（2026-10-24 02:47 EDT）：T1-a 取选项 A（只 M2，N = 115，149–192 CPU-h），并已在 BG-PARAM 交付后发射；PN b1/b4 取选项 B——不重跑，作为 T1-e 的显式缺口保留（lang2 源树已不在，完整 py 输入要约 397 GB）。The cost anchor corrected by 2.4\(\times\): the measured cost is 0.97 CPU-s/row (band 0.86–1.3), so T1-a runs M2 only (R218/R221/R225). The old 0.4 CPU-s/row anchor was arithmetically broken; the virgo measurement gives a mean of 0.97 (the py stage 0.55 at S 2.5 to 1.20 at S 320, log-linear; SAS 0.134; mock 0.02), independent of image size. Checked by hand on 621,000 rows: 0.86 s gives 148 CPU-h and 0.969 s gives 167.2 CPU-h. Recosted: T1-a M2-only (N = 115) 149–192 CPU-h, MDS80 0.165; pooled over three observations (N = 197) 255–329 CPU-h, MDS80 0.126; T1-b 697; T1-d 384; T1-e's py parse of 5 blocks 3,650 CPU-h (the old 765–1309 figure is superseded); re-running PN b1/b4 at 1000 reps would cost about 1,572 CPU-h. The PI's decision (2026-10-24 02:47 EDT): T1-a takes option A (M2-only, N = 115, 149–192 CPU-h) and was launched once BG-PARAM delivered; PN b1/b4 takes option B — not re-run, kept as an explicit T1-e gap (the lang2 source tree is gone and a full py input would be about 397 GB).
- 星表律：DET_ML 有序结构是标注同义反复；断幂律 5/6 通过但 M2 低 \(\mu_0\) 只能事后补；eta 振幅修好合并宽度、逐格形状失败（R213/R216/R219/R222/R223/R226）。RT-133 的法律精确零假设（同样的行、由模拟 \(\Delta C\) 重建 DET_ML、同样的窗口）复现了切片内 mean_z 展布，实/零之比约 0.93 → 所谓"DET_ML 有序残差"是标注同义反复（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（加宽拟合合同到含 \(\mu_0\) > 600 的拟合半）在 6 个主格中过 5 个（仅 M2 \(\mu_0\) 100–250 失败 +0.133），但 RT-137 的按 OBS_ID 簇自助说 M2 那一格通过比例 0.000（稳健失败）、PN 0.729（勉强）；M2 低 \(\mu_0\) 的补丁（"m2low"：\(\mu_0\) 250 以下退回冻结连续律）过 6/6，但候选是在对手分数已知之后选的，不是确认性证据。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，混合核心 0.75/0.81/0.83）；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，不是更多振幅参数。The catalogue law: the DET_ML-ordered structure is a labelling tautology; the broken power law passes 5/6 but M2's low \(\mu_0\) is patched only post hoc; the eta amplitude fixes the pooled width while the per-cell shape fails (R213/R216/R219/R222/R223/R226). RT-133's law-exact null (the same rows, DET_ML rebuilt from the simulated \(\Delta C\), the same windows) reproduces the within-slice mean_z spread, real/null about 0.93 → the alleged "DET_ML-ordered residual" is a labelling tautology (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 (only M2 \(\mu_0\) 100–250 fails, +0.133), but RT-137's per-OBS_ID cluster bootstrap puts that cell's pass fraction at 0.000 (a robust fail) and PN's at 0.729 (marginal); the M2 low-\(\mu_0\) patch ("m2low": below \(\mu_0\) 250 revert to the frozen continuous law) passes 6/6, but the candidate was chosen after its opponent's score was already known, so it is not confirmatory. 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, mixture cores 0.75/0.81/0.83); 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, not more amplitude parameters.
- M1 坏位置的原因仍然开放：PSF 加权参考被预注册检验驳回（R214），地图错配旗标只是筛选标记（R200/R205）。M1-BADPOS 说三个灾难位置（29/31/35）不是掩膜孔、热像素/CCD 缝、邻居或 PSF 形状离群；候选机制是参考把曝光/背景塌到中心像素值、而 emldetect 看得见盒内的地图梯度。旗标 \(F_p > 0.04\)（FIT 半选定）特异但不强：FIT 21/21、HELD 19/21（比什么都不标只多 1），p 约 0.04 过不了 26 特征 FIT 搜索的多重比较。PSFW-REF 直接检验机制：M1 能段 4 锚 0/12 → 0/12（门槛 \(\leq\) 3/12 = 驳回），5/6 个块的护栏失败（M2 能段 4 锚 9 → 0）→ 地图/中心错配是坏位置的伴随物，不是原因。在跑（不引用数字）：T1A-RUN（T1-a 主臂，M2-only、3 个新观测）、CAT-TAIL（星表残差尾行的身份）、RT-140（C2 的直接物理核对）、RT-141（星表逐格规则的正确模型零假设）、PY-PARSE（T1-e 剩余块的 py 解析）。The M1 bad positions' cause stays open: the PSF-weighted reference is refuted by a pre-registered test (R214) and the map-mismatch flag is a screening marker only (R200/R205). M1-BADPOS found the three catastrophic positions (29/31/35) are not a mask hole, a hot pixel/CCD gap, a neighbour or a PSF-shape outlier; the candidate mechanism is that the reference collapses exposure/background to the centre-pixel value while emldetect sees the map gradient inside the box. The flag \(F_p > 0.04\) (chosen on the FIT half) is specific but thin: FIT 21/21, HELD 19/21 (only +1 over flagging nothing), and p about 0.04 does not survive the 26-feature FIT search's multiplicity. PSFW-REF tested the mechanism directly: M1 band 4's anchor 0/12 → 0/12 (the bar \(\leq\) 3/12 = refutes) and the guard fails on 5/6 blocks (M2 band 4's anchor 9 → 0) → the map/centre mismatch is a correlate of the bad positions, not their cause. Running (no numbers quoted): T1A-RUN (the T1-a main arm, M2-only, 3 new observations), CAT-TAIL (what the catalogue residual's tail rows are), RT-140 (a direct physical check on C2), RT-141 (the correct-model null for the catalogue per-cell rule), PY-PARSE (the py parse of T1-e's remaining blocks).
本次更新（第 21 版，2026-10-23）What's New (version 21, 2026-10-23)
- 普适性 Tier 0：一个失败被撤回，一个新的失败出现（R158 \(\to\) R162/R167/R169）。能段/仪器轴上，A0-SCORE 报的失败是一次打分口径错误：它把 SS 位置项逐行放进 z（\(z = (R - a - \beta_i - ss)/(s\sqrt{V_{3p}})\)），而冻结打分手把它从合并偏移里减掉。按冻结口径、冻结能段 4 节点、零新参数：M2 能段 1 的 D+SS 在 \(\varepsilon\) 0.10 处 24/33 = 0.7273 \(\geq\) 门槛 0.6273（配对差 0.000 [−0.121, +0.121]）；逐行口径恰好复现 0.6061。RT-119 确认合并口径是物理上对的那个（逐行除法平均只施加约 37 % 的改正）。但 A0-RESCORE 给出全貌：M1 能段 1 为 8/21 = 0.381，配对差 −0.333 [−0.619, −0.048]，已分辨的中心失败；M1 能段 2/3/4 在 0.10 处分辨不了（MDS 约 0.30）；高计数锚 M2 能段 1 过 4/12 格、M1 能段 4 0/12。Tier 0 判定每个块都失败；Tier 1 压住等 PI，M1-ANCHOR 出发找原因。The universality Tier 0: one failure withdrawn, a new one found (R158 \(\to\) R162/R167/R169). On the band/instrument axis the failure A0-SCORE reported was a scoring-convention error: it put the SS position term per row into z (\(z = (R - a - \beta_i - ss)/(s\sqrt{V_{3p}})\)), whereas the frozen scorer subtracts it from the pooled offset. Under the frozen convention, the frozen band-4 nodes and zero new parameters: M2 band 1's D+SS at \(\varepsilon\) 0.10 is 24/33 = 0.7273 \(\geq\) the bar 0.6273 (paired difference 0.000 [−0.121, +0.121]); the row-level convention reproduces 0.6061 exactly. RT-119 confirmed the pooled convention is the physically right one (row-level division applies only about 37 % of the correction on average). But A0-RESCORE gave the full picture: M1 band 1 is 8/21 = 0.381, paired difference −0.333 [−0.619, −0.048], a resolved centre failure; M1 bands 2/3/4 cannot be resolved at 0.10 (MDS about 0.30); the high-count anchor passes 4/12 cells on M2 band 1 and 0/12 on M1 band 4. Tier 0 fails on every block; Tier 1 is held for the PI and M1-ANCHOR was launched to find the cause.
- 星表轴：冻结宽度律在 \(\mu_0 \lesssim 250\) 域内通过（R160/R163），\((f\cdot\mu)^2\) 恢复 MOS 高计数覆盖（R168），但没有偏差项单独管用（R172/R174）。留出 DR14 层上，DET_ML 6–10 至 50–100 的每个窗口、全部 15 个配置、两个层的 68 % 覆盖 0.69–0.70、95 % 覆盖 0.95–0.96；DET_ML 100 以上全部 15 个失败（cov68 0.33–0.86），其中只有 10 个 MOS 配置失败、PN 能段 1–3 反而过覆盖（0.74–0.86）；\(G^2 = 1 - 1.5785/\mu^{0.865}\) 在 1 处饱和，所以这个函数形式内部没有能把域推上去的重拟合。加 \(\sigma^2 \to \sigma^2 + (f\cdot\mu)^2\)（每相机一个 f）把 MOS 在 \(\mu_0\) 250 以上的 cov68 从 0.452 修到 0.678；f 随能段变化（M1 0.075–0.138、M2 0.052–0.100、PN 0.003–0.090）。但它恢复的是覆盖不是分布（MOS z 均值 \(\approx\) −0.75 仍在）；逐相机 \(\delta\cdot\mu\) 修好 250 以上却把 250 以下过修；pile-up 被拒（MOS/PN 比值 2.0–2.4 对预测 36）；PSF 定标的理由说过头了。单独哪一项都不是规律。The catalogue axis: the frozen width law passes within the \(\mu_0 \lesssim 250\) domain (R160/R163), \((f\cdot\mu)^2\) restores MOS coverage at high counts (R168), but no bias term works alone (R172/R174). On held-out DR14 layers, every DET_ML window from 6–10 through 50–100, all 15 configurations and both layers give 68 % coverage 0.69–0.70 and 95 % coverage 0.95–0.96; above DET_ML 100 all 15 fail (cov68 0.33–0.86), of which only the 10 MOS configurations fail while PN bands 1–3 over-cover (0.74–0.86); \(G^2 = 1 - 1.5785/\mu^{0.865}\) saturates at 1, so no refit inside this functional form can extend the domain. Adding \(\sigma^2 \to \sigma^2 + (f\cdot\mu)^2\) (one f per camera) restores MOS cov68 above \(\mu_0\) 250 from 0.452 to 0.678; f is band-dependent (M1 0.075–0.138, M2 0.052–0.100, PN 0.003–0.090). But it restores coverage, not the distribution (MOS z mean \(\approx\) −0.75 remains); a per-camera \(\delta\cdot\mu\) fixes above 250 but over-corrects below 250; pile-up is rejected (MOS/PN ratio 2.0–2.4 against a predicted 36); the PSF-calibration justification is overstated. Neither term alone is the law.
- 宽度：聚类标准误、粗锚假象、细锚到 oracle 下限（R157/R159/R164/R166），判据重新定标（R171），低 b 硬格指向掩膜截断（R170）。逐箱检验的行自助标准误忽略 (pid, S_true) 聚类设计（deff 1.26–1.8）；换聚类标准误后 V̄ 的失败从 7/9 降到 5/9，玩具从 1/8 降到 0/8，而"匹配功率 5 箱只失败 1 箱"是种子运气（P(\(\leq\) 1) = 0.067）。"窄核靠误差相消通过宽度"是粗锚假象：细锚下排序每个分半都反转（2.03 % 对 4–5 %）。细锚规律在 emldetect M2 能段 4 上失败 2/9、RMS 0.045，等于完美 oracle 的下限；约 2 % 的 RMS 带聚类区间 [1, 7] %。但"8 箱允许 1 箱失败"的 PI 规则按 8 箱定标，emldetect 窗口带 9–10 箱、完美 oracle 也失败 2/9，所以字面规则在那里够不到——重新定标给出 \(k^*\) 2–3（或更严的 0–1），PI 必须选零假设。低 b 高 Ŝ 的硬格（b = 0.0125）偏差指向少数低 MASKFRAC_EML 位置的探测器掩膜截断（丢 1–2 个离群位置后 std(z) 0.856 \(\to\) 0.960），但挑选是事后的，需要预登记规则。Width: the cluster SE, the coarse-anchor artefact, the fine anchor at the oracle floor (R157/R159/R164/R166), the rule recalibration (R171), and the low-b hard cell pointing at mask truncation (R170). The per-bin test's row-bootstrap SE ignores the (pid, S_true) cluster design (deff 1.26–1.8); with the cluster SE V̄'s failures fall from 7/9 to 5/9 and the toy's from 1/8 to 0/8, while the matched-power "1 failing bin of 5" was seed luck (P(\(\leq\) 1) = 0.067). "The narrow kernel passes width by error cancellation" is a coarse-anchor artefact: with fine anchors the ordering reverses in every split (2.03 % against 4–5 %). The fine-anchor law fails 2/9 bins on emldetect M2 band 4 with RMS 0.045, equal to the perfect oracle's floor; the about-2 % RMS carries a cluster interval [1, 7] %. But the PI's "8 bins with 1 failing bin allowed" rule was calibrated for 8 bins, emldetect windows carry 9–10 bins and the perfect oracle itself fails 2/9, so the literal rule is unreachable there — the recalibration gives \(k^*\) 2–3 (or 0–1 under the stricter null), and the PI must pick the null. The low-b high-Ŝ hard cell's (b = 0.0125) miss points at detector-mask truncation at a few low-MASKFRAC_EML positions (dropping 1–2 outliers moves std(z) 0.856 \(\to\) 0.960), but the picking is post hoc and needs a pre-registered rule.
- 位置项在新位置上 AMBIGUOUS，且这个测试不能读作对 SS 项的检验（R165/R173）。27 个新留出位置（冻结系数、合并 SS 口径）上，\(\varepsilon\) 0.10 处 D 10/27 = 0.370 \(\to\) D+SS 14/27 = 0.519，配对差 +0.148 [−0.037, +0.333]，区间含 0 且上端过 0.10 \(\to\) 预注册 AMBIGUOUS。但 RT-121 发现 27 个新位置里 8 个源太挤（最近邻最小 15.9 px 对原始 67 位置的 45.0 px），拥挤把高 S 的 \(\Delta C\) 压低最多 −4\(\sigma\)；不挤的 19 个位置上 +0.21 [−0.05, +0.47]（描述性、事后）。任何新位置臂（包括 Tier 1 的新观测位置）今后必须强制最近邻 \(\geq\) 45 px（04，2026-10-23）。The position term is AMBIGUOUS at fresh positions, and that test cannot be read as a test of the SS term (R165/R173). At 27 new held positions (frozen coefficients, pooled SS convention), at \(\varepsilon\) 0.10 D goes 10/27 = 0.370 \(\to\) D+SS 14/27 = 0.519, paired difference +0.148 [−0.037, +0.333]; the interval contains 0 with the upper end above 0.10 \(\to\) the pre-registered AMBIGUOUS. But RT-121 found 8 of the 27 fresh positions source-crowded (nearest-neighbour minimum 15.9 px against 45.0 px for the 67 originals), crowding depressing high-S \(\Delta C\) by up to −4\(\sigma\); on the 19 uncrowded ones +0.21 [−0.05, +0.47] (descriptive, post hoc). Any fresh-position arm (including Tier 1's new-observation positions) must from now on enforce a nearest neighbour \(\geq\) 45 px (04, 2026-10-23).
- B011 符号缺陷解释 0 % 的迁移失败；玩具 V̄ 面弃用（R161）。缺陷在 run_binconst.py:79（"beta-1.5" 解析成 +1.5），三个导出面全受影响、改正后差到 49.18 %；但 R151/R157 的满统计量失败用的是重建过的 V̄，不受影响（对照复跑逐位复现），R158 里改正符号只会让 vbar_min 更负、宽度判定维持 96/96 失败。宽度工作改用细锚（emldetect 重建、细锚）路线。The B011 sign defect explains 0 % of the transfer failures; the toy V̄ surface is dropped (R161). The defect sits at run_binconst.py:79 ("beta-1.5" parsed to +1.5), all three exported surfaces affected, differing by up to 49.18 % after the fix; but R151/R157's full-statistics failures used the rebuilt V̄ and are not exposed (a control rerun reproduces bit-for-bit), and in R158 correcting the sign only makes vbar_min more negative, the width verdict staying 96/96 failing. Width work proceeds on the fine-anchor (emldetect-rebuilt, fine anchors) route.
- 在跑（不引用数字）：M1-ANCHOR（M1 失败原因）、MASKFRAC-ANCHOR（掩膜份额条件下的宽度锚）、CAT-PSFMU（逐相机 \(c\cdot\mu^p\) PSF 错配项）、POS-ISOLATED（检验力 + 孤立规则设计）、RT-123（审 R170）、RT-124（审 R171 的零假设）。待 PI：Tier 1 是否派发（推荐 B：等 M1-ANCHOR 诊断）；宽度规则的零假设；星表高 \(\mu_0\) 域的偏差形式。Running (no numbers quoted): M1-ANCHOR (the cause of the M1 failure), MASKFRAC-ANCHOR (a width anchor conditioned on mask fraction), CAT-PSFMU (a per-camera \(c\cdot\mu^p\) PSF-mismatch term), POS-ISOLATED (a power + isolation-rule design), RT-123 (auditing R170), RT-124 (auditing R171's null). For the PI: whether to dispatch Tier 1 (recommending B: wait for the M1-ANCHOR diagnosis); the width rule's null; the bias form for the catalogue's high-\(\mu_0\) domain.
本次更新（第 20 版，2026-10-22）What's New (version 20, 2026-10-22)
- 进度指标定死：\(\varepsilon\) 0.05 比较规律、\(\varepsilon\) 0.10 跟踪进度（PI 04，2026-10-22 04:22 EDT）。此前"每条条目都在 \(\varepsilon\) 内"的判据噪声受限（R128：完美规律在能段 4 只解 28.3/33，位置盲规律约 0.3/33；R129：天花板 0.21–0.27），改成每个位置一个合并偏移（R131：天花板 1.000，但 \(\varepsilon\) 0.20 饱和）；"\(\varepsilon\) 0.10 最敏感"被 R137 证伪（半正态敏感度在 \(\varepsilon\) → 0 最大，分离度随 \(\varepsilon\) 单调下降），推荐拆成 0.05 比较 / 0.10 跟踪，PI 采纳。\(\varepsilon\) 0.10 \(\approx\) 0.12 DET_ML（6–10）。The progress metric is fixed: \(\varepsilon\) 0.05 to compare laws, \(\varepsilon\) 0.10 as the tracker (PI 04, 2026-10-22 04:22 EDT). The earlier "every entry within \(\varepsilon\)" rule was noise-limited (R128: a perfect law solves 28.3/33 on band 4, a position-blind one about 0.3/33; R129: ceiling 0.21–0.27), so the criterion became one pooled offset per position (R131: ceiling 1.000, but \(\varepsilon\) 0.20 saturated); "\(\varepsilon\) 0.10 is most sensitive" is refuted by R137 (the half-normal sensitivity peaks as \(\varepsilon\) → 0 and the separation falls monotonically in \(\varepsilon\)), and the PI adopted the split 0.05 compare / 0.10 track. \(\varepsilon\) 0.10 \(\approx\) 0.12 DET_ML at 6–10.
- 地板被改正，"规律就压在地板上"被推翻（R138，FLOOR-RECOMPUTE，已复算）。R129/R131 的位置盲地板只注入了声明 \(\tau^2\) 的 0.634 倍（能段 4）/ 0.497 倍（能段 1）（MAD 封顶 + EB 收缩，R133 定位）；解析完整 \(\tau\) 场给出 \(\varepsilon\) 0.05 / 0.10 / 0.15 / 0.20 处能段 4 地板 0.165 / 0.469 / 0.710 / 0.856、能段 1 为 0.187 / 0.426 / 0.623 / 0.768。\(\varepsilon\) 0.10 处 D 与 H2（各 0.545）高于地板 0.469——R131 的读法与地板数字 0.914 / 0.654 自此作废。The floors are corrected, and "the laws sit at the floor" is overturned (R138, FLOOR-RECOMPUTE, replayed). R129's and R131's position-blind floors injected only 0.634\(\times\) (band 4) / 0.497\(\times\) (band 1) of the declared \(\tau^2\) (MAD cap + EB shrinkage, located by R133); the analytic full-\(\tau\) field gives band-4 floors 0.165 / 0.469 / 0.710 / 0.856 at \(\varepsilon\) 0.05 / 0.10 / 0.15 / 0.20 and band-1 0.187 / 0.426 / 0.623 / 0.768. D and H2 at \(\varepsilon\) 0.10 (0.545 each) sit above the floor of 0.469 — R131's reading and the floor numbers 0.914 / 0.654 are superseded.
- 打分手冻结为 v1（R147，SCORER-FREEZE，哈希已核对）。score_pooled_v1.py（sha256 7192c947…8515）：18 个格、index {1.5, 2, 2.5, 3}、面积环 [0, 4, 8, 14.01]、分半种子 20261020（34 拟合 / 33 留出）、读法 = 位置真值校准的 95 % CI。留出表（能段 4）：\(\varepsilon\) 0.05 处 D 0.303 / H2 0.182 / C 0.091 / D+SS 0.394（地板 0.165，天花板 0.951）；\(\varepsilon\) 0.10 处 D 0.545 / H2 0.545 / C 0.545 / D+SS 0.727（\(\Delta\) +0.182 [+0.030, +0.333]，地板 0.470）。此后所有位置规律得分都用 v1。The scorer is frozen as v1 (R147, SCORER-FREEZE, hash verified). score_pooled_v1.py (sha256 7192c947…8515): 18 cells, indices {1.5, 2, 2.5, 3}, area annuli [0, 4, 8, 14.01], split seed 20261020 (34 fit / 33 held), reading = the position-true-calibrated 95 % CI. Held table (band 4): at \(\varepsilon\) 0.05 D 0.303 / H2 0.182 / C 0.091 / D+SS 0.394 (floor 0.165, ceiling 0.951); at \(\varepsilon\) 0.10 D 0.545 / H2 0.545 / C 0.545 / D+SS 0.727 (\(\Delta\) +0.182 [+0.030, +0.333], floor 0.470). All future position-law scores use v1.
- 位置效应：不是源表，是像素级几何，两条能段共享；第一个拟合前位置项抬升 D。每个位置注入同一份 54 源栅格，源表成分只解释 0.2–1.2 % 的位置间方差（R126）；预登记 5 因子模型在不相交半位置上 \(R^2\) 0.46–0.51（置换 p 0.0005），能段 1 对能段 4 偏移相关 r = +0.936 [0.904, 0.961]（R134；R145：去掉拟合输出协变量后 3 因子仍以 \(R^2\) 0.35–0.49 迁移；"相位"实为二元 x 相位拆分）。导出的相位修正被驳回（R140：留出相关约 0，减去它更差）；把注入点移到像素中心把偏移 sd 从 0.114 放大到 0.230（R144，预登记 AMBIGUOUS，方向与预测相反，未解释）。第一个拟合前几何项（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 配对 \(\Delta\) +0.182 [+0.030, +0.333]；R148 诚实全流程零假设 p 0.003–0.004，不是相位主导）——尚未在新位置上检验，POS-FRESH 在跑。The position effect: not the source list, but pixel-level geometry shared by both bands; a first pre-fit position term lifts D. Every position injects the same 54-source grid, and composition explains only 0.2–1.2 % of the between-position variance (R126); the pre-specified 5-factor model predicts a disjoint half of positions at \(R^2\) 0.46–0.51 (permutation p 0.0005), and the band-1 vs band-4 offset correlates r = +0.936 [0.904, 0.961] (R134; R145: without fit-output covariates the 3-factor model still transports at \(R^2\) 0.35–0.49; the "phase" is really a binary x-phase split). The derived phase correction is rejected (R140: held correlation about 0, subtracting it is worse); moving injections to pixel centres amplified the offset sd from 0.114 to 0.230 (R144, pre-registered AMBIGUOUS, opposite of the prediction, unexplained). A first pre-fit geometry 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 paired \(\Delta\) +0.182 [+0.030, +0.333]; R148 honest full-pipeline null p 0.003–0.004, not phase-led) — not yet tested on fresh positions; POS-FRESH is running.
- 宽度：严格判据本身没校准；可观测箱级规律 V̄ 在校正判据下通过（等 PI 决定）。Li et al. 2025 的 \(O(n^{-1})\) 展开只在 b 0.01 精确（R121，R130 维持）；可观测尝试失败（R127 的 V_exp 4/8 箱超 2\(\sigma\)；R136 的校准 V2 5/8 失败）。R143 找到两层原因：逐行误差与 u = (Ŝ − S_true)/\(\sigma\) 相关（+0.65，箱内 Eddington 混合），且完美方差规律在"全部 8 箱都过"的判据下也只有 66.8 % 通过（失败箱 0/1/2/\(\geq\)3 为 69/26/4.5/0.4 %）。R146 的可观测箱级规律 V̄(Ŝ, b) = E[V2 | Ŝ, b] 在"8 箱允许 1 箱失败"的校正判据下于 \(\beta\) −1.5 / −1.8 通过（1/8；严格判据失败），DET_ML 20–50 处 4/4；同行比较 V_3p 8/8 失败、V_exp 2/8。判据是否更改等 PI；emldetect 迁移检验（WIDTH-TRANSFER）在跑。Width: the strict rule itself is mis-calibrated; the observable bin-level law V̄ passes under the calibrated rule (awaiting the PI). The Li et al. 2025 \(O(n^{-1})\) expansion is exact only at b 0.01 (R121, upheld by R130); the observable attempts fail (R127's V_exp 4/8 bins beyond 2\(\sigma\); R136's calibrated V2 fails 5/8). R143 finds two layers: the per-row error correlates with u = (Ŝ − S_true)/\(\sigma\) (+0.65, the within-bin Eddington mixture), and a perfect variance law passes the "all 8 bins" rule only 66.8 % of the time (failing bins 0/1/2/\(\geq\)3 in 69/26/4.5/0.4 %). R146's observable bin-level law V̄(Ŝ, b) = E[V2 | Ŝ, b] passes at \(\beta\) −1.5 / −1.8 under the calibrated "1 failing bin of 8 allowed" rule (1/8; it fails the strict rule), 4/4 at DET_ML 20–50; same-rows, V_3p fails 8/8 and V_exp 2/8. The rule change awaits the PI; the emldetect transfer test (WIDTH-TRANSFER) is running.
- 在跑（不引用数字）：PHASE-DIAG（解释像素中心臂的放大）、WIDTH-TRANSFER（V̄ 迁到 emldetect 行）、UNIVERSALITY-DESIGN（PI 2026-10-22 的问题：什么多样性证明规律普适；只做设计）、POS-FRESH（\(\geq\) 24 个新位置、冻结系数）、RT-112（审 V̄）、RT-113（审打分手 v1）。已归档：DET_ML 临时盘 879 MB 入 archive/scratch/（MANIFEST.md，119 行；删除清单待 PI）。Running (no numbers quoted): PHASE-DIAG (explaining the pixel-centre amplification), WIDTH-TRANSFER (V̄ onto emldetect rows), UNIVERSALITY-DESIGN (the PI's 2026-10-22 question: what diversity proves universality; design only), POS-FRESH (\(\geq\) 24 new positions, frozen coefficients), RT-112 (auditing V̄), RT-113 (auditing scorer v1). Archived: the DET_ML scratch, 879 MB, into archive/scratch/ (MANIFEST.md, 119 rows; the deletion list awaits the PI).
本次更新（第 19 版，2026-10-21）What's New (version 19, 2026-10-21)
- 专家复核改写打分手（R114，已采纳）：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。后果："宽度可以推广"必须读作位置内；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），不是 72 次独立试验；Monte-Carlo p 必须写成 k/N 加一个界（0/2000 \(\to\) \(\leq\) 0.0015 单侧 95 %）。An expert review rewrites the scorer (R114, accepted): the old scorer mixes 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. Consequences: "the width generalises" must 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), not 72 independent trials; and a Monte-Carlo p must be reported as k/N with a bound (0/2000 \(\to\) \(\leq\) 0.0015 one-sided 95 %).
- 宽度指标对逐位置偏移精确不变（R117，RT-98，已复算）：位置内二阶矩对任何常数位置偏移恒等不变（1e-15），所以宽度的零假设里结构性地没有位置效应——一条带 0.30 z 逐位置偏移（实测位置簇 sd 的 2.5 倍）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 %（不是专家给的 25–93 % 最坏情形）。"宽度可以推广"自此必须带"位置内"三个字。The width metric is exactly invariant to per-position offsets (R117, RT-98, replayed): the within-position second moment is identically invariant (1e-15) to any constant per-position offset, so the width null is structurally empty of the position effect — a law carrying a 0.30 z per-position offset (2.5\(\times\) the measured cluster sd) still passes width 70/72, while cov68 failures rise from 4 to 29. On this dataset the adaptive null widens D's Q by only about 6.5 % (not the expert's 25–93 % worst case). "The width generalises" must from now on carry the words "within position".
- 新打分手分开两个目标，D 在两个目标上都不是答案（R118，SCORER-CONTRACT v2，已复算）：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。旧打分数的 48/72 等读数自此降级为历史诊断。The new scorer separates two targets, and D is not the answer under either (R118, SCORER-CONTRACT v2, replayed): 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 old scorer's 48/72 and similar readings are demoted to historical diagnostics.
- 中心项的主导成分是样本内表征误差（R116，CENTRE-DECOMP + RT-99）：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；位置专属残差只在 3-6/2、6-10/0 等少数格出现。专家的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——"先验混合协方差解释 D 的 index 增长误差"不成立为残差解释。The dominant centre component is an in-sample representation error (R116, CENTRE-DECOMP + RT-99): within one 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; a position-specific residual survives only in a few cells (3-6/2, 6-10/0). 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 — so "the prior-mixing covariance explains D's index-growing error" does not survive as a residual explanation.
- 宽度读数被限定（R115，RT-96b，已复算）：0.976 的留出宽度只是一次偏低的分半（40 个半径分层分半平均 0.998）；6 参数修正只通过 25/72 个宽度格，与 D 的 68/72 不是同一句话。今后每个宽度数字都要带估计量与分半集合。The width reading is qualified (R115, RT-96b, replayed): 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 — not the same statement as D's 68/72. From now on every width number must name its estimator and split ensemble.
- PI 定下进度指标并批准 WIDTH-2ND（04，2026-10-21）：目标（05:39 EDT）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；11:42 EDT 改为没有位置阈值——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。WIDTH-2ND（06:20 EDT 批准，当天 14:16 EDT 完成，R121）把宽度亏缺归因为条件方差 \(\mathrm{Var}[\Delta C \mid \hat{\theta}] < V_{3p}(\hat{\theta})\)，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\(\sigma\)）。The PI set the progress metric and approved WIDTH-2ND (04, 2026-10-21): the 05:39 EDT target was per-position accuracy at most positions over the largest possible domain (e.g. all positions within 14′ off-axis), with quality cuts such as MASKFRAC allowed; at 11:42 EDT it became no 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. WIDTH-2ND (approved 06:20 EDT, finished 14:16 EDT the same day, R121) re-attributes the width deficit to the conditional variance \(\mathrm{Var}[\Delta C \mid \hat{\theta}] < V_{3p}(\hat{\theta})\); a new \(O(n^{-1})\) closed-form factor is exact at \(b = 0.01\) and under-corrects at \(b = 0.04/0.16\), so the pre-registered verdict is FAIL (3 of 8 bins beyond 2\(\sigma\)).
本次更新（第 18 版，2026-10-20）What's New (version 18, 2026-10-20)
- 一支验证过的评分器拒收全部四条候选规律（R103，NULL-1，已复算）：评分器先在"恰好为真"的数据上验证（p 均匀，K-S p 0.59 / 250 个重复），再打分：M2 能段 4 上 C 10/72、D 48/72、DQ 42/72、MIX3d 35/72，能段 1 上 D 重拟合 37/72；零假设均值约 62–63/72，D 的 p_pass = 0.0015。此前"能不能接受完全取决于零假设"的分歧就此结束。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。One validated scorer rejects every candidate law (R103, NULL-1, replayed): the scorer is validated first (uniform p on exactly-true data, K-S p 0.59 over 250 replicas) and then scores: on M2 band 4 C 10/72, D 48/72, DQ 42/72 and MIX3d 35/72, and on band 1 the D refit 37/72, against a null mean of about 62–63/72 with D's p_pass = 0.0015. The earlier disagreement, where acceptance depended entirely on the null, ends here. 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.
- 打分手本身错了两次并修好：R094 给出噪声底——一条完美的规律在同样的 72 格里平均也只通过约 61 格（第 5 百分位 48），所以 10/72 远在噪声底之下；R100 撤回 R096 的"D 以 p = 0 被全局拒绝"——那个零假设是单一 index 的 72 项而观测 Q 是 288 项，p = 0.000 不是 p 值（对合并零假设 p = 0.75 或 0.0025）。教训：任何全局 Q 的 p 值必须先核对 n_terms(观测) 是否等于 n_terms(零假设)。The scoring itself was wrong twice and fixed: R094 gives the noise floor — a perfect law passes only about 61 of the same 72 cells on average (5th percentile 48), so 10/72 is far below the floor; R100 withdraws R096's "D globally rejected at p = 0" — that null was a single-index 72 terms against a 288-term observed Q, so p = 0.000 is not a p-value (against a pooled null, p = 0.75 or 0.0025). Lesson: before any global Q's p-value goes into a record, check that n_terms(observed) equals n_terms(null).
- 宽度这一半可以推广，中心这一半不能（R104，RT-91b，已复算）：D 的 24 个均值失败是真的（24/24 超过 2 倍正确抽样噪声，23/24 超过 3 倍），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（强制放到留出半则塌到 8/12）；没有拟合/留出泄漏。The width half generalises, the centre half does not (R104, RT-91b, replayed): D's 24 mean failures are real (24/24 beyond 2\(\times\) the correct sampling noise, 23/24 beyond 3\(\times\)), all negative, concentrated at mid count and growing with index; the 12/12 high-count anchor hinges entirely on which half position 14 falls into (forcing it onto the held half collapses it to 8/12); there is no fit/held leakage.
- 位置不是缺失的中心变量（R105，CENTRE-POS，已复算）：平滑的加性位置项（离轴 / g_pos / 支撑 / 方位角）在留出半把均值失败从 24 个推到 58–72 个；逐位置自由偏移不能推广。唯一有用的形式是逐格中心截距（18 个参数）：失败 24 \(\to\) 11、负和 −2.28 \(\to\) −0.80，而离轴斜率不加分。负均值 z 是逐格（计数箱）中心偏置，部分由少数视场边缘异常位置驱动，不是平滑的视场几何趋势。Position is not the missing centre variable (R105, CENTRE-POS, replayed): a smooth additive position term (off-axis / g_pos / support / azimuth) pushes the held-out mean failures from 24 to 58–72; a per-position free offset does not generalise. The only form that helps is a per-cell centre intercept (18 parameters): failures 24 \(\to\) 11 and the negative sum −2.28 \(\to\) −0.80, while the off-axis slope adds nothing. The negative mean z is a per-cell (count-bin) centre bias, partly driven by a few FoV-edge outlier positions, not a smooth FoV-geometry trend.
- 能段 1 不可迁移（R102，已复算）：D 在 M2 能段 1 上重拟合 37/72，能段 4 系数原样搬过去 17/72（对完美规律约 60.5，第 5 百分位 52）；负的中心残差重复出现且更大。Band 1 does not transfer (R102, replayed): D refit on M2 band 1 passes 37/72 and the unchanged band-4 coefficients 17/72 (against a perfect law of about 60.5, 5th percentile 52); the negative centre residual repeats and is larger.
- 在跑（不引用数字）：符号回归试点（PySR 与 SRHarness，LLM 后端限 NIM 模型，PI 2026-10-20 18:47 EDT 批准，均有中期报告）、混合规律任务 LAW-HYBRID（D 的宽度 + MIX3d 的中心）、RT-95（审 NULL-1 的验证）。CENTRE-POS、RT-93 与 RT-94 已完成（R105、R106、R107）。Running (no numbers quoted): the symbolic-regression trial (PySR and SRHarness, LLM backend restricted to NIM models, approved by the PI at 18:47 EDT on 2026-10-20, both with interim reports), the hybrid-law task LAW-HYBRID (D's width + MIX3d's centre) and RT-95 (auditing NULL-1's validation). CENTRE-POS, RT-93 and RT-94 are finished (R105, R106, R107).
本次更新（第 17 版，2026-10-20）What's New (version 17, 2026-10-20)
- 主线回到 emldetect 结果（PI 2026-10-20）：「现在我不关心星表。还是要结果emldetect的结果。」任务 0b（在 L0 同一批 stamp 图上跑 emldetect、与 L0 配对比较）已完成（R089，已复算）：高计数两臂一致（DET_ML \(\geq\) 50：−0.008/0.978 对 −0.012/0.971），DET_ML ≲ 7 处 emldetect 中心比 L0 低 0.16–0.53 z、宽 0.05–0.14——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。主线下一步 T-LAW 在 L1 输出上建 p(DET_ML | Ŝ, b, index)（显式含 b）。The main line returns to the emldetect result (PI 2026-10-20). “现在我不关心星表。还是要结果emldetect的结果。” Task 0b (emldetect on the same L0 stamp images, paired against L0) is completed (R089, replayed): the two arms agree at high counts (DET_ML \(\geq\) 50: −0.008/0.978 vs −0.012/0.971) and at DET_ML ≲ 7 emldetect's centre is 0.16–0.53 z lower and its width 0.05–0.14 larger than L0 — the L0 low-count centre excess is not an emldetect property, and the single-image emldetect law must be built on emldetect output. The next main-line step T-LAW builds p(DET_ML | Ŝ, b, index) on L1 output (with explicit b).
- L0 低计数中心项超额是 L0 流水线自己的性质（R082/R084/R085；归因经 R087 撤回）：匹配对照下超额仍在，但 72/28 的归因（FFT 模板族/v2t1 对 plug-in 均值）已被 R087 撤回——配对重拟合只补 0–27 %，成因开放；按可观测 \(\hat{S}\) 分箱 +0.389（\(\hat{S}\) < 4）、+0.14–0.18（4–11）、+0.058 \(\pm\) 0.007（\(\geq\) 11），高计数锚约 0。The L0 low-count centre excess is a property of the L0 pipeline itself (R082/R084/R085; the attribution is withdrawn by R087): the excess survives the matched control, but the 72/28 split (FFT family/v2t1 against plug-in mean) is withdrawn by R087 — the paired refit closes only 0–27 %, cause open; in observable \(\hat{S}\) bins +0.389 (\(\hat{S}\) < 4), +0.14–0.18 (4–11), +0.058 \(\pm\) 0.007 (\(\geq\) 11), about 0 at the high-count anchor.
- penalty 错配机制也解释 L1 宽度（R086 推翻 R083）：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，已推翻），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）；同一 \(V_{3p}\) 机制同时决定 L0 与 emldetect 的宽度。The penalty-mismatch mechanism does explain the L1 width (R086 reverses R083): 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, reversed), 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 same \(V_{3p}\) mechanism carries both the L0 and the emldetect widths.
- 配对实验 0b 完成（R089，已复算）：同一批 bit-exact L0 图像上，emldetect 与 L0 高计数一致（模型典型 DET_ML \(\geq\) 50：−0.008/0.978 对 −0.012/0.971），DET_ML ≲ 7 处 emldetect 中心比 L0 低 0.16–0.53 z、宽 0.05–0.14——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。Paired experiment 0b completed (R089, replayed): on the same bit-exact L0 images emldetect and L0 agree at high counts (model-typical DET_ML \(\geq\) 50: −0.008/0.978 vs −0.012/0.971) and at DET_ML ≲ 7 emldetect's centre is 0.16–0.53 z lower and its width 0.05–0.14 larger than L0 — the L0 low-count centre excess is not an emldetect property, and the single-image emldetect law must be built on emldetect output.
- index 规律须显式含 b（R088/R090，已复算）：中心对 index 线性（残差 \(\leq\) 0.003），但 a、b 随背景 b 变（10–20 斜率 −0.27、\(\geq\) 50 截距 +0.16，b 0.4 对 0.0125）；宽度结论依赖分箱约定（typ_bg 主约定下 10–20 为 +0.041 [+0.013, +0.080]，20–50 并非 index-free）。公共 2 参数规律被否定；T-LAW 建 p(DET_ML | Ŝ, b, index) 并检验平滑 b 参数化。The index law must carry b explicitly (R088/R090, replayed): the centre is linear in index (residual \(\leq\) 0.003) but a and b vary with background b (10–20 slope −0.27, \(\geq\) 50 intercept +0.16 between b 0.4 and 0.0125); the width result depends on the bin-label convention (under typ_bg primary, 10–20 is +0.041 [+0.013, +0.080] and 20–50 is not index-free). A common 2-parameter law is refuted; T-LAW builds p(DET_ML | Ŝ, b, index) and tests a smooth-in-b form.
- 按计数修正有条件成立（R091，RT-84 分诊）：在 emldetect 输出上按 Ŝ 箱拟合的 6 参数修正给留出 0.976 [0.956, 0.993]（DET_ML 6–10）；换独立模拟（8,640 行）过修（1.032/1.061）；“拟合器是原因”未证明。目标依赖估计量定义（逐格 0.908/0.919 对合并 0.940）。The count-dependent correction holds conditionally (R091, RT-84 triaged): a 6-parameter per-Ŝ-bin correction fitted on emldetect output gives held-out 0.976 [0.956, 0.993] at DET_ML 6–10; transferred from an independent 8,640-row simulation it over-corrects (1.032/1.061); "the fitter is the cause" is not shown. The target is estimand-dependent (within-cell 0.908/0.919 vs marginal 0.940).
- 计数依赖修订（R081）：共同掩膜上固定 \(\hat{S}\) 的错配差 +0.80 pp（\(\hat{S}\) 2.77 与 4.09）；R080 的“+0.86 pp 在 \(\hat{S}\) 约 5”是掩膜假象。Count dependence revised (R081): on a common mask the fixed-\(\hat{S}\) mismatch gap is +0.80 pp (\(\hat{S}\) 2.77 and 4.09); R080's “+0.86 pp at \(\hat{S}\) about 5” was a mask artefact.
- 物理页第 15 节算例更正：term1 = 63.6、Var(R) = 1.90、错配约 1.4 % 对应 \(V_{3p}(\hat{S})/Var(R)\) = 1.44（约 44 % 方差高估），不是 1.11。Physics §15 worked example corrected: term1 = 63.6, Var(R) = 1.90 and a mismatch of about 1.4 % give \(V_{3p}(\hat{S})/Var(R)\) = 1.44 (about a 44 % variance over-statement), not 1.11.
- 专家结论：目标未达成（astra，2026-10-20T03:25Z）：接受 P0（0b 验收标准）与 P1（0b 试点后的验收清单、emldetect 条件律表）；联合律（P2）已由 PI 决定：主线是单图 emldetect 条件律 p(DET_ML | Ŝ, b, index)（04，2026-10-20 04:51 EDT）。Expert verdict: goals not reached (astra, 2026-10-20T03:25Z): P0 (0b acceptance criteria) and P1 (the post-0b acceptance manifest and the emldetect conditional-law table) are accepted; the joint law (P2) is settled by the PI: the main line is the single-image emldetect conditional law p(DET_ML | S-hat, b, index) (04, 2026-10-20 04:51 EDT).
本次更新（第 16 版，2026-10-19）What's New (version 16, 2026-10-19)
- 真实 PSF 偏窄的机制被重新标注（R064）：公式在真实参数处是对的——固定 S_true 与真实位置时，残差的 std 除以 \(\sqrt{V_{3p}(S_{true})}\) 是 1.12 / 1.07 / 1.03 / 1.01 / 1.03（DET_ML 3 / 5 / 9 / 15 / 30，从不小于 1）。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（Ŝ 在 DET_ML 3 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），再加一个箱内 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（DET_ML 0.85 处 mean z +1.39）。The real-PSF narrowing's mechanism is re-labelled (R064): the formula is right at the true parameters — at fixed S_true and the true position the residual's std divided by \(\sqrt{V_{3p}(S_{true})}\) is 1.12 / 1.07 / 1.03 / 1.01 / 1.03 (DET_ML 3 / 5 / 9 / 15 / 30, never below 1). The observable-bin narrowing comes from evaluating the convex \(V_{3p}\) at a biased-high \(\hat{S}\) (\(\hat{S}\) is high by +14.5 % at DET_ML 3 and +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 (mean z +1.39 at DET_ML 0.85).
- R064 提示的偏差校正修正被推翻（R068）：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 从 0.88 / 0.92 / 0.94 / 0.96 只升到 0.975 / 0.959 / 0.970 / 0.979（消掉 45–79 % 的偏窄），回不到 1 \(\pm\) 0.02；DET_ML \(\lesssim\) 1 时该校正没有定义。08 §7 第 0a 项以否定结论关闭，目前没有任何可观测的点修正能闭合真实 PSF 的宽度。The bias-corrected correction R064 suggested is refuted (R068): evaluating \(V_{3p}\) at the observable Cox–Snell bias-corrected amplitude \(\hat{S} - b_S(\hat{S})\) lifts std z only from 0.88 / 0.92 / 0.94 / 0.96 to 0.975 / 0.959 / 0.970 / 0.979 at DET_ML 6.9 / 9 / 12 / 18.5 (45–79 % of the deficit removed) and not to 1 \(\pm\) 0.02; below DET_ML \(\lesssim\) 1 the correction is undefined. 08 §7 item 0a is closed negative, and no observable point correction closes the real-PSF width yet.
- 这条规律在 67 个位置上都成立（R065）：L0 生产数据（离轴 0–13′，我们的全局搜索拟合器）上，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。The law holds at all 67 positions (R065): on the L0 production data (off-axis 0–13', our global-over-support fitter) it sits \(\approx\) 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）没有这条低端偏窄；L1 与 L0 的差是两条流水线之间的差（R066、R067）：L1 按可观测 S_eml 分箱后 std z 0.94–0.98（DET_ML 1–6）、\(\approx\) 0.95（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（0.95 / 0.975 / 1.00 / 1.02 于 DET_ML 6–10 / 10–20 / 20–50 / 50–100）降为历史，只引用可观测箱。匹配比较（同 67 位置、同 S^−2.5 密度、同 (S, b) 格子权重）下 L1 − L0 = +0.24 / +0.19 / +0.12 / +0.08 / +0.07 于 DET_ML \(\approx\) 2 / 3 / 4 / 5 / 7（约 50 处持平）；但两边用的是不同打分（真实 PSF 模板族在拟合位置 对 e114 替代核在名义位置）与不同振幅估计量（S_true 2.5 时 \(E[\hat{S}]/E[S_{eml}]\) = 1.38，20 以上 0.92），所以这个差不归因于拟合器。L1 (emldetect) does not show this low-end narrowing, and the L1–L0 gap is between two pipelines (R066, R067): binned by the observable S_eml, L1's std z is 0.94–0.98 at DET_ML 1–6, \(\approx\) 0.95 at 6–10 and about 1 above 12–20, with the lowest bin wide (1.31); T228's b_true-labelled cell numbers (0.95 / 0.975 / 1.00 / 1.02 at DET_ML 6–10 / 10–20 / 20–50 / 50–100) are demoted to history and only the observable bins are quoted. In the matched comparison (the same 67 positions, the same S^−2.5 density, the same (S, b) cell masses) L1 − L0 = +0.24 / +0.19 / +0.12 / +0.08 / +0.07 at DET_ML \(\approx\) 2 / 3 / 4 / 5 / 7 (level by about 50); but the two sides use different scorings (the stamp family at the fitted offset against the e114 surrogate at the nominal position) and different amplitude estimands (\(E[\hat{S}]/E[S_{eml}]\) = 1.38 at S_true 2.5 and 0.92 at 20 and above), so the gap is not attributable to the fitter.
- 未解决问题清单更新：RT-60 至 RT-64 已分诊（R064–R068），RT-65 与 RT-66 也已分诊（R070、R069）：偏窄与逐行 \(\delta\) 约定无关（std 变化 \(\leq\) 0.005，DET_ML \(\approx\) 2 以上），也不是位置混合（\(\hat{r}\) \(\lt\) 0.5 px 的行照样偏窄）；R068 剩下的残差是先验无关的凸性，不是箱内 \(S_{true}\) 混合（三种先验下校正后 std z 变化 \(\leq\) 0.020）。RT-68 也已分诊（R071）：67 个 L0 位置上宽度随先验合并移动 \(\leq\) 0.015（2–4 SE，陡先验更窄），窄 \(\hat{S}\) 切片达 0.023、单个位置组 0.027，范围依赖（[2.5, 100] 上 0.052）——所以 R063/R069 的“宽度与先验无关”改述为小而非零（注册打分下 \(\leq\) 0.02 的系统项，引用时要带 S 范围与打分）。RT-67（校准偏差与 \(E[V \mid \hat{S}]\)）在跑；08 §7 第 0a 项以否定结论关闭，第 0b 项（在 L0 真实 PSF 图像上跑 emldetect 的决定性实验）是新的开放问题；真实 PSF 宽度配料研究（第 0 项）仍开放。Open problems updated: RT-60 through RT-64 are triaged (R064–R068) and so are RT-65 and RT-66 (R070, R069): the narrowing is independent of the per-row \(\delta\) convention (std change \(\leq\) 0.005 above DET_ML \(\approx\) 2) and is not a position mixture (rows at \(\hat{r}\) \(\lt\) 0.5 px are equally narrow); R068's remaining residual is prior-independent convexity, not a within-bin \(S_{true}\) mixture (corrected std z moves \(\leq\) 0.020 across three priors). RT-68 is triaged too (R071): across 67 L0 positions the width moves \(\leq\) 0.015 pooled over the priors (2–4 SE, steeper is narrower), up to 0.023 in narrow \(\hat{S}\) slices and 0.027 in one position group, and it is range-dependent (0.052 on [2.5, 100]) — so R063/R069's “the width is prior-independent” is restated as small, not zero (a \(\leq\) 0.02 systematic for the registered scoring; quote it with the S range and the scoring). RT-67 (calibrated bias and \(E[V \mid \hat{S}]\) ) is running; 08 §7 item 0a is closed negative and item 0b (the decisive run of emldetect on the L0 real-stamp images) is the new open problem; the real-PSF width ingredient study (item 0) stays open.
- \(\sigma_{ratio}\) 的 L0 近行通过被质疑（R054、R057）：\(\sigma_{ratio}\)'s L0 near-row pass is questioned (R054, R057):换成拟合模板族的精确位置导数后它在 DET_ML 6–10 过度修正（std z 1.67 / 1.83，两臂）；RT-54 用 5 点模板（1.31）、h = 2（0.81）与精确导数（1.68）证明原先的"18/19 格通过"只是 np.gradient 3 点差分里的偶然抵消。它是经验校准，不是推导出来的宽度。with the fitter template family's exact position derivatives it over-corrects at DET_ML 6–10 (std z 1.67 / 1.83, two arms); RT-54 showed with the 5-point stencil (1.31), h = 2 (0.81) and the exact derivatives (1.68) that the old "18 of 19 cells pass" was only an accidental cancellation inside np.gradient's 3-point stencil. It is an empirical calibration, not a derived width.
- 玩具低计数偏窄改写成可观测律（R059）：The toy low-count narrowing is rewritten as an observable law (R059):按可观测 \(\hat{S}\) 分箱（行在一个 \(S_{true}\) 先验上汇成总体）后，核 r̂ \(\leq\) 4 px 的 std z 在 DET_ML 1–3 是 0.82–0.86、\(\approx\) 5 是 0.93–0.95、\(\approx\) 7.5 是 0.94–0.96、\(\approx\) 12 及以上回到 1（1 SE 内）；核的均值在 DET_ML 1–3 是 +0.29。所以"窄 2–4 % 于 DET_ML 8–14"是固定 \(S_{true}\) 格子的效应。binned by the observable \(\hat{S}\) (rows pooled into one population over an \(S_{true}\) prior), the core r̂ \(\leq\) 4 px std z is 0.82–0.86 at DET_ML 1–3, 0.93–0.95 at \(\approx\) 5, 0.94–0.96 at \(\approx\) 7.5 and back to 1 (within 1 SE) at \(\approx\) 12 and above, with a core mean of +0.29 at DET_ML 1–3. So "narrow by 2–4 % at DET_ML 8–14" was a fixed-\(S_{true}\) cell effect.
- 真实 PSF：T219 被 R056 重新限定，T234 由 R062 改写为可观测律（Provisional）：Real PSF: T219 is rescoped by R056 and T234 is rewritten by R062 as an observable law (Provisional):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 \(V(S_{true})\) 才失败（IRON-RULE 失误）。位置自由（\(\nu\) = 3，样条模板族，全局网格 + NM，一个源位置，总体 S_true ~ dN/dS \(\propto\) S^−2.5 于 [5.58, 51.20]，N = 50,000）：按可观测 \(\hat{S}\) 分箱，插入式 \(V_{3p}(\hat{S})\) 高估宽度——std z 0.73（DET_ML \(\lesssim\) 2）、0.79（3）、0.83（4）、0.85（5.3）、0.88（6.9）、0.92（9）、0.94（12）、0.96（18.5）、\(\approx\) 1（42）；偏窄在 DET_ML \(\approx\) 17–22 以下都超过 2 SE，且窄 \(\hat{S}\) 切片（relSE 2–3 %）给出同样的值；低 DET_ML 处还右偏（最低箱 P(z > 2) = 19.9 %）。已删除（R064）：“\(V_{3p}(\hat{S})\) 在方差上最多高估约 1.9 倍”——那是一个均值被移过的分布的标准差比（DET_ML 0.85 处 mean z +1.39），R064 已把它删掉。逐行 \(\delta_{3p}\) 之后的均值是 +0.28 / +0.18 / +0.09 / 0.00 / −0.05 z（DET_ML 0.75 / 3 / 5.3 / 9 / 18–42）。Claude 的先验边界与斜率检验（S_true \(\geq\) 7 / \(\geq\) 8.5、斜率 −2.5 \(\to\) −1.5）每箱都在 1 SE 内不变，所以这是给定 \(\hat{S}\) 的律，不是先验截断；0.5 倍背景在匹配 DET_ML 的 4/5 箱内与 2 SE 一致。固定 \(S_{true}\) 格子的 0.914 / 0.955（DET_ML 8 / 15）与 0.992–1.000（30–600）只作历史。RT-60 已分诊（R064，见上方本次更新）。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 \(V(S_{true})\) is what failed (an IRON-RULE slip). With the position free (\(\nu\) = 3, a spline template family, a global grid + NM, one source position, a population S_true ~ dN/dS \(\propto\) S^−2.5 on [5.58, 51.20], N = 50,000): binned by the observable \(\hat{S}\), the plug-in \(V_{3p}(\hat{S})\) over-states the width — std z 0.73 (DET_ML \(\lesssim\) 2), 0.79 (3), 0.83 (4), 0.85 (5.3), 0.88 (6.9), 0.92 (9), 0.94 (12), 0.96 (18.5), \(\approx\) 1 (42); the deficit is above 2 SE through DET_ML \(\approx\) 17–22, and narrow \(\hat{S}\) slices (relSE 2–3 %) give the same values; at low DET_ML it is also right-skewed (P(z > 2) = 19.9 % in the lowest bin). Dropped (R064): “\(V_{3p}(\hat{S})\) over-states the variance by up to about 1.9\(\times\)” — that was the std ratio of a shifted distribution (mean z +1.39 at DET_ML 0.85) and R064 removes it. After the per-row \(\delta_{3p}\) the mean is +0.28 / +0.18 / +0.09 / 0.00 / −0.05 z (DET_ML 0.75 / 3 / 5.3 / 9 / 18–42). Claude's prior-edge and slope test (S_true \(\geq\) 7 / \(\geq\) 8.5, slope −2.5 \(\to\) −1.5) leaves every bin within 1 SE, so it is a law given \(\hat{S}\), not a prior truncation; the 0.5\(\times\) background agrees within 2 SE in 4 of 5 matched-DET_ML bins. The fixed-\(S_{true}\) cells' 0.914 / 0.955 (DET_ML 8 / 15) and 0.992–1.000 (30–600) are history only. RT-60 is attacking this new law.
- W5 关闭（判据未达成），RT-52 至 RT-55 分诊完毕，RT-56 与 RT-57 发出；W5 is closed (the key is not met), RT-52 to RT-55 are triaged and RT-56 and RT-57 are launched;W6 作为独立人口任务被撤掉（RT-55 已做）。新缺陷 B009：lib/beyond_fisher 的方差项漏掉泊松 \(\kappa_4\)，其 mean_curvature 是错的，E051 的均值侧结论不能当已确立。W6 is dropped as a separate population task (RT-55 did it). New defect B009: lib/beyond_fisher's variance term omits the Poisson \(\kappa_4\) and its mean_curvature is wrong, so E051's mean-side conclusion cannot be quoted as established.
- 新增：为什么按真实流量分箱会把偏窄放错位置（physics.html）New: why binning by the true flux puts the narrowing in the wrong place (physics.html)固定 \(S_{true}\) 的格子与可观测 \(\hat{S}\) 箱的对照，附新图 fig_core_vs_ml.png。fixed-\(S_{true}\) cells versus observable \(\hat{S}\) bins, with the new figure fig_core_vs_ml.png.