词汇表与阅读清单Glossary & Reading List

符号Symbols

ΔC（DeltaC）ΔC（DeltaC）

似然比统计量：C_null − C_min ≥ 0。emldetect 内部真正计算的检测量；项目的一切理论都在 ΔC 空间做。The likelihood-ratio statistic: C_null − C_min ≥ 0. The actual detection quantity computed inside emldetect; all project theory lives in ΔC space.

DET_MLDET_ML

检测似然：−ln Q(ν/2, ΔC/2)，自然对数。4XMM 目录的入选阈值是 6。注意 ln 不是 log10；Q 的第二参数是 ΔC/2。Detection likelihood: −ln Q(ν/2, ΔC/2), natural log. The 4XMM catalogue acceptance threshold is 6. Note ln is not log10; Q's second argument is ΔC/2.

Ŝ（S_hat）Ŝ（S_hat）

拟合出的源流量（counts）。理论的条件化变量；catalog 里 = RATE × EXP。The fitted source flux (counts). The theory's conditioning variable; in the catalogue = RATE × EXP.

S（S_true）S（S_true）

真实流量。只在模拟里知道；理论中只能通过声明的先验 π 积分进入，永远不作逐行输入。True flux. Only known in simulation; in theory it enters only through the declared prior π, never as a row-by-row input.

bb

背景，单位 counts/像素（4″ 像素）。catalog 里取 BG 列；模拟里是 BG_MAP（BG_MAP_EML）。Background, units counts/pixel (4″ pixels). In the catalogue it comes from the BG column; in simulation from BG_MAP (BG_MAP_EML).

ν（nu）ν（nu）

χ² 自由度 = n_images + 2（位置自由时）。单图像 ν = 3，shape = ν/2 = 1.5。χ² degrees of freedom = n_images + 2 (when position is free). Single image ν = 3, shape = ν/2 = 1.5.

Q(a, x)Q(a, x)

上不完全伽马函数（gammaincc）：P(χ²(2a) > 2x)。Upper incomplete gamma function (gammaincc): P(χ²(2a) > 2x).

μ₃p（mu_3p）μ₃p（mu_3p）

3 参数拟合 ΔC 的条件均值（Asimov/期望），由 PSF 模板逐像素算：2 Σ (μ ln(μ/b) − (μ−b))。Conditional mean of ΔC for the 3-parameter fit (Asimov/expected), computed pixel by pixel from the PSF template: 2 Σ (μ ln(μ/b) − (μ−b)).

V₃pV₃p

3 参数拟合 ΔC 的 Fisher 投影方差（近似高计数极限）。Fisher-projected variance of ΔC for the 3-parameter fit (approximate high-count limit).

δ（delta_E086）δ（delta_E086）

条件均值修正：E[ΔC | θ̂] − μ₃p。E090 的 δ₁ = 3 + ε_Lawley − ½tr(H_μ I⁻¹) − ∇μ₃p·bias_CS 是零系数的解析形式。Conditional-mean correction: E[ΔC | θ̂] − μ₃p. E090's δ₁ = 3 + ε_Lawley − ½tr(H_μ I⁻¹) − ∇μ₃p·bias_CS is the zero-coefficient analytic form.

G² = 1 − c20/μ₃pG² = 1 − c20/μ₃p

零参数宽度律；c20 = 1.794461 是 E140 推导的常数。Zero-parameter width law; c20 = 1.794461 is the constant derived in E140.

zz

标准化残差 (ΔC_obs − μ_model) / σ_model。理论正确时 z ~ N(0,1)。Standardized residual (ΔC_obs − μ_model) / σ_model. If the theory is right, z ~ N(0,1).

ww

联合拟合中某图像的位置 Fisher 信息占比 Φᵢ/ΣⱼΦⱼ。二维损失 L_2D = tr(Φᵢ(Φᵢ⁻¹ − Φ_tot⁻¹)) ≈ 2(1−w)。The share of the position Fisher information in a joint fit, Φᵢ/ΣⱼΦⱼ. The two-dimensional loss L_2D = tr(Φᵢ(Φᵢ⁻¹ − Φ_tot⁻¹)) ≈ 2(1−w).

π（prior）π（prior）

源计数先验 dN/dS ∝ S^−α。v2 要求每个评分都声明 π；α 网格 0…2.5 + 天空先验。Source-count prior dN/dS ∝ S^−α. v2 requires every score to declare π; α grid 0…2.5 plus sky prior.

HPDHPD

半功率直径（half power diameter）：PSF 包含一半功率的直径，XMM PN 约 15″。Half-power diameter: the diameter containing half the PSF power; XMM PN ~ 15″.

A1 / A2 / A3 / A4A1 / A2 / A3 / A4

基准的四个轴：均值偏差、宽度偏差、过阈值校准（L=6/8/10）、参数个数。The benchmark's four axes: mean bias, width bias, threshold calibration (L=6/8/10), parameter count.

CRPS / PITCRPS / PIT

连续分级概率评分 / 概率积分变换：分布校准的汇总诊断（报告但不进前沿）。Continuous ranked probability score / probability integral transform: distribution-calibration summary diagnostics (reported but not entering the front).

缩写与软件Acronyms & Software

XMM-NewtonXMM-Newton

欧空局 X 射线望远镜；三台相机：PN、MOS1、MOS2。ESA's X-ray telescope; three cameras: PN, MOS1, MOS2.

SASSAS

XMM 官方分析软件（Science Analysis System）。只在 virgo 上运行。XMM's official analysis software (Science Analysis System). Runs only on virgo.

emldetectemldetect

SAS 的最大似然 PSF 拟合任务；输出 DET_ML、位置、流量。对 15 幅图像（3 相机 × 5 波段）联合拟合一个位置。SAS's maximum-likelihood PSF fitting task; outputs DET_ML, position, flux. Jointly fits one position across 15 images (3 cameras × 5 bands).

eboxdetecteboxdetect

emldetect 上游的滑窗候选源粗筛。emldetect's upstream sliding-window candidate pre-filter.

edetect_stackedetect_stack

多观测叠加检测（DR14s 用；eml_ecut 默认 15）。Stacked multi-observation detection (used in DR14s; eml_ecut defaults to 15).

esensmapesensmap

灵敏度图/上限任务；官方承认其 Poisson 近似在阈值附近不准。Sensitivity map / upper-limit task; the official documentation admits its Poisson approximation is inaccurate near the threshold.

esplinemapesplinemap

样条背景图任务（DR14 背景配方：fitmethod=smooth, scut=2e-4, snrmin=12, smoothsigma=10）。Spline background map task (DR14 background recipe: fitmethod=smooth, scut=2e-4, snrmin=12, smoothsigma=10).

psfgenpsfgen

SAS 的 PSF 生成任务；level=ELLBETA 是项目模板的来源。SAS's PSF generation task; level=ELLBETA is the source of the project's templates.

4XMM / DR144XMM / DR14

第 4 次 XMM 巡天源目录的第 14 个数据发布（Webb et al. 2020 是 DR9；DR14 由同一巡天项目发布）。The 14th data release of the 4XMM survey catalogue (Webb et al. 2020 is DR9; DR14 is from the same survey project).

OBS_IDOBS_ID

观测编号；catalog 层的两级自助聚类单元。Observation ID; the catalogue layer's two-level bootstrap clustering unit.

epoch（观测年代，OBS_ID 的代理）epoch (an OBS_ID proxy)

观测的时间。4XMM 行没有直接的年代列，分析里用 OBS_ID 的顺序（十分位/三分位）作代理。T11_023：全部 12 个配置的乘性亏损 m 都随年代下降。When an observation was taken. 4XMM rows have no direct epoch column, so analyses use OBS_ID ordering (deciles/tertiles) as a proxy. T11_023: the multiplicative deficit m falls with epoch on all 12 configs.

MJD_STARTMJD_START

观测的真实开始日期（修正儒略日），来自 FITS 文件的 TCOMM。T11_025 起用它代替 OBS_ID 作年代代理，把 m 按真实年份拟合漂移/阶跃。The real start date of an observation (Modified Julian Date), quoted from the FITS TCOMM. From T11_025 on, it replaces OBS_ID as the epoch proxy, fitting m against the real year for drift/step.

漂移 / 阶跃（Δχ² 判读）drift / step (Δχ² reading)

m 随时间变化的两种形式判读：漂移 = 线性斜率；阶跃 = 在某个仪器事件处的跳变。用 Δχ² 比较两个模型：阶跃比漂移好 Δχ² > 9 才记为阶跃。T11_025 评审更正：M1 = 阶跃（2012-12，MOS1 CCD3 损失，−0.016），M2/PN = 漂移（约 −0.001/年）。The two reading forms for m's time behaviour: drift = a linear slope; step = a jump at an instrument event. Compared by Δχ²: the step beats the drift only if Δχ² > 9. T11_025 review correction: M1 = step (2012-12, MOS1 CCD3 loss, −0.016), M2/PN = drift (~−0.001/yr).

E114 核网格E114 kernel grid

μ₃p、V₃p、δ 在 (S, b) 上的预计算网格。原网格只覆盖 S = 1–400，超出部分线性外推——这导致 pred ≳ 2000 时核失效（T5_009）。T5_010/011 把全部 15 个网格扩展到 S = 5000（E116 生成器，只改 S 轴；旧节点逐一复现，差值 0.0），新模拟 |frac| < 0.01 到 S = 4000。The precomputed grid of μ₃p, V₃p, δ over (S, b). The original grid covered only S = 1–400, extrapolated linearly above — this caused the kernel failure at pred ≳ 2000 (T5_009). T5_010/011 extended all 15 grids to S = 5000 (E116 generator, only the S axis changed; every old node reproduces to 0.0), with fresh simulations giving |frac| < 0.01 up to S = 4000.

erfcxerfcx

缩放补余误差函数 erfcx(x) = e^(x²)·erfc(x)。对 ν = 3 的 ML 换算有精确恒等式 ln Q(3/2,x) = −x + ln((2/√π)√x + erfcx(√x))，x = ΔC/2——用它在 ML ≳ 745 处不溢出（T5_013），替代会下溢的 exp(−ML) 路径。The scaled complementary error function, erfcx(x) = e^(x²)·erfc(x). For the ν = 3 ML transform there is an exact identity ln Q(3/2,x) = −x + ln((2/√π)√x + erfcx(√x)), x = ΔC/2 — used to avoid the overflow at ML ≳ 745 (T5_013), replacing the exp(−ML) path that underflows.

管线版本（PPSVERS）pipeline version (PPSVERS)

生成星表产品所用的 SAS 管线版本。4XMM DR14 用管线 18/19，而档案（NXSA）现在提供 2024 年重处理产品（PPSVERS 21.51，SAS 21）。T13 试点发现：同一观测的新旧产品不能直接混用——0512180101 的图像计数约为 DR14 星表值的 3.5 倍。The SAS pipeline version that produced the catalogue products. 4XMM DR14 used pipeline 18/19, while the archive (NXSA) now serves the 2024 reprocessing (PPSVERS 21.51, SAS 21). The T13 pilot found: products of the two versions cannot be mixed directly — 0512180101's image counts are about 3.5× the DR14 catalogue values.

m_newcat / m_localm_newcat / m_local

T13 的三个每源测量量中的两个（第三个是 m_cat）。m_cat：DR14 数据、DR14 管线；m_newcat：2024 图像、2024 管线自己的 ΔC（用稳定反解）；m_local：2024 图像、我们的冻结拟合器。m_newcat 对 m_local（同一批图像）把星表的拟合与处理 (c) 从数据 (b) 分开；m_cat 对 m_newcat 量出重处理改变了什么。Two of T13's three per-source measurements (the third is m_cat). m_cat: DR14 data, DR14 pipeline; m_newcat: 2024 images, the 2024 pipeline's own ΔC (via the stable inversion); m_local: 2024 images, our frozen fitter. m_newcat vs m_local (same images) isolates the catalogue's fitting and processing (c) from the data (b); m_cat vs m_newcat measures what the reprocessing changed.

bootstrap 区间类型（percentile / basic / BCa）bootstrap interval types (percentile / basic / BCa)

配对差的置信区间算法。percentile：直接取 bootstrap 分布的分位数；basic（反百分位）：2θ̂ − q，对有偏统计量更稳健；BCa：再加偏度和加速度校正。T5_014 发现绝对值统计量之差的 percentile 区间在小窗口有偏（不含点估计），T5 第 18 阶段改报 basic/BCa 并注明主区间。Confidence-interval algorithms for paired differences. percentile: take quantiles of the bootstrap distribution directly; basic (reverse-percentile): 2θ̂ − q, more robust for biased statistics; BCa: adds bias and acceleration corrections. T5_014 found the percentile interval of a difference of absolute-value statistics is biased in small windows (does not contain the point estimate); T5 stage 18 switches to basic/BCa and states the headline.

红队（red team）red team

受派攻击本项目承重结论的独立执行者：给它一条结论和记录文件，让它找反例和漏洞；主代理（Claude）重跑其检查后分诊每条意见（实质性 / 次要 / 误报 / 重复），分诊可以推翻原结论。2026-10-15 起队列保持 2–4 个活跃红队（cognition-os/REDTEAM_QUEUE.md），第 1 轮 11 队（RT-01 到 RT-11）。战果示例：RT-05 发现 T13 位置文件 bug，推翻了"约一半来自共享位置"的读数。An independent executor assigned to attack the project's load-bearing claims: given a claim and its record file, it hunts counterexamples and holes; the main agent (Claude) re-runs its checks and triages each objection (material / minor / false alarm / duplicate), and a triage can overturn the claim. Since 2026-10-15 the queue keeps 2–4 active red teams (cognition-os/REDTEAM_QUEUE.md); round 1 has 11 teams (RT-01 through RT-11). Example result: RT-05 found the T13 position-file bug and overturned the "about half from the shared position" reading.

50% 完备度（S_50）50% completeness (S_50)

探测流程以 50% 概率发现源的计数阈值 S_50——灵敏度比较的正确口径（RT-08：固定 Ŝ 的均值越过检出阈值不是）。算例（用 RT-08 的真实数字）：某方法在 ML = 6 处宣称的阈值是 S = 14.34，除以 E094 模拟的真实 S_50 得到比值 1.15，即该方法比真实 50% 完备度阈值高 15%；附录公式 1.15、kernel 1.24、esensmap deltac 1.30、esensmap Poisson 1.82（中位数）；ML = 20 时附录公式约 1.01——原表"附录公式与数据 1.004"是两种相反偏差的抵消。初步，图修正 3 运行中。The count threshold S_50 at which a detection pipeline finds a source with 50% probability — the correct basis for sensitivity comparisons (RT-08: a mean crossing the detection threshold at fixed Ŝ is not). Worked example (with RT-08's real numbers): a method's claimed threshold at ML = 6 is S = 14.34; dividing by the true S_50 from the E094 simulations gives the ratio 1.15, i.e. the method sits 15% above the true 50% completeness threshold; appendix formula 1.15, kernel 1.24, esensmap deltac 1.30, esensmap Poisson 1.82 (medians); at ML = 20 the appendix formula is about 1.01 — the old table's "appendix vs data 1.004" was a cancellation of two opposite biases. Preliminary; figure correction 3 is running.

MASKFRACMASKFRAC

未遮挡比例：4XMM 中是探测掩膜在源位置附近的 PSF 加权平均（T11_023 评审更正——不是"被遮挡的比例"）。M2 b4 中位数 0.981、IQR 0.961–0.986；本项目的干净样本约定：T109 用 ≥ 0.97，较早的 T56 用 > 0.8（来源：cognition-os/03_TRUST_TABLE.md）。高 MASKFRAC = 完全覆盖；部分遮挡源的 MASKFRAC 低。两组只差约 0.02，不足以直接造成 0.05 的 m 差异——它更可能是年代或探测器位置的代理变量。The unmasked fraction: in 4XMM it is the PSF-weighted mean of the detection mask near the source position (T11_023 review correction — not "the masked fraction"). On M2 b4 the median is 0.981, IQR 0.961–0.986; this project's clean-sample convention is ≥ 0.97 in T109 and > 0.8 in the earlier T56 (Source: cognition-os/03_TRUST_TABLE.md). High MASKFRAC = fully covered; partly masked sources have low MASKFRAC. The two groups differ by only ~0.02, too small to produce a 0.05 difference in m directly — it is probably a proxy for epoch or detector position.

DIST_NNDIST_NN

最近邻距离；C-iso 要求 ≥ 60″。Nearest-neighbour distance; C-iso requires ≥ 60″.

EP_EXTENT / SUM_FLAGEP_EXTENT / SUM_FLAG

延展度 / 汇总标志；点源样本要求 EP_EXTENT = 0、SUM_FLAG = 0。Extent / summary flag; point-source sample requires EP_EXTENT = 0, SUM_FLAG = 0.

ecut / scutecut / scut

emldetect 的拟合/掩膜半径参数。standalone 默认 ecut = 0.68；DR14 的 CUTRAD = 60″ 相当于 4″ 像素下的 ecut = 15。emldetect's fit/mask radius parameters. Standalone default ecut = 0.68; DR14's CUTRAD = 60″ corresponds to ecut = 15 at 4″/pixel.

ECFECF

能量转换因子（energy conversion factor），counts → flux。Energy conversion factor (counts → flux).

logN–logSlogN–logS

源计数分布：单位面积上亮于流量 S 的源数；斜率陡 ⇒ Eddington 偏差大。Source-count distribution: number of sources per unit area brighter than flux S; steeper slope ⇒ larger Eddington bias.

MC / AsimovMC / Asimov

蒙特卡洛 / 期望数据集（Yᵢ = μᵢ 的假想数据，Cowan et al. 2011）。Monte Carlo / Asimov dataset (the hypothetical dataset with Yᵢ = μᵢ, Cowan et al. 2011).

LOPO / LOTOLOPO / LOTO

留一位置 / 留一目标验证（leave-one-position/target-out）：防止循环论证的标准做法。Leave-one-position/target-out: the standard practice to avoid circular reasoning.

bootstrapbootstrap

有放回重采样估计不确定度；v2 用两级（先簇后行）。Resampling with replacement to estimate uncertainty; v2 uses two levels (first cluster then row).

关键文献（手稿 i044_references.bib 中引用）Key Papers (cited in the manuscript's i044_references.bib)

来源：paper/i044_references.bib。完整书目见该文件。Source: paper/i044_references.bib. Full bibliography in that file.

项目内关键文件（先读顺序）Key Project Files (Suggested Reading Order)

1. README.md → GOAL.md（项目总目标，中文）README.md → GOAL.md (project overall goal, Chinese)
2. CLAUDE.md（铁律、DET_ML 约定、目录陷阱——每次动手前重读）CLAUDE.md (iron rule, DET_ML convention, catalogue traps — re-read before every hands-on session)
3. cognition-os/PARETO_BENCHMARK_V2.md（当前规则）cognition-os/PARETO_BENCHMARK_V2.md (current rules)
4. cognition-os/BENCHMARK_V2_TASKS.md（任务卡与最新状态）cognition-os/BENCHMARK_V2_TASKS.md (task cards and latest status)
5. cognition-os/04_DECISION_LOG.md（每个决定的背景与理由）cognition-os/04_DECISION_LOG.md (background and rationale for every decision)
6. cognition-os/PITFALLS.md（坑清单）cognition-os/PITFALLS.md (pitfalls list)
7. wiki/index.md（33 页概念 wiki，从 index 进）wiki/index.md (33 concept wiki pages, start from index)
8. paper/i044_manuscript_v32.tex（当前手稿）paper/i044_manuscript_v32.tex (current manuscript)
9. benchmark/leaderboard_v2.json（当前成绩）benchmark/leaderboard_v2.json (current scores)