考试：固定天域，多个观测

The exam: one fixed cell, many observations

整个项目的方法论核心是一个"考试"（exam）：一块固定天域（4′ 孔径，r240），被多个观测以不同的离轴角、姿态角、年代、SP/QPB 状态看到。如果流水线是好的，这些观测必须给出同一个答案。它们不一致的程度就是系统误差的测量值。设计冻结于 D019/D020：7 个 problem、18 个 ObsID、45 个 case（40 PASS、5 EXCLUDED）。

The methodological core of the project is an "exam": one fixed sky cell (4′ aperture, r240), seen by many observations at different off-axis angles, roll angles, epochs and SP/QPB states. If the pipeline is good, these observations must give the same answer. The size of their disagreement is the measured systematic. The design is frozen by D019/D020: 7 problems, 18 ObsIDs, 45 cases (40 PASS, 5 EXCLUDED).

三个字段做考场：WCDFS（最暗，纯天空）、M31（有源星系盘/晕）、LMC（有最强 SP 与一个反常观测）。分数定义在 MosaicSelfCal/BENCHMARK.md：每个 problem 算加权平均附近的 \(\chi^2_\nu\)、DerSimonian–Laird 超额散布 \(\tau/\bar{I}\)、以及面亮度对 \(\theta\) 的加权斜率。更好的流水线 = 更小的 \(\tau/\bar{I}\) 与 \(|\text{slope}|\)，且 \(\sigma/\bar{I}\) 不变大。

Three fields serve as exam rooms: WCDFS (faintest, sky only), M31 (a disc/halo with sources), LMC (the strongest SP and one anomalous observation). The score definitions are in MosaicSelfCal/BENCHMARK.md: per problem, \(\chi^2_\nu\) about the weighted mean, the DerSimonian–Laird excess scatter \(\tau/\bar{I}\), and the weighted slope of surface brightness against \(\theta\). A better pipeline = smaller \(\tau/\bar{I}\) and \(|\text{slope}|\) at no larger \(\sigma/\bar{I}\).

来源：Source: MosaicSelfCal/BENCHMARK.md, MosaicSelfCal/results/exam_v4_r240/SUMMARY.md, cognition-os/04_DECISION_LOG.md D019/D020

两个被检验的工具

The two tools under test

项目不修改 SAS 本身；它检验并迭代两个自有工具（D015）：发现问题 → 工具 issue → 测试过的版本化发布 → DXC 用固定版本重测。版本固定在 cognition-os/tool_iteration.md。

The project does not modify SAS itself; it tests and iterates two in-house tools (D015): a finding becomes a tool issue, the fix lands as a tested, versioned release, and DXC re-tests pinned to that version. Versions are pinned in cognition-os/tool_iteration.md.

来源：Source: cognition-os/tool_iteration.md （版本表与 item 19/37/45）(version table and items 19/37/45), cognition-os/04_DECISION_LOG.md D015/D023/D024/D025

三台机器与分工

Three machines and the division of labour

与如何用手继续页相同：lang 分析与拟合、Virgo 减数据（SAS/ESAS）、lang2 只做谱拟合。细节与命令见那一页。

As on the how to continue by hand page: lang for analysis and fits, Virgo for SAS/ESAS reduction, lang2 for spectral fits only. Details and commands are there.

来源：Source: cognition-os/08_RUNNING.md §2

信任分级与红队

Trust tiers and the red teams

每个结论被分级为 Trusted / Provisional / Questioned / Deprecated（cognition-os/00_AI_BRIEFING.md）。红队（red team）是独立的再分析：每条重要结论都会被另一个执行者用可运行的检验复核，发现的问题记入 cognition-os/REDTEAM_QUEUE.md（RT-M01 到 RT-M78）。红队的发现以记录为准：例如 R18a 的 +12.9 % 因 MOS2 旋转错误被撤回，R20i 的 7\(\sigma\) 被修正为 3.1\(\sigma\)。本手册所有数字都在现状页的信任表里带等级。

Every conclusion is tiered Trusted / Provisional / Questioned / Deprecated (cognition-os/00_AI_BRIEFING.md). A red team is an independent re-analysis: each material claim is re-checked by another executor with runnable checks, and findings go to cognition-os/REDTEAM_QUEUE.md (RT-M01 through RT-M78). The red team's findings win: e.g. R18a's +12.9 % was withdrawn for a MOS2 rotation error, and R20i's 7\(\sigma\) was corrected to 3.1\(\sigma\). Every number in this handbook carries its tier in the trust table on the state page.

来源：Source: cognition-os/REDTEAM_QUEUE.md, cognition-os/00_AI_BRIEFING.md

版本固定

Version pinning

任何分数都只在"哪个工具版本 + 哪个校准文件"下才有意义。固定方式：esas 用 releases/vX.Y.Z/ + SHA256；xsherpa 用 releases/current 指针；CCF 用日期目录 + sas_ccf_latest 指针（D027：永远是最新公共 CCF，两个 SAS 版本共用）。每个任务目录记录它用的版本；考试集（D020 冻结）是所有比较的基准面。

A score only means something under "which tool version + which calibration files". Pinning: esas by releases/vX.Y.Z/ + SHA256; xsherpa by the releases/current pointer; the CCF by dated directories plus the sas_ccf_latest pointer (D027: always the latest public CCF, shared by both SAS versions). Every job directory records the versions it used; the exam set (frozen, D020) is the baseline of every comparison.

来源：Source: cognition-os/tool_iteration.md, cognition-os/04_DECISION_LOG.md D027, cognition-os/shared/version_registry.md