P001:预注册设计（MOS1 相机迁移 + CDFS 空白场对照）P001: the pre-registered design (MOS1 camera transfer + CDFS blank-field control)
P001 回答一个问题：继承下来的 R 族 GTI 筛选，在 (a) 另一台 MOS 相机上、（b） 一个没有亮可变源的天空上，是否仍然有效且有用?2026-10-13 PI 批准（"同意"）；代码在看到任何盲测结果之前冻结（git 46d8df3）。
P001 answers one question: does the inherited R-family GTI screen stay valid and useful (a) on the other MOS camera and (b) on a sky without bright variable sources? Approved by the PI on 2026-10-13; the code was frozen (git 46d8df3) before any blind outcome was computed.
1. 数据与盲性（P001 §2）1. Data and blindness (P001 §2)
| 集合Set | 内容Content | 盲性Blindness |
|---|---|---|
| M31-MOS1 | 30 个 ObsID(0800730101–0800733101),MOS1 S001（U002 曝光是 CalClosed,排除）。30 ObsIDs (0800730101-0800733101), MOS1 S001 (U002 exposures are CalClosed and excluded). | 0800730101 不盲(2.1 时代做过 MOS1);其余对 R 盲，但与 MOS2 不独立:相同指向、相同耀斑 —— 这是相机迁移测试，不是新天空测试。0800730101 is not blind (2.1-era MOS1 work); the rest are blind for R but not independent of MOS2: same pointings, same flares — a camera-transfer test, not a fresh-sky test. |
| CDFS | MosaicSelfCal CDFS 列表的 34 个 ObsID,MOS1+MOS2,stage-1/2 产物。34 ObsIDs of the MosaicSelfCal CDFS list, MOS1+MOS2, stage-1/2 products. | 不盲:0108060601、0108060701;部分看过:0555780501、0604960501、0604961301（只进敏感性分析）。主集 = 其余 29 个。Not blind: 0108060601, 0108060701; partly seen: 0555780501, 0604960501, 0604961301 (sensitivity analysis only). Primary set = the other 29. |
2. 冻结的方法（P001 §3,与 g1 完全一致，除列出项外）2. The frozen method (P001 §3; identical to g1 except as listed)
- 100 s cell;选择能段 FOV 2.5–8.0 keV;测量窗口 1.000–2.395 keV、93 × 15 eV 整数构造 bin;固定 Al-K 1.4879 / Si-K 1.7466 keV;单个对数抛物线；Cash;每候选 200 次零假设抽样；门边界 0.02;候选族 = archived_2filter + 率排名（保留 0.3–1.0）。100 s cells; selection band FOV 2.5-8.0 keV; measurement window 1.000-2.395 keV, 93 × 15 eV integer-constructed bins; fixed Al-K 1.4879 / Si-K 1.7466 keV; a single log-parabola; Cash; 200 null draws per candidate; gate boundary 0.02; candidate family = archived_2filter + rate-rank at retention 0.3-1.0.
- 相机专属校准重算而非调参:κ（门误差标度）来自同一窗口恒定性程序，按相机与场重算。耦合修正不应用;门内与反事实线比都报告。Camera-specific calibrations are recomputed, not tuned: κ (gate-error scale) comes from the same within-observation window-constancy procedure, per camera and field. The coupling correction stays unapplied; both as-gated and counterfactual line ratios are reported.
- 规则化零控制（取代 g1 的手工清单）：若观测的归档 2.filter 拒绝 < 2% 活时间且其 FOV 2.5–8 keV 100 s 光变与常数相容（χ² p > 0.05），它就是零控制。规则在读任何 R 之前在代码里运行。Zero controls by rule (replacing g1's hand list): an observation is a zero control if its archived 2.filter rejects < 2% of live time and its FOV 2.5-8 keV 100 s light curve is consistent with constant (χ² p > 0.05). The rule runs in code before any R is read.
- 新增：匹配保留率候选:率排名截在该观测归档 2.filter 的保留活时间比例 f_T(2f) 上 —— 这才是一次公平的"同保留率"比较。New: the matched-retention candidate: rate-rank cut at exactly that observation's archived 2.filter retained live fraction f_T(2f) — the fair "equal retention" comparison.
3. 终点 E1–E6 与阈值（P001 §4）3. Endpoints E1-E6 and thresholds (P001 §4)
| 终点Endpoint | 定义与阈值Definition and threshold |
|---|---|
| E1 有效性Validity | 零控制族假阳性分数（|z_null| ≥ 3,全体候选）≤ 0.05;零假设居中 1.00 ± 0.02（f_T ≥ 0.5 的候选）。A3 后：E1 = 零控制 FWER 单项，居中只报告。Zero-control family-wise false-positive fraction (|z_null| ≥ 3 over all candidates) ≤ 0.05; null-centring 1.00 ± 0.02 for candidates with f_T ≥ 0.5. After A3: E1 is the zero-control FWER alone; centring is reported only. |
| E2 线门Line gate | E3 集合上，匹配候选的 PROVISIONAL_FAIL 比例 ≤ 归档 2.filter 的 FAIL 比例 + 0.05（A1 修正后的定义；原文在前沿上恒为 0,是空话）。On the E3 set, the matched candidate's PROVISIONAL_FAIL fraction ≤ archived 2.filter's FAIL fraction + 0.05 (the A1-corrected definition; the original wording was vacuous — always 0 on the frontier). |
| E3 有用性（主终点）Usefulness (primary) | 每观测 Δ = ln R（匹配率排名） − ln R(2f),配对循环移位零假设给 z_Δ。成功 = 单侧符号检验 p < 0.01 且 Stouffer Z > 3 且 无 z_Δ < −3 且门 FAIL 的观测。Per observation Δ = ln R(matched rate-rank) − ln R(2f), with z_Δ from a paired circular-shift null. Success = one-sided sign test p < 0.01 and Stouffer Z > 3 and no observation with z_Δ < −3 and a gate FAIL. |
| E4 相机一致性Camera consistency | MOS1 与 MOS2 每观测 Δ 的 Spearman 相关（仅描述，无通过/失败）。Spearman correlation of per-observation Δ between MOS1 and MOS2 (descriptive only, no pass/fail). |
| E5 耦合/亏损与角落Coupling/deficit and corner | 报告加权 ln(L_K/L_B)（门内与反事实）；角落测试：主集上叠加 ln(Q_K/Q_B) 的 σ，σ ≤ 0.010 才宣布角落可用，否则 T017 以"太吵"关闭。Report weighted ln(L_K/L_B), both as gated and counterfactual; corner test: stacked σ of ln(Q_K/Q_B) over the primary set — the corner is declared usable iff σ ≤ 0.010, otherwise T017 closes as "too noisy". |
| E6 空白场期望（仅描述）Blank-field expectation (descriptive) | 若移除成分是软质子，耀斑 CDFS 观测的 R 增益应像 M31;大差异指向 M31 源。报告，不打分。If the removed component is soft protons, R gains on flared CDFS observations should look like M31's; large differences point at M31 sources. Reported, not scored. |
结果的含义（§5）：E1 失败 → 该相机的 E3 解释停止；E1 过、E3 失败 → 负结果，如实记录；全过 → 可采纳的迁移证据，仍需 D002 第 3 步才采纳。
What a result means (§5): E1 FAIL → stop interpreting E3 for that camera; E1 PASS with E3 FAIL → a negative result, recorded as such; all pass → admissible transfer evidence, still gated on D002 step 3 before adoption.
4. 结果（P001 §9：两半都已完成，2026-10-13）4. Results (P001 §9: both halves complete, 2026-10-13)
| 终点Endpoint | M31 MOS1 （盲测，E002）(blind, E002) | CDFS MOS1 （盲测，E003）(blind, E003) | CDFS MOS2 （盲测，E003）(blind, E003) |
|---|---|---|---|
| E1 | PASSPASS 0/780/78 | NOT_EVALUABLENOT_EVALUABLE （敏感性集 PASS,0/9）(sensitivity set PASS, 0/9) | NOT_EVALUABLENOT_EVALUABLE （敏感性集 PASS,0/9）(sensitivity set PASS, 0/9) |
| E2 | PASSPASS 0 对 00 vs 0 | PASSPASS 0 对 00 vs 0 | PASSPASS 0 对 00 vs 0 |
| E3 （主终点）(primary) | FAILFAIL 13/25,Z = 0.0213/25, Z = 0.02 | FAILFAIL 24/29,Z = 1.5524/29, Z = 1.55 | FAILFAIL 27/29,Z = 1.7327/29, Z = 1.73 |
| E4 / E6 （描述）(descriptive) | ρ = 0.08;中位 Δ +0.0005ρ = 0.08; median Δ +0.0005 | 中位 Δ +0.0035（M1）与 +0.0029（M2）median Δ +0.0035 (M1) and +0.0029 (M2) | |
| E5 角落corner | TOO_NOISYTOO_NOISY σ = 0.013σ = 0.013 | USABLEUSABLE σ = 0.0044σ = 0.0044 | USABLEUSABLE σ = 0.0038σ = 0.0038 |
§5 结论（D006）:在任何主集上 E3 都没有通过。按预注册判据，等保留率下的率排名在 M31 MOS1 上、也不在 CDFS 空白场上优于标准 2.filter —— 这是一个负结果。CDFS 上增益方向一致但只有约 0.3%（R），描述性而已。角落按数据集回答：CDFS 两台相机可用，M31 MOS1 太吵。
§5 conclusion (D006): E3 passes on no primary set. By the pre-registered criterion, rate-rank at matched retention does not beat the standard 2.filter on M31 MOS1 or on the CDFS blank field — a negative result. On CDFS the gain is consistent in sign but only about 0.3% in R, descriptive only. The corner is answered per set: usable on both CDFS cameras, too noisy on M31 MOS1.
由此（D006 的推论）:R 族的表面优势来自"选保留率"。要让任何不同的 GTI 政策变得值得，只剩 D002 第 3 步 —— 那个把曝光损失与连续谱增益定价的下游判据。没有 PI 决定，不改方法、不跑新东西。
Consequence (D006): the R-family's apparent advantage was a retention choice. The only way a different GTI policy could now be justified is D002 step 3 — the downstream criterion that prices exposure loss against continuum gain. No method change and no new run without a PI decision.
prereg/P001_mos1_and_blank_field.md §9; cognition-os/04_DECISION_LOG.md D005, D006; cognition-os/experiments/E002_p001_m31_mos1.md; E003_p001_cdfs_blank_field.md.5. 代码冻结与回归门（§6）5. Code freeze and the regression gate (§6)
- 冻结 commit:
46d8df35c41ce11712972192bec91bef0d917b00（分支feat/pareto-gti-cal,2026-10-13）。Frozen commit:46d8df35c41ce11712972192bec91bef0d917b00(branchfeat/pareto-gti-cal, 2026-10-13). - sha256:
m31_mos2_line_continuum_invariant.py=ff41b1aa…;p001_endpoints.py=679a9af2…。sha256:m31_mos2_line_continuum_invariant.py=ff41b1aa…;p001_endpoints.py=679a9af2…. - 回归门：PASS —— 冻结代码与未修改 g1 脚本在全部 30 个 M31 MOS2 ObsID 上完全一致（26,111 个数字，最大相对差 0）。与存档 g1 收据的差异只有 1 个元数据键与 1 个描述性显著度（4.6e-5），且 g1 收据在那一点与它自己存的 slope/stderr 不一致 —— 差异来自 g1 的存档脚本，不是 P001 代码。Regression gate: PASS — the frozen code equals the unmodified g1 script on all 30 M31 MOS2 ObsIDs (26,111 numbers, max relative difference 0). vs the stored g1 receipt the only differences are 1 metadata key and 1 descriptive significance (4.6e-5), and the g1 receipt disagrees with its own stored slope/stderr there — the difference comes from g1's archived script, not from P001 code.
- P001 相对 g1 的改动仅限：相机/曝光参数化、输入根布局、规则化零控制、匹配保留率候选、配对差零假设、角落 σ。Changes from g1 are limited to: camera/exposure parameterisation, input-root layouts, rule-based zero controls, the matched-retention candidate, the paired-difference null, and the corner σ.
prereg/P001_mos1_and_blank_field.md §6; results/p001_regression/regression_check_a2_vs_orig.json.6. 修正案 A1–A3（全部在看到任何盲测结果之前）6. Amendments A1-A3 (all before any blind outcome)
| 修正案Amendment | 内容Content |
|---|---|
| A1 | E2 原文是空话（前沿上恒为 PASS）。改为：E3 集合上匹配候选的 FAIL 比例 ≤ 2.filter + 0.05。另定：E1 空集 → NOT_EVALUABLE;曝光选择、corner 几何（MOS1 不查 2 圆 4 框，因 2.filter 写过畸形数 15&00）、corner Q 定义、配对零假设与种子、匹配候选的 ID 与旗标。阈值一个没改。E2 as written was vacuous (frontier rows are PASS by construction). Now: the matched candidate's FAIL fraction on the E3 set ≤ 2.filter's + 0.05. Also fixes: empty E1 set → NOT_EVALUABLE; exposure choice; corner geometry (MOS1 is not checked for 2 circles + 4 boxes, because 2.filter once wrote the malformed number 15&00); corner Q definition; paired null and seeds; matched-candidate id and flag. No threshold changed. |
| A2 | 相同掩码 = 平局：匹配掩码与归档 2.filter 逐 cell 相同时 Δ = 0、零假设宽度为 0,该观测从 E3 中剔除并计数（符号检验剔平局）。发现于 M31 MOS2 0800733101（已消费的参考数据）。另：共有的 CDFS ObsID 0604960101 按默认根取 codexnim 副本，Hermes 副本记为重复（两次还原因 emchain 帧内随机化而不同）。Identical masks are ties: if the matched mask equals the archived 2.filter cell for cell, Δ = 0 exactly with null width 0, and the observation is dropped from E3 and counted (the sign test drops ties). Found on M31 MOS2 0800733101 (consumed reference data). Also: the shared CDFS ObsID 0604960101 is taken from the codexnim copy under the default roots, the Hermes copy recorded as duplicate (the two reductions differ because emchain randomises within frames). |
| A3 | PI 决定（"选择A"）：E1 = 零控制 FWER 单项。居中仍计算并报告，但不再是 E1 的通过条件，因为 z_null 与 z_Δ 都减去了零假设均值。起因：在已消费的 M31 MOS2 参考数据上，4 个 f_T ≥ 0.5 的候选零假设均值 1.020–1.037,而 FWER 是 0/58;按批准原文 E1 会"按构造"失败并作废 E3。PI decision ("choose A"): E1 is the zero-control FWER alone. Centring is still computed and reported but is no longer an E1 pass condition, because z_null and z_Δ subtract the null mean. Trigger: on consumed M31 MOS2 reference data, 4 candidates with f_T ≥ 0.5 had null means 1.020-1.037 while the FWER was 0/58; under the approved wording E1 would have failed by construction and voided E3. |
7. 为什么只认预注册7. Why only pre-registration counts
这个项目的历史里有一个直接教训：r16 规则是 0800730101 单观测前沿上的真实成员，但在 30 个观测上迁移时只有 10/8 正（p = 0.50）—— 单观测前沿挖掘选出的是唯一带噪声坐标（对比度）上的向上涨落。2026-09-07 的决定把"单观测前沿点"降级为假设：只有一条在未参与其选择的观测上原样应用的预注册规则，才算证据。P001 把规则、统计量、阈值、集合、修正案全部写在看到结果之前，并把代码连同 git 哈希一起冻结，正是为了让"有效但不优于"这样的负结果也无可辩驳。
The project's own history contains the direct lesson: the r16 rule was a genuine member of the 0800730101 single-observation frontier, yet on transfer to 30 observations it was only 10/8 positive (p = 0.50) — mining the upper-right envelope of one observation selects upward noise excursions in the only noisy coordinate (the contrast). The 2026-09-07 decision demoted single-observation frontier points to hypotheses: only a pre-registered rule applied unchanged to observations that took no part in its selection is evidence. P001 writes the rule, the statistic, the thresholds, the sets and the amendments down before any outcome, and freezes the code with its git hash, precisely so that a negative result like "valid but not better" is unimpeachable.
legacy/cognition-os/04_DECISION_LOG.md (2026-09-07); cognition-os/00_AI_BRIEFING.md invariant 7.动画：单次观测的 Pareto 前沿（教学收据数据）Animation: the Pareto front within one observation (tutorial receipt data)
0800731101 MOS2 的 33 个候选（教学收据）。点一个点看它的数字。前沿 = 不被任何其他候选同时在两个坐标上超过的点。注意 corner 排名（故意错误）与随机 cell 都被前沿支配。
The 33 candidates of 0800731101 MOS2 (tutorial receipt). Click a point to read its numbers. The front = points not beaten in both coordinates at once. Note the corner ranking (deliberately wrong) and the random cells are both dominated.