成果 / Results

Results

• R20v（临时，已复算）：自由每相机 SP 谱形下，M31 配对斜率 \(+1.51\to-0.73\pm0.29\ \%/\text{arcmin}\)、平衡边缘步长 \(+9.86\to+0.88\ \%\)。边缘增亮不是稳健检测，取决于 SP 分解。临时R20v (provisional, recomputed): with per-camera free SP shapes the M31 pair slope goes \(+1.51\to-0.73\pm0.29\ \%/\text{arcmin}\) and the balanced edge step \(+9.86\to+0.88\ \%\). Edge brightening is not a robust detection; it depends on the SP decomposition. provisional
• R20r/R20s/R20t（已验证）：主失配源是冻结的 SP 谱形（自由形状 \(\Delta\chi^2\ 3026\)）；自由 SP 谱形后 QPB 标度只再买 252（原 1001）——QPB 标度不是主要缺陷（RT-M83：共享、顺序依赖，对称归因下 QPB 保留 62.6 %）。已确立R20r/R20s/R20t (verified): the main misfit is the frozen SP spectral shape (\(\Delta\chi^2\ 3026\) for free shapes); once SP shapes are free the QPB scale buys only 252 more (was 1001) — the QPB scale is not the main defect (RT-M83: shared and order-dependent; under symmetric attribution QPB keeps 62.6 %). established
• R20u（已验证）：角区数据既不能证明也不能否证每相机 ESAS QPB 水平；LMC 0137551401 的 7–12 keV FoV/QPB 1.48/1.52 是硬 FoV 超额（SP 类），不是 QPB 误差；\(s=1\) 保持名义值。已确立R20u (verified): the corner data neither certify nor refute the per-camera ESAS QPB level; LMC 0137551401's 7–12 keV FoV/QPB of 1.48/1.52 is a hard FoV excess (SP-like), not a QPB error; \(s=1\) stays nominal. established
• R21 step 0 / R21-1（已验证）：MOS2 全链计时 1641 s（evselect 34 %、backscale 32 %、arfgen 17 %）；pn 3580 s（backscale 63 %、evselect 32 %）；快速抽谱原型 0.09–0.12 s/区域，像素直方图与 DET 图与 SAS 逐位相同（BACKSCAL \(-1.1…+3.0\ \%\)、QPB \(+25…+35\ \%\)、ARF \(-4.1/-2.7\ \%\) → R21-2）。已确立R21 step 0 / R21-1 (verified): the full MOS2 chain takes 1641 s (evselect 34 %, backscale 32 %, arfgen 17 %); pn 3580 s (backscale 63 %, evselect 32 %); the fast-extraction prototype reaches 0.09–0.12 s per region with pixel histograms and DET maps bit-identical to SAS (BACKSCAL \(-1.1…+3.0\ \%\), QPB \(+25…+35\ \%\), ARF \(-4.1/-2.7\ \%\) → R21-2). established
• esas v1.8.0-dev.6（已验证并提交）：112 测试 + 6 子测试通过，commit 8832d72，未 READY。已确立esas v1.8.0-dev.6 (verified and committed): 112 tests + 6 subtests pass, commit 8832d72, not READY. established

来源：Source: cognition-os/08_RUNNING.md revision 9–14，revisions 9–14, MosaicSelfCal/RESEARCH_PLAN.md R20r–R20v、R21/R21-1 行the R20r–R20v, R21/R21-1 rows

工作 / Work

Work

• 跑了/答完：R20n（luna）、R20o（sol）、R20p（sol）、R20q（sol）、R20r（sol）、R20s（sol）、R20t（sol）、R20u（sol）、R20v（sol）、R21 step 0（codexnim）、R21-1（sol）；红队 RT-M77–RT-M84 全部归档；文献检索 S01–S09 完成。Ran and answered: R20n (luna), R20o (sol), R20p (sol), R20q (sol), R20r (sol), R20s (sol), R20t (sol), R20u (sol), R20v (sol), R21 step 0 (codexnim), R21-1 (sol); red teams RT-M77–RT-M84 all archived; literature searches S01–S09 done.
• 派出：R20w（sol）、R21-2（sol）、RT-M85（codexnim）、S10（luna）、手册刷新 1（Hermes）。Dispatched: R20w (sol), R21-2 (sol), RT-M85 (codexnim), S10 (luna), handbook refresh 1 (Hermes).
• 部署：手册发布于 https://dxc-handbook.pages.dev/（15:10 EDT，密码门）；RT-M82b 复核 11 条发现全部受理。Deployed: the handbook at https://dxc-handbook.pages.dev/ (15:10 EDT, password gate); RT-M82b's 11 findings all accepted.
• 工具：esas v1.8.0-dev.6 由 sol 生成、Claude 复核并提交（8832d72）；lesson promotion 完成。Tools: esas v1.8.0-dev.6 built by sol, verified and committed by Claude (8832d72); lesson promotion done.

来源：Source: cognition-os/06_AGENT_LOG.md 2026-10-19 各行the 2026-10-19 rows

问题 / Problems

Problems

• RT-M82 停滞：Hermes 会话以一行裸工具调用结束、无报告 → 以新会话 RT-M82b 重派，11 条发现全部受理（刷新 1 逐条修复）。RT-M82 stalled: the Hermes session ended on a raw tool-call line with no report → relaunched fresh as RT-M82b; all 11 findings accepted (fixed one by one in refresh 1).
• R20t 的归因顺序依赖（RT-M83）：1001→252 取决于归因顺序；对称归因下 QPB 保留 62.6 %；"大部分被 SP 拿走"改述为"共享、顺序依赖"。R20t's attribution is order-dependent (RT-M83): 1001→252 depends on the attribution order; under symmetric attribution QPB keeps 62.6 %; "mostly taken by SP" restated as "shared, order-dependent".
• R20o 的 78 % 是一个观测（RT-M79）：LMC 0137551401 占 A→C \(\Delta\chi^2\ 973.3\) 中的 768.6；R20r 的 33 % 保留同理（93 % 在同一观测）。已按单观测驱动重述。R20o's 78 % is one observation (RT-M79): LMC 0137551401 carries 768.6 of the 973.3 A→C \(\Delta\chi^2\); R20r's 33 % retention likewise (93 % in the same observation). Restated as single-observation driven.
• 手册自身的缺陷（RT-M82b，11 条）：用户名泄露（已删）、工具版本过期、运行状态过期、pn QPB 成因写成被否定的解释、CDFS 中位数 n 与值不一致、论文/讨论班空壳（本次补齐）。The handbook's own defects (RT-M82b, 11 findings): a username leak (removed), stale tool versions, stale running statuses, the pn QPB cause written as the refuted explanation, the CDFS median's n and value inconsistent, the paper and seminar pages empty stubs (completed in this refresh).

来源：Source: cognition-os/REDTEAM_QUEUE.md RT-M79/RT-M82b/RT-M83 行the RT-M79/RT-M82b/RT-M83 rows

待 PI 决定 / For the PI

For the PI

• 论文草稿的中心贡献句在 R20v 之后是否仍用"主导"一词（本稿已改弱）。Whether the paper draft's central-contribution sentence keeps the word "dominate" after R20v (this draft weakened it).
• 若 R20w 确认边缘增亮不稳健，第 3 步成像的验收标准是否改写。If R20w confirms the edge brightening is not robust, whether the step-3 acceptance criterion is rewritten.
• 手册部署（本刷新不部署；Claude 复审后部署）。Handbook deployment (this refresh does not deploy; Claude deploys after review).