BIL16 第33章来源映射与本批读取边界
2026-10-16。本批以接受I3、已应用C15集中校订、当前工作Markdown与主控完整准备笔记为语义基线。下方分隔线后保留A6/C15历史来源记录，“已读/实测”只指当时执行者，不表示本批重新读完或运行。
C15-E02已删除过程句，正文8.1承接接受语义不回退。新增normal/t尾积分中间步骤与Kupiec计数似然推导，P02由明确权重/beta重算，P05改为8损失/3窗口的确定性局部例。三份旧/tmp PDF/文本当前缺失，根有界归档查找无恢复指针；下方页码/SHA保留历史身份，不宣称本批重新取得或逐页核读。
本批主作者完整读取两章工作正文/source_notes、两语题答设计用中文Loom、正式派发/范围/140项绑定和根ROOT_READER_NOTES/brief；完整读取risk.py、定向读取inference.py相关接口与performance.py财富域、33CLI暴露输入和I3/工作analyze差异、I3两Notebook各16cells全部source及runner输出时钟。PBO本地2015-02-27配套txt的Definition2.2与Algorithm2.3作定向核对；不声称全文/后续期刊版复核，无网络抓取。
复杂推导由Sol作者整合。根8组独立算术及四点边际联合反例仅按适用参数沿用，不重复计成本批新研究。5个新显示小代码块、两个标量尾积分及似然算术为本批局部核验；工程辅助只负责两护栏/直接测试和首图显示。候选隔离模块加载及实际结果见build/batches/BIL16/r1_20261016/LOCAL_CHECKS.json与engineering/ENGINEERING_CHECKS.json，作者结论不替代主控独立接受。
I3默认32Notebook fresh一seed1250×200冠军及252CSCV；33Notebook调用include_french=False，fresh3000/MC50000并写JSON/4图。这里是静态读入口结果，不是本批执行；工作33无参analyze仍总调用French。完整CLI/Notebook/runner/French/冠军/AR/20,000×12厚尾模拟均未重跑。原始/抽取/处理French数据不随默认I3包，公开可得不代替权利或执行授权；A10-B未授权。英语Notebook、便携装配/冷启动/全局引用归BIL23，BIL17消费者待后续。P4真实研究planned，PIT/权利/借券/日历/容量闸门不变。
以下为历史记录（读取身份保留）
Chapter 33 source and evidence notes
Scope: this file records sources actually opened for the chapter-author pass. It is a partial reading log, not a claim of comprehensive literature review. The parent Codex agent will independently check and update this record before final acceptance.
Opened source documents
| Key | Source and exact read scope | How it is used | Evidence limits |
|---|---|---|---|
AT2002-v5 |
Carlo Acerbi and Dirk Tasche, “On the Coherence of Expected Shortfall,” arXiv:cond-mat/0104295v5 (2002), local copy /tmp/a6_acerbi_tasche_v5.pdf; SHA-256 ee6e09d81619caac419bbacc568db17c8fcecc0b3cc4057b5baee653c9e1ecd1; URL: https://arxiv.org/pdf/cond-mat/0104295v5. Opened printed pp. 5–9: Definition 2.6; Proposition 3.2 and Eq. 3.3; Proposition 4.1 and adjacent discussion. |
The threshold-atom correction, quantile-integral representation and stated integrability condition; the empirical worst-tail interpretation is described with the paper's sampling qualification. | Only these pages and adjacent paragraphs were read. Do not describe this as a full-paper or journal-version review. The chapter formulas are restated in loss-sign convention. |
KUPIEC1995-FEDS95-24 |
Paul H. Kupiec, “Techniques for Verifying the Accuracy of Risk Measurement Models,” Finance and Economics Discussion Series 95-24, Federal Reserve Board. Local copy /tmp/a6_kupiec_1995.pdf; SHA-256 99d8cb03e47bb9c18bb00194ea4a6e880d930b06fd34adb3aab7bfc0428d3cbc; official PDF: https://fraser.stlouisfed.org/files/docs/publications/bog_feds/feds_1995-024.pdf. Opened printed pp. 6–8, 18–19, and 23–24. |
Independent Bernoulli failure framing and the one-day monitoring-horizon discussion; count likelihood and its likelihood-ratio form (printed Eq. 5–6); binomial coefficient cancellation (Eq. 7–8); historical-simulation critical-value estimation error. | The cover states May 1995 and the interior title page April 1995; the document identifies itself as preliminary FEDS discussion material, not a journal article. Only the listed pages were read. The chapter uses a course derivation and does not repeat the source's older “uniformly most powerful” wording for the composite two-sided likelihood-ratio comparison. The chi-square reference is asymptotic; non-rejection is not proof of model accuracy, especially at zero or few exceptions. |
SCIPY-T |
SciPy v1.18.0 Manual, scipy.stats.t; opened local extracted text /tmp/a6_scipy_t.txt; SHA-256 1141370485adc8bc14c32d78e12d59836b0f7cc16da3a60a2609fa19808ea979; URL: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.t.html. Read the distribution API and Notes on df, standardized density, loc, and scale; checked ppf, pdf, rvs, and survival-function method descriptions. |
Confirms API semantics and that loc/scale parameterize a shifted/scaled standard t distribution; scale is not its standard deviation. The variance and finite-tail-mean conditions are stated from the distribution's moment integral in the chapter. |
No claim that SciPy documentation derives the chapter's ES formula. SciPy's scale is converted explicitly to a standard-deviation-matched scale when comparing models. |
Project sources opened
| Key | Local source and read scope | Use |
|---|---|---|
A6-33-PLAN |
qa/a6/claude/plan.md, Chapter 33 risk contract and Part V spine. |
Frozen sign convention, VaR/ES definition, scenario revaluation, liquidity, constructed margin, separated backtest windows. Read-only. |
A6-CONTRACT |
Sibling handoff directory ../quant_course_handoffs/2026-10-07_claude_a6/A6_CONTRACT.md, Chapter 33 row and evidence wording; LEARNING_QUALITY_CHECKLIST.md, relevant teaching requirements. |
Chapter and learning acceptance boundaries. |
CHAPTER-TEMPLATE |
templates/chapter_template.md, all required twenty section slots and chapter editing notes. |
Structure and exercise-answer expectations. |
TEACHING-LAYOUT |
qa/TEACHING_LAYOUT_STANDARD.md, layout and teaching checks L01–L06. |
Hand example to general definition, proof/algorithm/boundary separation, and figure placement. |
CH29-LEDGER |
course/zh/chapters/ch29/content.md, ledger contract and relevant example; src/quant_course/backtest.py, BacktestConfig, signal-to-execution lag, cash/position/NAV accounting and modeled limitations; src/quant_course/portfolio_construction.py, exposure_report. |
Revalues the Chapter 29 teaching ledger without changing its engine; uses its delayed target execution for the limit example. These shared files were read-only. |
CH30-32-BRIDGE |
course/zh/chapters/ch30/content.md section 5; course/zh/chapters/ch31/content.md sections 3–6; course/zh/chapters/ch32/content.md sections 5–6; course/zh/chapters/ch15/content.md, course/zh/chapters/ch18/content.md, and course/zh/chapters/ch21/content.md relevant sections. |
Avoids re-deriving earlier ledger, timing, validation, volatility, factor and exposure material; adds risk measures at the Chapter 33 interface. |
FRENCH-FROZEN |
src/quant_course/french_frozen.py, data/manifests/french_industries_2026-10-05.json, local frozen processed CSV and linked FF3 manifest/data. |
Reproducible descriptive illustration; runtime verifies all six raw/extracted/processed hashes. The manifest records the official French Data Library source URL and local download time. |
CH31-32-CLI-PATTERN |
Chapter 31/32 CLI, workflow-test and notebook-builder patterns; qa/a6/claude/notebook_builders/common.py (read-only). |
Follows the existing project pattern for deterministic CLI output, refusal to overwrite, workflow checks and a non-executing notebook builder. |
Numerical evidence and claims
The authored numerical claims, including chart labels and solution checks, are sourced from qa/a6/codex_resume/ch33/results/ch33_results.json after the final default CLI run. The top-level keys are teaching_inputs, pinned_discrete_case, same_portfolio_risk_methods, exposures, ledger_stress_and_limit, risk_model_backtest, french_real_example, and artifacts. Every displayed constructed parameter is under teaching_inputs; every computed output is in its corresponding result key. The four generated figure filenames are in artifacts.figure_files.
The constructed DGP uses the declared seeds in teaching_inputs.portfolio_dgp; it is not market data. The ledger prices, positions, gap returns, ADV, margin rates and limit are explicitly teaching inputs. The French data are published monthly portfolio returns, not point-in-time vintages and not instruments with observed fills. Their use makes no claim about a tradable strategy or realized execution.
Further reading not opened for this pass
- Christoffersen conditional-coverage and exception-independence methods — 未核读. Consult the original work before adding an independence test or making claims about its finite-sample distribution.
- Basel market-risk standards and supervisory backtesting guidance — 未核读. No regulatory threshold or capital rule is asserted in this chapter.
- Additional definitions of tail conditional expectation, worst conditional expectation and expected shortfall — 未核读 beyond the listed Acerbi–Tasche pages. The chapter uses the explicit fractional-tail-mass definition above and does not equate it with a strict or inclusive conditional tail mean when the distribution has atoms.