Clarify decision-making citation usage

docs/research/README.md

@@ -22,6 +22,8 @@ perenne y lo que el código lee/escribe.
   óptimo). Generado por `scripts/build_bound_tightening_audit.py`.
 - `ijds_rebaseline_2026-06-07.md` — registro del run-tag congelado vigente
   (`ijds-rebaseline-2026-06-07`).
+- `literature_reference_audit_2026-06-14.md` — auditoría de evidencia local de
+  lectura para las referencias citadas en el cuerpo IJDS.
 
 ## Registros de gobernanza (decisiones; no se re-ejecutan sin permiso)
 
@@ -36,6 +38,13 @@ perenne y lo que el código lee/escribe.
 - `crpto_pyepo_dfl_intake_2026-05-26.md` — PyEPO 1.3.7 y cierre del comparador
   DFL/SPO+ aislado.
 
+## `literature/` — notas de lectura versionadas
+
+Notas curadas sobre fuentes específicas. Los PDFs de literatura se mantienen en
+`Papers_tesis/`, que está ignorado por Git, para no commitear material con
+copyright. En `docs/research/literature/` deben quedar solo notas, hashes,
+decisiones editoriales y trazabilidad de uso.
+
 ## `foundations/` — referencia técnica perenne
 
 Material de fundamento que puede volver a consultarse para la tesis: conformal

docs/research/literature/fernandez_loria_provost_2022_causal_decision_making_note.md

@@ -0,0 +1,54 @@
+# Fernandez-Loria and Provost (2022) - reading note
+
+Source PDF:
+
+- Local ignored archive: `Papers_tesis/paper/Fernandez-Loria Provost 2022 - Causal Decision Making and Causal Effect Estimation Are Not the Same and Why It Matters.pdf`
+- User download checked: `C:\Users\carlos\Downloads\papers_nuevos\2104.04103v3.pdf`
+- SHA-256: `72551FD3AD3FEA779BC114578680750ABBB2A619C02ECF56CB7705FB2900DF69`
+- Bib key: `fernandezloria2022causaldecision`
+- DOI: `10.1287/ijds.2021.0006`
+- Pages inspected with PyMuPDF: 24
+
+## What the paper contributes
+
+The paper argues that causal decision making (CDM) and causal effect estimation
+(CEE) are not the same task. Its key claim is that accurate effect-size
+estimation is not necessary for accurate decision making when the operational
+goal is treatment assignment. In the authors' framing, the estimand changes:
+the model should be judged by whether it assigns the right action, not only by
+whether it estimates individual effects precisely.
+
+The abstract and introduction emphasize three implications:
+
+1. The modeling objective should, where possible, optimize treatment assignment
+   rather than effect-size accuracy.
+2. Confounding affects decision quality differently from effect estimation
+   quality; confounded data can sometimes be decision-useful.
+3. A non-causal proxy target can sometimes support a useful decision rule,
+   which explains why firms often use predictive models for intervention
+   targeting even when those models are not causal models.
+
+The paper does not mention credit, loans, Lending Club, portfolio selection, or
+conformal prediction. Its value for CRPTO is conceptual rather than empirical.
+
+## How CRPTO should use it
+
+Use this source to support a narrow point in related work: prediction, effect
+estimation, and downstream action are different objects. CRPTO uses the same
+separation in a non-causal credit setting. The PD model is an input artifact;
+the manuscript's claim is about the auditable portfolio decision and the
+funded-set risk certificate.
+
+Do not use this source as evidence that CRPTO estimates causal treatment
+effects, that Lending Club decisions are causal interventions, or that the
+portfolio policy has experimental policy value. Those would require a separate
+design.
+
+Action taken on 2026-06-14:
+
+- Kept `fernandezloria2022causaldecision` in the manuscript.
+- Rewrote the related-work sentence so the citation supports the estimand /
+  assignment-rule distinction rather than a generic "prediction vs action"
+  slogan.
+- Kept the PDF in the ignored local literature archive instead of committing a
+  copyrighted PDF under `docs/`.

docs/research/literature_reference_audit_2026-06-14.md

@@ -0,0 +1,77 @@
+# Literature reference audit - 2026-06-14
+
+Scope: cited references in the IJDS manuscript body, using
+`paper/submission/CRPTO_ijds_submission.aux` as the current body-citation
+surface. This is not a full supplement audit.
+
+The audit question is deliberately strict: does the repository contain evidence
+that a source was actually read or curated, beyond being present in BibTeX?
+
+## Status labels
+
+- `strong`: local PDF/read-note/deep-audit evidence exists, or the source was
+  inspected directly in this session.
+- `partial`: the source is discussed in the book or manuscript, but no standalone
+  read-note or deep-audit row was found.
+- `citation-only`: no local reading evidence found beyond BibTeX and citation
+  use. These are not automatically wrong, but they should be spot-checked before
+  freeze/submission if they carry a specific claim.
+
+## Strong local reading evidence
+
+These body references have local reading evidence from
+`docs/research/papers_tesis_deep_audit_2026-06-06.md`, the new
+Fernandez-Loria/Provost note, or direct PDF inspection in this session:
+
+`albanesi2024credit`, `angelopoulos2023`, `angelopoulos2024foundations`,
+`angelopoulos2024risk`, `bao2025croms`, `bates2021rcps`, `bental2007oce`,
+`bertsimas2004`, `cresswell2024`, `donti2017`, `elmachtoub2022`,
+`fernandezloria2022causaldecision`, `gibbs2024`, `hu2026crc`,
+`jagtiani2019altdata`, `johnstone2021`, `liu2026portfolio`, `mandi2024`,
+`patel2024`, `rockafellar2000cvar`, `schutte2024robust`, `sun2024ptc`,
+`aior2025lendingclub`, `yang2026multidistribution`, `yeh2025training`,
+`zhao2025robust`, `zhou2025credo`, `zhou2026creme`.
+
+Supplement-only check: `fernandezloria2025observational` was also inspected in
+this session. It supports the supplement's causal/experimental-design boundary,
+but it is not credit or loan evidence.
+
+## Partial evidence
+
+These sources are used in plausible places and are discussed in the book or
+manuscript, but I did not find a standalone local read note:
+
+`ayari2026`, `bostrom2021`, `chen2024creditrisk`, `lessmann2015`, `vovk2005`,
+`xia2017`, `yang2025costaware`, `zhou2024`.
+
+Recommended action before freeze/submission: spot-check the exact sentence each
+one supports, especially if it claims recency, a benchmark frontier, or a credit
+domain fact.
+
+## Citation-only / spot-check before freeze
+
+These references appear in the manuscript body, but I did not find local
+evidence that they were read in detail:
+
+`boucheron2013concentration`, `das2023creditgraph`, `delage2010dro`,
+`ghosh2002`, `goldfarb2003robustportfolio`, `hoeffding1963`,
+`boosting2025default`, `serrano2016profitscoring`, `zhao2016p2pportfolio`,
+`zheng2026twostage`.
+
+The most important items to verify are `hoeffding1963`,
+`boucheron2013concentration`, and `ghosh2002` because they support theory or
+inequality language; `goldfarb2003robustportfolio` and `delage2010dro` because
+they anchor robust-optimization positioning; and `das2023creditgraph`,
+`yang2025costaware`, `zheng2026twostage`, and `boosting2025default` because they
+are recent IJDS/credit references that make the manuscript look current.
+
+## Bottom line
+
+The core CRPTO spine is well supported by read/curated sources: conformal
+prediction/risk control, conformal robust optimization, SPO+/DFL comparison,
+Lending Club fintech context, and the new decision-vs-estimation framing.
+
+The weaker area is not the main claim; it is citation hygiene. Several classical
+or recent references are being used as positioning anchors without a local note
+proving close reading. They can stay for now, but they deserve a targeted
+pre-freeze source check rather than another broad literature expansion.

paper/CRPTO_ijds.qmd

@@ -158,9 +158,10 @@ interval artifact with DVC/MLflow-style lineage, paper tables, and model-risk
 documentation.
 
 The third foundation is predict-then-optimize and decision-focused learning.
-IJDS work on decision making warns that estimating a model and choosing an
-action are different objects, even before causal claims are introduced
-[@fernandezloria2022causaldecision]. SPO+ and modern decision-focused learning
+IJDS work on causal decision making sharpens the same warning: once an estimate
+feeds an action, the relevant target can become the assignment rule rather than
+only the intermediate effect-size estimate [@fernandezloria2022causaldecision].
+SPO+ and modern decision-focused learning
 ask models to respect the loss surface induced by the downstream decision
 [@elmachtoub2022; @donti2017; @mandi2024]. CRPTO is intentionally more
 conservative. It does not retrain the PD model end-to-end through the optimizer.

paper/submission/CRPTO_ijds_submission.tex

@@ -194,9 +194,9 @@ \section{Related Work}\label{sec:related}
 The third foundation is predict-then-optimize and decision-focused learning. SPO+
 and modern decision-focused learning ask models to respect the loss surface induced
 by the downstream decision \citep{elmachtoub2022,donti2017,mandi2024}. IJDS work on
-decision making also warns that estimating a model and choosing an action are
-different objects, even before causal claims are introduced
-\citep{fernandezloria2022causaldecision}. CRPTO is intentionally more
+causal decision making sharpens the same warning: once an estimate feeds an
+action, the relevant target can become the assignment rule rather than only the
+intermediate effect-size estimate \citep{fernandezloria2022causaldecision}. CRPTO is intentionally more
 conservative. It does not retrain the PD model end-to-end through the optimizer.
 Instead, it asks what can be achieved when a calibrated predictive system is
 already frozen and the decision layer must remain explainable to credit-risk