Promote validated IJDS selector v7

.codex/skills/crpto/SKILL.md

@@ -1,191 +1,125 @@
 # CRPTO Skill
 
-Use this skill when working inside `C:\Users\carlos\Documents\Paper_CRPTO`.
-CRPTO is a standalone academic paper project for Conformal Robust
-Predict-Then-Optimize in credit risk. The repository is public-facing, but the
-paper champion artifacts are frozen.
+Use this skill inside `C:\Users\carlos\Documents\Paper_CRPTO`. CRPTO is a
+single-author IJDS paper and reproducibility bundle, not a production service.
+Prefer simple code, frozen evidence, and one coherent manuscript claim.
 
-## Required Context
-
-Before structural work, read these files in order:
+## Read First
 
 1. `docs/ACADEMIC_CONTEXT.md`
 2. `docs/SCOPE_AND_GOVERNANCE.md`
 3. `docs/research/active_claims_2026-07-04.md`
 4. `CONTRIBUTING.md`
 5. `EXTRACTION_MANIFEST.md`
 6. `configs/crpto_publication_targets.yaml`
-7. `docs/research/README.md`
-
-Use the local project context over generic habits. This is a single-author,
-static-dataset, no-production academic repo. Keep code simple, functional, and
-close to the existing style.
-
-## Platform
-
-- Work Windows-first in PowerShell.
-- Prefer `uv run ...` for Python, Quarto, dbt, DVC, MLflow, pytest, and ruff.
-- The local venv is `.venv/Scripts/python.exe`.
-- Do not introduce Unix-only shell assumptions.
-
-## Champion Rules
-
-The current IJDS paper-facing body claim is the promoted pool93 finite-grid
-frontier closure. It is a deterministic policy-grid re-evaluation over the
-frozen upstream PD/calibration/conformal artifacts, not a retraining run.
-
-- Terminal run tag:
-  `champion-reopen-2026-06-19__pool93__ijds-claim-bound-terminal`
-- Body point source run:
-  `champion-reopen-2026-06-19__pool93__ijds-claim-micro-ext`
-- Policy family: `claim_micro_ext_body_cap345`
-- Policy mode: `capped_blended_uncertainty`
-- Robust return: `$184,832.48`
-- Return-floor surplus: `$14,367.94`
-- `V(alpha=0.01)`: `0.035350`
-- `Gamma_CP(alpha=0.01)`: `0.162616`
-- Endpoint budget upper at `alpha=0.01`: `0.24508374`
-- Markov cap at `alpha=0.01`: `0.34508374`
-- Exact alpha violation: `0.0`
-- Declared alpha-grid pass: `8/8`
-- Consolidated frontier: `50,010` deduplicated semantic policies, `27,508`
-  eligible all-alpha above-floor policies.
-- Terminal exact search: `37,068/37,068` all-alpha passers and `296,544`
-  completed exact candidate-alpha checks.
-
-The frozen upstream baseline remains retained for provenance and as the
-declared return floor:
-
-- Run tag: `ijds-rebaseline-2026-06-07`
-- Policy: `bound_aware_276k_economic_champion`
-- Robust return: `$170,464.54`
-- `V(alpha=0.01)`: `0.028875`
-- `Gamma_CP(alpha=0.01)`: `0.187987`
-- Exact pass: `True`
-- Former robust region: `45/45`
-
-The older run tag `paper-thesis-final-economic-2026-04-06` is historical
-provenance only. Do not use it as the active body claim.
-
-Never overwrite these frozen artifacts unless the user explicitly asks for a
-champion rebuild:
+
+Use Windows PowerShell and `uv run`. Do not introduce Unix-only workflow
+assumptions.
+
+## Active IJDS Policy
+
+- Run tag:
+  `champion-reopen-2026-06-19__pool93__ijds-calibration-selected-endpoint28-v7`
+- Exact conformal replay: target `alpha=0.10`, frozen used alpha `0.095`.
+- Decision score: `q=(p+u)/2`.
+- Risk tolerance: `tau=0.17`.
+- Objective: point-PD expected net return; conformal `q` is the risk guardrail.
+- Selector: nine round-number policies on November 2017; five satisfy full
+  budget, effective-PD, and deterministic `B_u<=0.28` screens. Outcomes are
+  stored separately from the 12-column selector frame.
+- Audit: an outcome-free December replay selects the same policy; opening
+  outcomes afterward gives weighted default `0.145650` and miscoverage
+  `0.124925`, so stability is not reported as selected-set validity.
+- Full OOT: 276,869 candidates, 308 funded, `$179,327.59` realized return,
+  `0.039375` weighted default, `0.036875` weighted miscoverage.
+- `Gamma_CP=0.176102`, `Gamma_residual=0.088051`, endpoint `0.258051`.
+- Markov sensitivity: threshold `0.574279` with probability bound `0.316228`
+  under weighted funded-set validity.
+- Matched point-PD: `$196,369.14`, `0.118400` weighted default, endpoint
+  `0.921317`, threshold `1.237545`.
+
+The exact paper-facing source is:
+
+`models/experiments/champion_reopen/<run_tag>/portfolio/ijds_policy_governance.json`
+
+A35--A40 provide exact-alpha, selector, temporal, grade, bootstrap, and matched
+comparison evidence. The final selector does not read OOT outcomes, but prior
+project development inspected the static OOT corpus. Say "retrospective
+lockbox replay," not "preregistered," "prospective," or "untouched holdout."
+
+Do not revive these as active claims: approximate alpha-0.01 scaling, `8/8`,
+the 50,010-policy frontier, `0.345084`, capped/tail-selected policies, or OOT
+outcome-selected hyperparameters.
+
+## Frozen Provenance
+
+Never overwrite the manifest-protected upstream or historical pool93 files:
 
 - `models/pd_canonical.cbm`
 - `models/pd_canonical_calibrator.pkl`
 - `models/final_project_promotion.json`
 - `models/conformal_policy_status.json`
 - `data/processed/conformal_intervals_mondrian.parquet`
 - `data/processed/portfolio_bound_aware/rank1_alpha01_bound_aware_276k_full_2026-04-05-1734/`
-- `reports/crpto/tables/crpto_tableA35_pool93_ijds_frontier.csv`
-- `reports/crpto/tables/crpto_tableA36_pool93_body_funded_grade_audit.csv`
-- `reports/crpto/tables/crpto_tableA37_pool93_body_tail_risk.csv`
-- `reports/crpto/tables/crpto_tableA38_pool93_body_cluster_bound_audit.csv`
-- `reports/crpto/tables/crpto_tableA39_pool93_body_bootstrap_metrics.csv`
-- `models/experiments/champion_reopen/champion-reopen-2026-06-19__pool93__ijds-claim-consolidated-definitive/portfolio/pool93_ijds_consolidated_governance.json`
-- `models/experiments/champion_reopen/champion-reopen-2026-06-19__pool93__ijds-claim-bound-terminal/portfolio/pool93_ijds_claim_governance.json`
+- historical `crpto_tableA35..A40_pool93_*` files
+- historical pool93 governance JSON files
 - `EXTRACTION_MANIFEST.json`
 
-Protected DVC stages:
-
-- Do not run without explicit permission: `crpto.portfolio.bound_exact_eval`
-  and any Optuna/HPO or policy search.
-- Treat as protected revalidation stages: `crpto.pd.champion`,
-  `crpto.conformal.intervals`, `crpto.conformal.validation`, and
-  `crpto.portfolio.optimization`. If the user explicitly permits them, run them
-  as a drift check and validate hashes before/after.
-- Safe paper stages: `crpto.paper.export_tables`, `crpto.paper.evidence`,
-  `crpto.paper.journal_package`, `crpto.paper.tail_satisficing_audit`,
-  `crpto.paper.figures`, `crpto.paper.spo_stability`, and `crpto.book.render`.
-
-If the user says a change may touch the champion, isolate the work:
-
-- Use a dedicated branch and a clear run label.
-- Write new outputs under a new path; do not replace frozen champion files.
-- Produce a drift report with expected tolerances.
-- Run `just validate-champion` before and after.
-- Do not update `EXTRACTION_MANIFEST.json` unless the user explicitly approves
-  a new frozen release.
-
-## Current Journal Scope
-
-The active paper scope is the IJDS pool93 certificate plus bounded diagnostics:
-
-- A19/Fig15: regret-auditability frontier in the main narrative.
-- A35: promoted pool93 finite-grid return-bound frontier.
-- A36: funded-set grade composition audit for the selected pool93 allocation.
-- A37: pool93 selected-allocation LGD/CVaR/OCE tail-risk repricing.
-- A38: pool93 selected-allocation cluster-bound sensitivity; Markov remains
-  the body theorem because the cluster thresholds are not tighter here.
-- A39: fixed-allocation bootstrap diagnostic; it does not resample model,
-  solver, conformal intervals or policy search.
-- A20--A22: legacy tail-risk/OCE/CVaR diagnostic package retained in the
-  supplement, not as the promoted pool93 selector.
-- A23--A24: multi-distribution/online coverage diagnostics, not universal
-  conditional-coverage claims.
-- A25--A34: Prosper/Freddie external economic replication, not new Lending Club
-  champions.
-
-Do not re-run champion search, HPO, conformal interval generation or protected
-portfolio stages to support this pack. Use existing artifacts unless the user
-explicitly asks for a new isolated experiment with a claim target, evidence
-gate, artifact sink and stop rule.
-
-## Objective Experiments
-
-OCE/CVaR/satisficing can be implemented as an isolated research objective or
-scoring layer. The default rule is:
-
-- It may read existing allocations, predictions, intervals, or shortlist
-  artifacts.
-- It may generate new diagnostic tables, figures, configs, tests, and docs.
-- It must not replace the champion objective, rank-1 policy, or frozen outputs.
-- A tail-satisficing challenger audit may read or re-score frozen allocations
-  under a new paper/audit stage if outputs are new and the status marks
-  `champion_promotion_changed=false`. It must not replace the A35 pool93 body
-  selector without a new promotion protocol.
-- If used for a new search, store results under a new experiment path and make
-  the comparison explicit against the active pool93 body claim and the frozen
-  upstream rebaseline.
-
-## Submission Closeout
-
-For the current submission, keep these gates visible:
-
-- Consolidate A19/Fig15, A20--A39, paper/supplement, docs, manifest hashes and
-  `dvc.lock`.
-- Sweep the manuscript for stale numbers, captions, body-vs-appendix placement,
-  and IJDS length.
-- Convert the final `.qmd` into the official IJDS LaTeX template when the PDF
-  is ready for submission.
-- Keep GitHub/DagsHub/MLflow disclosure compatible with double-anonymous
-  review: anonymize in the manuscript, reveal in cover letter or after review
-  according to venue policy.
-- Create the reproducible release tag/bundle only after the final PDF passes.
-- Treat `dvc push` as optional and credential-dependent.
-
-## Standard Checks
-
-Use focused checks while editing, then close with the strongest feasible set:
+New experiments must use a distinct run tag and write under
+`data/processed/experiments/champion_reopen/` and
+`models/experiments/champion_reopen/`. Never replace frozen paths.
+
+Protected DVC stages are `crpto.pd.champion`,
+`crpto.conformal.intervals`, `crpto.conformal.validation`,
+`crpto.portfolio.optimization`, and `crpto.portfolio.bound_exact_eval`.
+Experiments may read their outputs. Run a protected stage only with explicit
+permission and a drift report.
+
+## Method Boundary
+
+The submitted method has one linear policy. Capped, tail, OCE/CVaR, SPO+,
+multi-distribution, online, causal, and external-data variants are comparators
+or diagnostics, not additional CRPTO methods.
+
+Keep these distinctions explicit:
+
+- exact conformal quantile replay versus approximate width scaling;
+- point PD in the economic objective versus conformal `q` in the constraint;
+- deterministic `weighted outcome <= B_u + V` versus the
+  assumption-conditional Markov statement;
+- calibration-only final ranking versus historical OOT-aware development;
+- full-OOT averages versus temporal heterogeneity;
+- retrospective contrasts versus causal or universal dominance.
+
+## Paper Workflow
+
+Safe paper work may regenerate active A35--A40, figures, Quarto outputs, and the
+official IJDS PDF. Keep body, supplement, submission TeX, governance JSON, and
+claim-sync tests numerically aligned.
+
+Standard closeout:
 
 ```powershell
-uv run python scripts/build_crpto_journal_package.py
-uv run pytest tests/test_publication_targets.py -q
-uv run pytest tests/test_scripts/test_build_crpto_journal_package.py -q
-uv run pytest tests/test_pool93_body_claim_sync.py -q
-just smoke
+uv run python scripts/build_ijds_calibration_selected_evidence.py
+uv run pytest tests/test_ijds_active_claim_sync.py -q
 just lint
+just type-check
+just type-advisory-full
+just smoke
 just validate-champion
 just paper-submission
+just paper-submission-official
 uv run dvc status --no-updates
 ```
 
-For a full milestone, also run `just test`, `just type-check`, and the relevant
-Quarto render. Do not bypass hooks.
+Run `just drift-gate` after changes to conformal or PD semantics. Do not bypass
+hooks, commit secrets, or alter `EXTRACTION_MANIFEST.json`.
 
-## Writing And Docs
+## Writing
 
-- Spanish for book, paper prose, and research notes.
-- English for code, identifiers, CI, and docstrings.
-- Do not reformat the whole book in one pass.
-- Avoid broad refactors unless they directly reduce risk for the current task.
-- Never commit secrets, `.env`, local credentials, or heavyweight raw data.
+- Paper and code identifiers are English; project notes may be Spanish.
+- Lead with data, method, decision, and managerial implication.
+- Report the price of robustness and temporal failures as prominently as wins.
+- Treat reproducibility as evidence quality, not as the sole novelty.
+- Keep the main IJDS body within 25 pages; move proofs and diagnostics to the
+  separate supplement.

.github/workflows/tests-full.yml

@@ -72,6 +72,7 @@ jobs:
             tests/test_features/ \
             tests/test_crpto_final_sync.py \
             tests/test_quarto_book_guardrails.py \
+            tests/test_publication_integrity.py \
             -q
 
       - name: Upload coverage if generated

.gitignore

@@ -92,6 +92,8 @@ models/feature_artifacts_runtime_status.json
 
 # Reports (large generated outputs)
 reports/run_logs/
+reports/api-docs/
+reports/ci/
 reports/notebook_images/
 reports/paper_material/
 reports/crpto/experiments/
@@ -149,6 +151,9 @@ models/experiments/**
 !models/experiments/champion_reopen/champion-reopen-2026-06-19__pool93__ijds-claim-bound-terminal/portfolio/pool93_ijds_claim_governance.json
 !models/experiments/champion_reopen/champion-reopen-2026-06-19__pool93__ijds-claim-consolidated-definitive/portfolio/pool93_ijds_consolidated_governance.json
 !models/experiments/champion_reopen/champion-reopen-2026-06-19__pool93__ijds-claim-consolidated-definitive/portfolio/pool93_ijds_consolidated_frontier.json
+!models/experiments/champion_reopen/champion-reopen-2026-06-19__pool93__ijds-certificate-semantics-v2/portfolio/pool93_ijds_consolidated_frontier.json
+!models/experiments/champion_reopen/champion-reopen-2026-06-19__pool93__ijds-certificate-semantics-v2/portfolio/pool93_ijds_consolidated_governance.json
+!models/experiments/champion_reopen/champion-reopen-2026-06-19__pool93__ijds-certificate-semantics-v2/portfolio/pool93_point_pd_baseline_audit.json
 configs/experiments/
 !configs/experiments/
 configs/experiments/*

.pre-commit-config.yaml

@@ -47,15 +47,15 @@ repos:
         stages: [pre-push]
 
       - id: pytest-smoke
-        name: pytest smoke (final-sync + quarto guardrails)
-        entry: uv run pytest tests/test_crpto_final_sync.py tests/test_quarto_book_guardrails.py -q
+        name: just smoke (publication guardrails)
+        entry: just smoke
         language: system
         pass_filenames: false
         stages: [pre-push]
 
       - id: validate-champion
         name: validate champion artefacts
-        entry: uv run python -c "import json; from pathlib import Path; m=json.loads(Path('EXTRACTION_MANIFEST.json').read_text()); print('manifest OK, keys:', len(m))"
+        entry: just validate-champion
         language: system
         pass_filenames: false
         stages: [pre-push]

AGENTS.md

@@ -10,8 +10,8 @@ Lectura obligatoria, en este orden:
 1. [`CLAUDE.md`](CLAUDE.md) — reglas de operación, champion congelado
    (esquema dual-tag pool93), stages prohibidos, comandos y convenciones.
 2. [`docs/research/active_claims_2026-07-04.md`](docs/research/active_claims_2026-07-04.md)
-   — registro vigente de claims del paper (body point pool93, denominadores
-   finite-grid, reopen gate).
+   — registro vigente de claims del paper (replay exacto al 90%, selector de
+   calibración 3x3, política lineal 50/50 y reopen gate).
 3. [`.codex/skills/crpto/SKILL.md`](.codex/skills/crpto/SKILL.md) — skill de
    Codex con el detalle operativo del certificado y sus artefactos.