CRPTO

CRPTO is a research repository for one manuscript targeted to the INFORMS Journal on Data Science. It studies what happens when a binary conformal score is used as a coefficient in a monthly credit-allocation linear program. The contribution is an exact geometric analysis and identification audit of the machine-learning--conformal--optimization interface, not a promoted lending policy or a new credit-scoring leaderboard.

Active Result

The active design uses the Lending Club 2007--2020Q3 archive and declares all time roles before evaluation.

Quantity	Active value
Raw archive	2,925,493 rows
Eligible 36-month design universe	640,543 loans
Primary OOT candidates	376,890 loans
Resolved / unresolved at the six-month endpoint	364,814 / 12,076
Coverage controls	5 frozen learner specifications x 8 windows
Largest all-candidate coverage upper bound	0.897726
Joint-block rank diagnostic	31 / 40 cells meet the locked nominal thresholds
Individual-age origin sensitivity	16 / 16 upper bounds below 0.90
Label-Mondrian sensitivity	27 / 40 marginal and 109 / 400 category shortfalls
Closed CatBoost calibrator family	18 / 32 pooled upper endpoints below 0.90
Calibrator evaluation / pairwise cells	192 / 288
Two-ruler optimization solves	6,240
Registered finite point-cap support	3,067 caps
Finite registered-cap endpoint envelopes including zero	216 / 216
Structural sensitivity	36 complete scenarios
Joint fit-label completion stress	32 / 32 upper bounds below 0.90
USD 25 floor diagnostic	maximum rate perturbation 0.001284 pp
Set-preserving embedding audit	80 / 80 binary-set cells unchanged
Positive-gamma allocation changes	9,659 / 11,520 monthly contrasts
Fixed theta tracks with both separated signs	77 / 576
Clean binary-phase census	200 / 200 exact reconciliations; 87 below-half cells
Dual set-native certificates	208 / 208; 0 new optimization solves

Under the declared six-month outcome-availability rule, all 40 sharp all-candidate coverage upper bounds are below 0.90. That is a deterministic finite-archive fact, not by itself a rejection of split conformal. In a separate post-inspection combined-rank diagnostic, 31 of the 40 frozen learner-window cells meet the locked Bonferroni-within-cell and Holm-across-cell nominal thresholds. Its Beta--Binomial reference law assumes the stronger joint exchangeability of each calibration stratum together with its entire target block; it does not test or refute the usual one-future-point marginal split-conformal guarantee. Because the family and pattern were inspected before locking, the thresholds do not provide post-selection or study-wide FWER control. The primary CatBoost eight-window finding also recurs under three feature-semantics-preserving missingness encodings and at both April--June origins with cutoffs 39 months after each issue-month end (equal whole-month administrative age, not exact day-level age). Label-Mondrian calibration changes the 40 marginal states to 27 shortfalls, 12 crossings, and one lower endpoint at or above nominal; it is a complete sensitivity, not a repair. A separate closed four-map CatBoost sensitivity holds one common uncalibrated-probability (q_raw) taxonomy fixed. Platt and beta retain upper endpoints below 0.90 in 8/8 pooled windows, while isotonic and the IVAP Venn--Abers scalar do so in 1/8 each; because only 18/32 pooled cells remain below, a uniform closed-family shortfall is not established. This selects no calibrator, transfers no Venn guarantee to the scalar, and supplies no portfolio result. Portfolio direction changes with the outcome-blind ruler, coordinate, and comparator support; no model, encoding, gamma, ruler, coordinate, scenario, or policy is selected. Exact downstream score-order invariance requires positive-affine equivalence on the span of feasible allocation differences. Preserving binary sets or the coordinatewise loan ranking is not sufficient, while failure of that condition does not require every fixed cell to change.

These are retrospective, archive-specific identification results. They are not prospective validity, selected-set conformal coverage, causal lending effects, cash-flow returns, or deployment evidence.

Sources Of Truth

Read these in order:

1. docs/research/active_claims_2026-07-14.md
2. configs/ijds_active_evidence_sources.yaml
3. configs/ijds_claim_ledger.yaml
4. reports/crpto/ijds_binary_geometry_frontier_v4_evidence.json
5. paper/CRPTO_ijds.qmd
6. paper/supplement_ijds.qmd

The source registry owns lineage identities, 53 DVC pointers, and 11 scientific Git-native artifact lineages. The evidence builder emits 45 paper-facing CSV tables and five figure families. The evidence manifest is the only numeric paper-facing manifest. The body QMD is the only editable source for the official submission TeX.

Architecture

raw archive + frozen experiment roots
              |
              v
  configs/ijds_active_evidence_sources.yaml
              |
              v
  scripts/build_ijds_binary_geometry_frontier_v4_evidence.py
              |
              +--> reports/crpto/ijds_binary_geometry_frontier_v4_evidence.json
              +--> reports/crpto/tables/crpto_ijds_v4_*.csv
              +--> reports/crpto/figures/crpto_ijds_v4_*.{png,pdf}
              |
              v
  paper/CRPTO_ijds.qmd + paper/supplement_ijds.qmd
              |
              v
  generated HTML/PDF + official INFORMS TeX/PDF

Current reusable code is under src/ijds_audit, src/ijds_challengers, and the retained data, model, evaluation, and optimization modules they import. Current experiment entrypoints are exactly the scripts/experiments/run_ijds_* files named by configs/crpto_publication_targets.yaml.

dvc.yaml, dvc.lock, and paths fixed by EXTRACTION_MANIFEST.json form a sealed compatibility capsule. They preserve old hashes and path-bound replay metadata but are not active workflows or manuscript evidence. The complete pre-consolidation repository is archived outside the project at D:\crpto_legacy.

The current package and evidence flow is mapped in docs/ARCHITECTURE.md.

Setup

Requirements: Python 3.11 or 3.12, uv, just, Quarto, Git, and TeX Live.

uv sync --group dev --locked
just smoke

Use Windows PowerShell and uv run for Python commands. Known transitive advisories and their containment rules are recorded in docs/security/DEPENDENCY_RISK_REGISTER.md.

Main Commands

just test                    # complete local/Git-native suite; no DVC bytes required
just coverage                # local/Git-native suite plus branch coverage XML
just lint                    # Ruff check and format check
just type-check              # mypy
just type-check-fast         # blocking ty check on the active surface
just publication-integrity   # source, claim, and artifact contracts
just drift-gate              # read-only PD/conformal/evidence regression
just ijds-active-science-tests # auto-discovered IJDS tests that need no DVC bytes
just ijds-active-dvc-tests   # materialized IJDS tier; run only after just ijds-pull
just ijds-active-check       # scientific and manuscript synchronization
just submission-build       # evidence, HTML, TeX, PDFs, previews
just submission-check       # pre-freeze read-only gates; no page cap
just submission-freeze-check # final-only gate; enforces configured page cap
just submission-closeout    # build, check, and remote DVC verification
just dependency-audit        # fail on unregistered dependency advisories

The manual GitHub workflow named clean-clone-reproducibility (kept at .github/workflows/tests-full.yml to preserve its existing GitHub identity) reconstructs the locked publication capsule on a disposable Windows CPU runner. That operating system deliberately matches the canonical figure-generation surface because the workflow compares PNG and PDF evidence byte for byte. The lightweight lint workflow remains an independent Ubuntu portability check; CRPTO does not require WSL, CUDA, a GPU, or cuOpt. The clean-clone workflow pulls only active evidence and strict-manifest DVC targets, then runs the same claim, drift, test, and coverage gates used locally. It is a reproducibility audit, not the final reviewer-PDF page freeze. It also rebuilds the reviewer HTML, generated TeX, and deterministic machine-readable supplement. The official publisher PDF and browser-printed preview PDFs remain the Windows closeout surface because the INFORMS style kit and local Chromium installation are not distributed by this repository.

The manual fallback for official TeX compilation is intentionally pdflatex -> bibtex -> pdflatex -> pdflatex: the first pass writes citation and cross-reference metadata, BibTeX writes the bibliography, and the final two passes stabilize references and pagination.

Protected Boundary

Do not run these historical DVC stages without explicit permission:

• crpto.pd.champion
• crpto.conformal.intervals
• crpto.conformal.validation
• crpto.portfolio.optimization
• crpto.portfolio.bound_exact_eval

Do not modify EXTRACTION_MANIFEST.json or its protected model/data artifacts. Use just validate-champion for ordinary work and just validate-champion-strict when every protected artifact is available.

Paper Editing

Edit paper/CRPTO_ijds.qmd, then generate the official TeX with:

just paper-tex

Never edit paper/submission/CRPTO_ijds_submission.tex directly. The paper is double-anonymous; author-identifying material belongs only in the separate submission forms.

Data And Literature

The raw CSV and experiment roots are DVC-managed and excluded from Git. The local Papers_tesis/ corpus is also excluded because it contains copyrighted PDFs; bibliographic metadata belongs in paper/references.bib.

License And Citation

Code is MIT licensed. Narrative text, figures, tables, and documentation are currently governed by LICENSE-CONTENT. Because that prior public grant may conflict with a journal copyright transfer, publication rights remain a human-confirmation blocker rather than a resolved repository setting; see docs/research/ijds_submission_rights_anonymity_audit_2026-07-21.md. See CITATION.cff for citation metadata.