Integrate IJDS full-data credit controls and comparator audit

[EigenCharlie]

Summary

• Integrates the active V4 binary-geometry audit, the locked V1c/V2 two-ruler diagnostic, and a full-data credit-risk control lineage into one IJDS manuscript and evidence manifest.
• Keeps CRPTO as one ML -> conformal -> optimization research object, framed as an identification audit rather than a winning credit policy.
• Uses all rows eligible under the declared term, date, schema, maturity, and observability contracts; there is no computational sample or OOT model selection.
• Remains deliberately pre-freeze: no learner, window, gamma, ruler, coordinate, policy, cap, or comparator is promoted.

Full-data contract and credit controls

A chunked audit reads all 2,925,493 raw rows and 142 columns. The archive contains 2,060,077 36-month and 865,415 60-month loans. The active 640,543-loan population is exhaustive for the declared 36-month binary terminal-default estimand:

• PD development: 17,433 rows.
• 2011 Platt/taxonomy block: 14,101 rows.
• Residual pool: 49,007 rows.
• Outcome-free policy development: 94,885 rows.
• Primary OOT: 376,890 rows, including 11,551 unresolved outcomes.
• Censored extension: 88,227 rows.

Using the remaining archive as ordinary labeled training data would change the estimand or condition on maturity: 60-month loans have a different horizon, 48 fields are late-schema, and only 36.85%, 10.20%, and 1.15% of the 2014, 2015, and 2016Q1 cohorts have labels available by the March 2016 cutoff.

Five predeclared full-data scores are frozen before OOT outcomes:

1. active CatBoost + Platt;
2. numeric logistic + Platt;
3. CatBoost with 20 domain-signed monotonic constraints + Platt;
4. a 26-field platform-signal OptBinning WOE/IV scorecard + Platt; and
5. a 19-field borrower-only WOE/IV scorecard + Platt.

All 45 OptBinning problems are OPTIMAL. Platform pricing/grade dominates its IV ranking, while FICO and recent inquiries lead the borrower-only model. The controls never enter portfolio optimization and are not a predictive leaderboard.

Scientific result

All 40 model-window all-candidate coverage upper bounds are below 0.90:

• CatBoost maximum: 0.882167.
• Numeric logistic maximum: 0.895654.
• Monotonic CatBoost maximum: 0.885991.
• Platform WOE maximum: 0.894317.
• Borrower-only WOE maximum: 0.896973.

OOT AUC ranges from 0.612712 to 0.652244. Every calibration slope is below one and every calibration-in-the-large value is negative. Numeric logistic has development-to-OOT score PSI 0.005168 yet still fails coverage in every window, so marginal score stability alone is not a validity certificate.

CatBoost stratum 2 crosses alpha=0.10 from prevalence 0.101703 to 0.097147; its residual quantile changes from 0.888435 to 0.111801 and mean width from 0.984263 to 0.207631.

The locked two-ruler diagnostic contains 6,240 solves and 622,455 funded rows. Objective-matched coordinate .25 is one repeated four-month contrast: 44 loan-month positions, USD 155,937.27 one-way turnover, USD 5,603.66 higher standardized payoff, and 0.0068 pp lower default/miscoverage. Coordinate .50 is adverse and .75 is mostly unidentified for payoff/default. All normalized-score tracks are adverse, but do not match plug-in opportunity cost.

The exact point-cap frontier contains 3,067 caps. All 216 broad-stress envelopes cross zero, as do all 72 development-support default envelopes and all 27 W8 development envelopes. No universal policy direction is active.

Theory and claim boundary

The manuscript formalizes the binary residual phase transition, positive-affine cap equivalence, normalized-ruler invariance, same-cap nesting, C2 plug-in dominance, sharp common-outcome bounds, and basis-endpoint sufficiency over declared supports.

It does not claim selected-set conformal validity, latent-PD confidence intervals, causal effects, cash-flow return, a Markov certificate, prospective confirmation, a continuous joint frontier, scorecard superiority, universal robustness, fair-lending certification, or deployment readiness.

Implementation and reproducibility

• One evidence builder verifies three immutable outcome-free/evaluation pairs and emits one active manifest, ten tables, and six figure files.
• Twelve DVC pointers preserve V4-v1/v2, two-ruler V1c/V2, and credit-control V1b/V2b.
• The raw CSV pointer is healthy; the local file has the declared 1,773,470,505 bytes and DVC reports it up to date.
• The manuscript, supplement, generated INFORMS TeX, claim registry, disclosure surfaces, book corrections, governance, and CRPTO skill are synchronized.
• The active DVC capsule and dagshub remote are in sync; 26 new objects were pushed.
• Core active scientific packages are already current in the resolved environment. Major pandas/PyArrow/OR-Tools/mypy upgrades are documented as new-run sensitivities, not invisible maintenance.

Validation

• just submission-check: pass.
• Full pytest suite: pass with two expected skips.
• Ruff/format: 299 files clean.
• Mypy: 316 source files clean.
• Full pinned ty: zero diagnostics.
• Strict historical manifest gate: pass.
• Champion drift gate: exact zero drift for predictions, intervals, score edges, coverage cells, and floor multipliers.
• just smoke and just book: pass.
• Official INFORMS PDF: 29 pages; references begin on page 25, with 24 complete pre-reference pages and the conclusion on page 25.
• Preview PDFs: body 22 pages, supplement 22 pages.
• Main body: five tables, two figures, 218-word abstract, seven keywords.
• BibTeX: zero warnings; no undefined citations or references.
• Every page of all three PDFs was rendered and visually inspected; no clipping, overlap, blank page, broken table, missing glyph, or illegible figure was found.

Boundaries

No protected DVC stage was executed, EXTRACTION_MANIFEST.json was not modified, and no frozen champion or protocol artifact was overwritten. The rolling-origin evidence was deterministically rebuilt only to refresh two aggregation-source hashes; all derived tables and metrics remained byte-identical.

This PR remains draft. Identity insertion, clean-clone release rehearsal, submission-week guideline recheck, sanitized upload archive, immutable release tag, and ScholarOne proof remain intentionally deferred until an explicit freeze decision.