closed-loop-default-detection 0.1.0__tar.gz → 0.2.0__tar.gz

{closed_loop_default_detection-0.1.0 → closed_loop_default_detection-0.2.0}/PKG-INFO

@@ -1,7 +1,7 @@
 Metadata-Version: 2.4
 Name: closed-loop-default-detection
-Version: 0.1.0
-Summary: CLUE-style closed loop that measures selective-labels default detection on synthetic SMB lending cohorts and finds the PD model's operating frontier.
+Version: 0.2.0
+Summary: Closed loop that measures selective-labels default detection on synthetic SMB lending cohorts and finds the PD model's operating frontier.
 Author-email: Hossain Pazooki <hossain@pazooki.com>
 License-Expression: MIT
 Project-URL: Homepage, https://github.com/hossainpazooki/closed-loop-default-detection
@@ -55,7 +55,7 @@ hides labels the way real approval policies do, and grades every correction agai
 
 The loop escalates selection severity until correction fails and reports the **operating
 frontier** — the last severity at which declined-cohort calibration still holds (target
-ECE ≤ 0.10). From the committed runs (`artifacts/clue_frontier*.csv`, seed 42):
+ECE ≤ 0.10). From the committed runs (`artifacts/loop_frontier*.csv`, seed 42):
 
 | Selection severity | 0.0 | 0.2 | 0.4 | 0.6 |
 |---|---|---|---|---|
@@ -63,17 +63,90 @@ ECE ≤ 0.10). From the committed runs (`artifacts/clue_frontier*.csv`, seed 42)
 | **IPW-corrected** (flat world) | 0.020 | 0.038 | **0.086 ✓** | **0.154 ✗** |
 | **IPW-corrected** (SCM world) | 0.036 | 0.038 | **0.097 ✓** | **0.244 ✗** |
 
-Both worlds land the frontier at **severity 0.4**, and the counterfactual deliverable breaks at
-the same boundary: across 25 seeds, g-computation cuts strong-propagation counterfactual MAE
-from 0.099 to 0.086 (−13.5%, positive on 24/25 seeds, Wilcoxon p = 1.5e-7) *inside* the
-frontier — and collapses to a negligible +0.0017 at full severity, where **no deployable
-advantage is claimed**. One cause explains both: selection through an **unobserved
-confounder**, which backdoor adjustment and IPW cannot fix. That single measured limit — not
-an unverifiable score — is the deliverable.
+On **this seed** both worlds land the frontier at severity 0.4, and the counterfactual
+deliverable breaks at the same boundary: across 25 seeds, g-computation cuts
+strong-propagation counterfactual MAE from 0.099 to 0.086 (−13.5%, positive on 24/25 seeds,
+Wilcoxon p = 1.5e-7) *inside* the frontier — and collapses to a negligible +0.0017 at full
+severity, where **no deployable advantage is claimed**. One cause explains both: selection
+through an **unobserved confounder**, which backdoor adjustment and IPW cannot fix. That
+single measured limit — not an unverifiable score — is the deliverable.
+
+### The frontier is a distribution, not a point (v2)
+
+A frontier quoted from one seed is a figure published without an error bar. `v0.2.0` runs the
+loop across the full 25-seed set (`artifacts/frontier_sweep.csv`) — and **seed 42 turns out to
+sit at the optimistic end**:
+
+| World | min | median | max | seeds at 0.4 | seeds at 0.2 |
+|---|---|---|---|---|---|
+| flat | 0.2 | **0.4** | 0.4 | 15/25 | 10/25 |
+| SCM | 0.2 | **0.2** | 0.4 | 11/25 | 14/25 |
+
+In the SCM world the *majority of seeds fail one step earlier than the headline*: the median
+frontier is **0.2**, and seed 42's 0.4 is a minority outcome (11/25). The honest statement is
+that the operating frontier is **0.2–0.4 depending on the draw**, not a clean 0.4 — the
+single-seed table above is a valid instance of it, not its center. Nothing about the
+mechanism changes (the unobserved confounder still explains the failure); what changes is how
+precisely the boundary can be quoted.
+
+*Caveat on the sweep:* loop seed `s` consumes generator seeds `s..s+7`, and the 25-seed set has
+gaps smaller than 8, so some runs share feature draws. No two runs duplicate a cohort, but the
+25 rows are not fully independent — the set is kept for comparability with the counterfactual
+sweep that uses it.
 
 Reproduce the headline from committed evidence: `python scripts/paired_significance.py`.
 The full independent assessment (methodology, all numbers, what didn't hold) is the
-accompanying article, [`FABLE.md`](FABLE.md).
+accompanying article, [`docs/assessment.md`](docs/assessment.md) — a **dated snapshot**, written
+against the single-seed frontier and not retro-fitted with the distribution above.
+
+## Pricing the frontier (v2)
+
+Calibration says *whether* the model is wrong on the declines; it does not say **what being
+wrong costs**. v2 adds an expected-maximum-profit (EMP) reporting axis — computed from the
+same in-process scores the loop already measures, and **never** a loop-control input (ECE
+still decides pass/fail and the frontier). Two variants run side by side:
+
+| Variant | Parameters | What it answers |
+|---|---|---|
+| `empc` — literature EMPC | Verbraken et al. (2014) prior: `p0=0.55, p1=0.10, ROI=0.2644` | benchmark-comparable: what the standard measure says |
+| `emp_h` — harness-derived | this harness's own economics + **planted** default timing | what the loan structure actually pays |
+
+**The two disagree — and the disagreement is the finding.** In the SCM world they move in
+opposite directions as selection severity rises (from `artifacts/loop_frontier_scm.csv`,
+declined subpopulation, seed 42):
+
+| Selection severity | 0.0 | 0.2 | 0.4 | 0.6 |
+|---|---|---|---|---|
+| Literature `empc` (naive) | 0.0217 | 0.0331 | 0.0416 | **0.0420 ↑** |
+| Harness `emp_h` (naive) | 0.0387 | 0.0297 | 0.0139 | **0.0024 ↓** |
+
+The convenience prior **misprices this loan structure**: it assumes `ROI = 0.2644` where a
+60-day daily-ACH loan actually returns `0.0875` (3% origination fee + 5.75% term interest) —
+a **3.0× overstatement** — and puts **55% of defaults at full recovery where the harness plants
+1.2%** (prior mean λ = 0.275 vs harness 0.419). Priced honestly, the profit on the declined
+pool collapses to near-zero at the frontier; priced by the literature prior, it appears to
+*grow*. Same model, same scores, opposite conclusion.
+
+**Reading caveats — the boundary of what this measures:**
+
+- **Raw EMP moves with world hardness.** A riskier declined pool changes EMP even for a
+  perfect model, so a cross-severity EMP delta is *not* pure model signal. Read the two
+  variants against each other at fixed severity, not the trend in isolation.
+- **`emp_h` rests on unfitted timing.** `days_to_default` is planted but independent of
+  features and risk given default — spec-shaped, never validated against real recovery data.
+  So `emp_h` is a **verified experiment, not a verified result**: the arithmetic is exact, the
+  timing distribution is an assumption.
+- **Post-term defaults are imputed.** ~22.5% of planted defaults land at day 90, past the
+  60-day term, and the generator does not model their payment history; they are priced at the
+  cohort's mean in-term loss fraction (a stated convention, not measured truth).
+- **`emp_h` is SCM-only.** The flat generator plants no timing, so its `emp_h` columns are
+  empty by design; only `empc` is reported there.
+
+Exploration now carries a price tag too. From `artifacts/exploration_frontier.csv` (SCM,
+seed 42, 10% budget, severity 0.6): the lever buys **157 labels for $439,578** — about
+**$2,800 per label**, of which 56 are realized defaults. That is the cost of *identification*,
+in dollars a lender can see, on the only lever that buys it rather than reweighting what is
+already identified.
 
 ## Install
 
@@ -139,6 +212,7 @@ Everything is importable from top-level `cldd` (full reference: the [Sphinx docs
 | `SyntheticBorrowerGenerator`, `StructuralBorrowerGenerator` | the flat and fitted-SCM synthetic worlds |
 | `run_counterfactual_eval`, `GComputationEstimator` | counterfactual validator (g-computation vs naive conditioning) |
 | `FeedbackLoop` | model-in-the-loop selective-labels simulation |
+| `empc_literature`, `emp_harness` | the two EMP variants — price the frontier in profit (reporting only, never loop control) |
 | `positivity_diagnostics` | observable regime/drift alarm — needs **no** declined-row labels |
 | `CalibratedPDClassifier` | the calibrated PD detector as a scikit-learn estimator |
 | `cldd.fidelity.run_fidelity_gate` | SCM-vs-real **marginal**-fidelity gate (univariate marginals only) |
@@ -163,7 +237,8 @@ scores = cross_val_score(CalibratedPDClassifier(random_state=42), X, y, scoring=
 Each driver runs without install (adds `src/` to the path) and writes to `artifacts/`:
 
 ```bash
-python scripts/run_clue.py                    # the closed loop → frontier table + plot (--generator scm for the SCM world)
+python scripts/run_loop.py                    # the closed loop → frontier table + plot (--generator scm for the SCM world)
+python scripts/run_frontier_sweep.py --quick  # frontier distribution across the 25-seed set (drop --quick for all)
 python scripts/run_seed_sweep.py --quick      # counterfactual certification (drop --quick for all seeds)
 python scripts/run_reject_inference.py        # reject-inference levers vs the frontier
 python scripts/run_exploration_sweep.py       # frontier vs exploration budget
@@ -173,7 +248,7 @@ python scripts/paired_significance.py         # recompute the headline stat from
 
 ## Validation
 
-`pytest` — 123 tests, all synthetic, no real data needed. CI runs a pinned-repro job (exact
+`pytest` — 149 tests, all synthetic, no real data needed. CI runs a pinned-repro job (exact
 pins), a cross-version/OS compat matrix, and a strict docs build. Six float-sensitive tests
 reproduce only under the pins in `requirements-dev.txt`; the optional marginal-fidelity gate
 compares the SCM against a **private** real dataset via `CLDD_DATA_DIR` and is the only thing
@@ -189,17 +264,21 @@ that needs it. Details, reproducibility, and troubleshooting:
 | [docs/configuration.md](docs/configuration.md) | every knob (`config.py`) and the one env var |
 | [docs/validation.md](docs/validation.md) | tests, gates, reproducibility, troubleshooting |
 | [docs/reject_inference.md](docs/reject_inference.md) | the four RI methods and their honest (modest) results |
-| [`FABLE.md`](FABLE.md) | **the accompanying article** — independent results & methodology assessment |
+| [`docs/assessment.md`](docs/assessment.md) | **the accompanying article** — independent results & methodology assessment (dated snapshot) |
 
 Build locally: `pip install -e ".[docs]" && sphinx-build -b html -W docs docs/_build/html`.
 
 ## Status
 
-`0.1.0` **alpha** on [PyPI](https://pypi.org/project/closed-loop-default-detection/),
-changelog in [CHANGELOG.md](CHANGELOG.md). Shipped: the loop, both worlds, all levers, the
-fidelity gate, the sklearn estimator, CI on three gates. CLDD began as a validation harness
-for the Intuit TechWeek SMB Underwriting Challenge; it is not a submission and does not
-alter challenge files.
+`0.2.0` **alpha** on [PyPI](https://pypi.org/project/closed-loop-default-detection/),
+changelog in [CHANGELOG.md](CHANGELOG.md).
+
+Shipped in 0.1.0: the loop, both synthetic worlds, all correction levers, the fidelity gate,
+the sklearn estimator, CI on three gates. Added in 0.2.0: `cldd.emp` (both EMP variants),
+priced exploration, the EMP-optimal cutoff, and the 25-seed frontier sweep.
+
+CLDD began as a validation harness for the Intuit TechWeek SMB Underwriting Challenge; it is
+not a submission and does not alter challenge files.
 
 ## Citation
 
@@ -211,7 +290,7 @@ Metadata in [`CITATION.cff`](CITATION.cff) (GitHub's "Cite this repository" read
   title   = {{closed-loop-default-detection}: measuring selective-labels default
              detection and the PD model's operating frontier},
   year    = {2026},
-  version = {0.1.0},
+  version = {0.2.0},
   license = {MIT},
   url     = {https://github.com/hossainpazooki/closed-loop-default-detection}
 }

{closed_loop_default_detection-0.1.0 → closed_loop_default_detection-0.2.0}/README.md

@@ -21,7 +21,7 @@ hides labels the way real approval policies do, and grades every correction agai
 
 The loop escalates selection severity until correction fails and reports the **operating
 frontier** — the last severity at which declined-cohort calibration still holds (target
-ECE ≤ 0.10). From the committed runs (`artifacts/clue_frontier*.csv`, seed 42):
+ECE ≤ 0.10). From the committed runs (`artifacts/loop_frontier*.csv`, seed 42):
 
 | Selection severity | 0.0 | 0.2 | 0.4 | 0.6 |
 |---|---|---|---|---|
@@ -29,17 +29,90 @@ ECE ≤ 0.10). From the committed runs (`artifacts/clue_frontier*.csv`, seed 42)
 | **IPW-corrected** (flat world) | 0.020 | 0.038 | **0.086 ✓** | **0.154 ✗** |
 | **IPW-corrected** (SCM world) | 0.036 | 0.038 | **0.097 ✓** | **0.244 ✗** |
 
-Both worlds land the frontier at **severity 0.4**, and the counterfactual deliverable breaks at
-the same boundary: across 25 seeds, g-computation cuts strong-propagation counterfactual MAE
-from 0.099 to 0.086 (−13.5%, positive on 24/25 seeds, Wilcoxon p = 1.5e-7) *inside* the
-frontier — and collapses to a negligible +0.0017 at full severity, where **no deployable
-advantage is claimed**. One cause explains both: selection through an **unobserved
-confounder**, which backdoor adjustment and IPW cannot fix. That single measured limit — not
-an unverifiable score — is the deliverable.
+On **this seed** both worlds land the frontier at severity 0.4, and the counterfactual
+deliverable breaks at the same boundary: across 25 seeds, g-computation cuts
+strong-propagation counterfactual MAE from 0.099 to 0.086 (−13.5%, positive on 24/25 seeds,
+Wilcoxon p = 1.5e-7) *inside* the frontier — and collapses to a negligible +0.0017 at full
+severity, where **no deployable advantage is claimed**. One cause explains both: selection
+through an **unobserved confounder**, which backdoor adjustment and IPW cannot fix. That
+single measured limit — not an unverifiable score — is the deliverable.
+
+### The frontier is a distribution, not a point (v2)
+
+A frontier quoted from one seed is a figure published without an error bar. `v0.2.0` runs the
+loop across the full 25-seed set (`artifacts/frontier_sweep.csv`) — and **seed 42 turns out to
+sit at the optimistic end**:
+
+| World | min | median | max | seeds at 0.4 | seeds at 0.2 |
+|---|---|---|---|---|---|
+| flat | 0.2 | **0.4** | 0.4 | 15/25 | 10/25 |
+| SCM | 0.2 | **0.2** | 0.4 | 11/25 | 14/25 |
+
+In the SCM world the *majority of seeds fail one step earlier than the headline*: the median
+frontier is **0.2**, and seed 42's 0.4 is a minority outcome (11/25). The honest statement is
+that the operating frontier is **0.2–0.4 depending on the draw**, not a clean 0.4 — the
+single-seed table above is a valid instance of it, not its center. Nothing about the
+mechanism changes (the unobserved confounder still explains the failure); what changes is how
+precisely the boundary can be quoted.
+
+*Caveat on the sweep:* loop seed `s` consumes generator seeds `s..s+7`, and the 25-seed set has
+gaps smaller than 8, so some runs share feature draws. No two runs duplicate a cohort, but the
+25 rows are not fully independent — the set is kept for comparability with the counterfactual
+sweep that uses it.
 
 Reproduce the headline from committed evidence: `python scripts/paired_significance.py`.
 The full independent assessment (methodology, all numbers, what didn't hold) is the
-accompanying article, [`FABLE.md`](FABLE.md).
+accompanying article, [`docs/assessment.md`](docs/assessment.md) — a **dated snapshot**, written
+against the single-seed frontier and not retro-fitted with the distribution above.
+
+## Pricing the frontier (v2)
+
+Calibration says *whether* the model is wrong on the declines; it does not say **what being
+wrong costs**. v2 adds an expected-maximum-profit (EMP) reporting axis — computed from the
+same in-process scores the loop already measures, and **never** a loop-control input (ECE
+still decides pass/fail and the frontier). Two variants run side by side:
+
+| Variant | Parameters | What it answers |
+|---|---|---|
+| `empc` — literature EMPC | Verbraken et al. (2014) prior: `p0=0.55, p1=0.10, ROI=0.2644` | benchmark-comparable: what the standard measure says |
+| `emp_h` — harness-derived | this harness's own economics + **planted** default timing | what the loan structure actually pays |
+
+**The two disagree — and the disagreement is the finding.** In the SCM world they move in
+opposite directions as selection severity rises (from `artifacts/loop_frontier_scm.csv`,
+declined subpopulation, seed 42):
+
+| Selection severity | 0.0 | 0.2 | 0.4 | 0.6 |
+|---|---|---|---|---|
+| Literature `empc` (naive) | 0.0217 | 0.0331 | 0.0416 | **0.0420 ↑** |
+| Harness `emp_h` (naive) | 0.0387 | 0.0297 | 0.0139 | **0.0024 ↓** |
+
+The convenience prior **misprices this loan structure**: it assumes `ROI = 0.2644` where a
+60-day daily-ACH loan actually returns `0.0875` (3% origination fee + 5.75% term interest) —
+a **3.0× overstatement** — and puts **55% of defaults at full recovery where the harness plants
+1.2%** (prior mean λ = 0.275 vs harness 0.419). Priced honestly, the profit on the declined
+pool collapses to near-zero at the frontier; priced by the literature prior, it appears to
+*grow*. Same model, same scores, opposite conclusion.
+
+**Reading caveats — the boundary of what this measures:**
+
+- **Raw EMP moves with world hardness.** A riskier declined pool changes EMP even for a
+  perfect model, so a cross-severity EMP delta is *not* pure model signal. Read the two
+  variants against each other at fixed severity, not the trend in isolation.
+- **`emp_h` rests on unfitted timing.** `days_to_default` is planted but independent of
+  features and risk given default — spec-shaped, never validated against real recovery data.
+  So `emp_h` is a **verified experiment, not a verified result**: the arithmetic is exact, the
+  timing distribution is an assumption.
+- **Post-term defaults are imputed.** ~22.5% of planted defaults land at day 90, past the
+  60-day term, and the generator does not model their payment history; they are priced at the
+  cohort's mean in-term loss fraction (a stated convention, not measured truth).
+- **`emp_h` is SCM-only.** The flat generator plants no timing, so its `emp_h` columns are
+  empty by design; only `empc` is reported there.
+
+Exploration now carries a price tag too. From `artifacts/exploration_frontier.csv` (SCM,
+seed 42, 10% budget, severity 0.6): the lever buys **157 labels for $439,578** — about
+**$2,800 per label**, of which 56 are realized defaults. That is the cost of *identification*,
+in dollars a lender can see, on the only lever that buys it rather than reweighting what is
+already identified.
 
 ## Install
 
@@ -105,6 +178,7 @@ Everything is importable from top-level `cldd` (full reference: the [Sphinx docs
 | `SyntheticBorrowerGenerator`, `StructuralBorrowerGenerator` | the flat and fitted-SCM synthetic worlds |
 | `run_counterfactual_eval`, `GComputationEstimator` | counterfactual validator (g-computation vs naive conditioning) |
 | `FeedbackLoop` | model-in-the-loop selective-labels simulation |
+| `empc_literature`, `emp_harness` | the two EMP variants — price the frontier in profit (reporting only, never loop control) |
 | `positivity_diagnostics` | observable regime/drift alarm — needs **no** declined-row labels |
 | `CalibratedPDClassifier` | the calibrated PD detector as a scikit-learn estimator |
 | `cldd.fidelity.run_fidelity_gate` | SCM-vs-real **marginal**-fidelity gate (univariate marginals only) |
@@ -129,7 +203,8 @@ scores = cross_val_score(CalibratedPDClassifier(random_state=42), X, y, scoring=
 Each driver runs without install (adds `src/` to the path) and writes to `artifacts/`:
 
 ```bash
-python scripts/run_clue.py                    # the closed loop → frontier table + plot (--generator scm for the SCM world)
+python scripts/run_loop.py                    # the closed loop → frontier table + plot (--generator scm for the SCM world)
+python scripts/run_frontier_sweep.py --quick  # frontier distribution across the 25-seed set (drop --quick for all)
 python scripts/run_seed_sweep.py --quick      # counterfactual certification (drop --quick for all seeds)
 python scripts/run_reject_inference.py        # reject-inference levers vs the frontier
 python scripts/run_exploration_sweep.py       # frontier vs exploration budget
@@ -139,7 +214,7 @@ python scripts/paired_significance.py         # recompute the headline stat from
 
 ## Validation
 
-`pytest` — 123 tests, all synthetic, no real data needed. CI runs a pinned-repro job (exact
+`pytest` — 149 tests, all synthetic, no real data needed. CI runs a pinned-repro job (exact
 pins), a cross-version/OS compat matrix, and a strict docs build. Six float-sensitive tests
 reproduce only under the pins in `requirements-dev.txt`; the optional marginal-fidelity gate
 compares the SCM against a **private** real dataset via `CLDD_DATA_DIR` and is the only thing
@@ -155,17 +230,21 @@ that needs it. Details, reproducibility, and troubleshooting:
 | [docs/configuration.md](docs/configuration.md) | every knob (`config.py`) and the one env var |
 | [docs/validation.md](docs/validation.md) | tests, gates, reproducibility, troubleshooting |
 | [docs/reject_inference.md](docs/reject_inference.md) | the four RI methods and their honest (modest) results |
-| [`FABLE.md`](FABLE.md) | **the accompanying article** — independent results & methodology assessment |
+| [`docs/assessment.md`](docs/assessment.md) | **the accompanying article** — independent results & methodology assessment (dated snapshot) |
 
 Build locally: `pip install -e ".[docs]" && sphinx-build -b html -W docs docs/_build/html`.
 
 ## Status
 
-`0.1.0` **alpha** on [PyPI](https://pypi.org/project/closed-loop-default-detection/),
-changelog in [CHANGELOG.md](CHANGELOG.md). Shipped: the loop, both worlds, all levers, the
-fidelity gate, the sklearn estimator, CI on three gates. CLDD began as a validation harness
-for the Intuit TechWeek SMB Underwriting Challenge; it is not a submission and does not
-alter challenge files.
+`0.2.0` **alpha** on [PyPI](https://pypi.org/project/closed-loop-default-detection/),
+changelog in [CHANGELOG.md](CHANGELOG.md).
+
+Shipped in 0.1.0: the loop, both synthetic worlds, all correction levers, the fidelity gate,
+the sklearn estimator, CI on three gates. Added in 0.2.0: `cldd.emp` (both EMP variants),
+priced exploration, the EMP-optimal cutoff, and the 25-seed frontier sweep.
+
+CLDD began as a validation harness for the Intuit TechWeek SMB Underwriting Challenge; it is
+not a submission and does not alter challenge files.
 
 ## Citation
 
@@ -177,7 +256,7 @@ Metadata in [`CITATION.cff`](CITATION.cff) (GitHub's "Cite this repository" read
   title   = {{closed-loop-default-detection}: measuring selective-labels default
              detection and the PD model's operating frontier},
   year    = {2026},
-  version = {0.1.0},
+  version = {0.2.0},
   license = {MIT},
   url     = {https://github.com/hossainpazooki/closed-loop-default-detection}
 }

{closed_loop_default_detection-0.1.0 → closed_loop_default_detection-0.2.0}/pyproject.toml

@@ -4,8 +4,8 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "closed-loop-default-detection"
-version = "0.1.0"
-description = "CLUE-style closed loop that measures selective-labels default detection on synthetic SMB lending cohorts and finds the PD model's operating frontier."
+version = "0.2.0"
+description = "Closed loop that measures selective-labels default detection on synthetic SMB lending cohorts and finds the PD model's operating frontier."
 readme = "README.md"
 requires-python = ">=3.10"
 license = "MIT"

{closed_loop_default_detection-0.1.0 → closed_loop_default_detection-0.2.0}/src/cldd/__init__.py

@@ -1,7 +1,7 @@
-"""Closed-loop default-rate detection — a CLUE-style harness for selective labels.
+"""Closed-loop default-rate detection — a selective-labels validation harness.
 
 generate (``synthetic``) -> measure (``eval_default``) -> improve/frontier
-(``loop``). See README.md for the CLUE mapping and how to run.
+(``loop``). See README.md for how to run.
 """
 
 from __future__ import annotations
@@ -11,7 +11,7 @@ from importlib.metadata import PackageNotFoundError, version as _pkg_version
 try:  # pyproject `version` is the single source of truth (read from installed metadata)
     __version__ = _pkg_version("closed-loop-default-detection")
 except PackageNotFoundError:  # running from a source tree with no install
-    __version__ = "0.1.0"
+    __version__ = "0.2.0"
 
 from .correctors import (
     CorrectionOutcome,
@@ -36,6 +36,13 @@ from .reject_inference import (
     ReclassificationCorrector,
     RejectInferenceCorrector,
 )
+from .emp import (
+    EMPCParams,
+    EMPResult,
+    LoanEconomics,
+    emp_harness,
+    empc_literature,
+)
 from .eval_default import PdDetectionResult, fit_observed_model, score_pd_detection
 from .model_pd import CalibratedPDClassifier, CalibratedPDModel
 from .feedback import FeedbackLoop, FeedbackResult, GenerationResult
@@ -87,6 +94,11 @@ __all__ = [
     "PdDetectionResult",
     "fit_observed_model",
     "score_pd_detection",
+    "EMPCParams",
+    "EMPResult",
+    "LoanEconomics",
+    "empc_literature",
+    "emp_harness",
     "CalibratedPDModel",
     "CalibratedPDClassifier",
     "CounterfactualResult",

{closed_loop_default_detection-0.1.0 → closed_loop_default_detection-0.2.0}/src/cldd/config.py

@@ -24,8 +24,7 @@ RANDOM_SEED = 42
 
 # Disjoint train-cohort offset for the retrain lever. Large enough that train
 # seeds (RANDOM_SEED + TRAIN_SEED_OFFSET + iteration) never collide with measure
-# seeds (RANDOM_SEED + iteration). This is the no-leakage discipline ported from
-# upstream-label-correction/clue/loop.py.
+# seeds (RANDOM_SEED + iteration).
 TRAIN_SEED_OFFSET = 1000
 
 # --------------------------------------------------------------------------- #
@@ -59,6 +58,19 @@ TARGET_DECLINED_ECE = 0.10
 
 DEFAULT_N_APPLICANTS = 4000
 
+# --------------------------------------------------------------------------- #
+# EMP reporting layer (cldd.emp) — v2 measurement axis, never loop control
+# --------------------------------------------------------------------------- #
+# Literature EMPC prior. Verified against the primary source before entry
+# (2026-07-13 spike): Verbraken, Bravo, Weber & Baesens (2014), EJOR
+# 238(2):505-513 — h(lambda) = p0*delta(0) + p1*delta(1) + (1-p0-p1)*U(0,1),
+# constant ROI, computed over the ROC convex hull (paper Eqs. 13/15). The
+# defaults are the benchmark-standard CRAN EMP::empCreditScoring values, so
+# empc columns are comparable to the published literature.
+EMPC_P0 = 0.55       # P(lambda = 0): full recovery
+EMPC_P1 = 0.10       # P(lambda = 1): total loss
+EMPC_ROI = 0.2644    # per-loan return on investment, fraction of principal
+
 # --------------------------------------------------------------------------- #
 # Observable positivity-diagnostic thresholds (cldd.diagnostics)
 # --------------------------------------------------------------------------- #

{closed_loop_default_detection-0.1.0 → closed_loop_default_detection-0.2.0}/src/cldd/correctors.py

@@ -11,9 +11,9 @@ used to live as ``_measure_*`` methods on ``SelectiveLabelsLoop`` — identical
 numpy calls, RNG seeds, sklearn calls and ordering — so the loop's numeric
 output is unchanged.
 
-This module imports only ``config``, ``eval_default``, ``model_pd`` and numpy;
-it must NOT import ``loop`` (loop re-exports ``LeverMetrics`` from here, so the
-reverse dependency would be circular).
+This module imports only ``config``, ``emp``, ``eval_default``, ``model_pd`` and
+numpy; it must NOT import ``loop`` (loop re-exports ``LeverMetrics`` from here,
+so the reverse dependency would be circular).
 """
 
 from __future__ import annotations
@@ -24,12 +24,16 @@ from dataclasses import dataclass, field
 
 import numpy as np
 
-from . import config, eval_default, model_pd
+from . import config, emp, eval_default, model_pd
 
 
 @dataclass
 class LeverMetrics:
-    """Declined-subpopulation focus + the all-population calibration for one lever."""
+    """Declined-subpopulation focus + the all-population calibration for one lever.
+
+    EMP fields (v2 reporting axis) mirror the existing declined_*/all_ece
+    pattern and are ``None`` where inapplicable (see ``SubgroupMetrics``).
+    """
 
     declined_ece: float
     declined_auc: float
@@ -38,6 +42,13 @@ class LeverMetrics:
     declined_base_rate: float
     declined_f1: float
     all_ece: float
+    declined_empc: float | None = None
+    declined_empc_fraction: float | None = None
+    declined_emp_h: float | None = None
+    declined_emp_h_fraction: float | None = None
+    declined_f1_emp_cutoff: float | None = None
+    all_empc: float | None = None
+    all_emp_h: float | None = None
 
     @classmethod
     def from_scored(cls, scored: dict) -> "LeverMetrics":
@@ -51,6 +62,13 @@ class LeverMetrics:
             declined_base_rate=d.base_rate,
             declined_f1=d.f1,
             all_ece=a.ece,
+            declined_empc=d.empc,
+            declined_empc_fraction=d.empc_fraction,
+            declined_emp_h=d.emp_h,
+            declined_emp_h_fraction=d.emp_h_fraction,
+            declined_f1_emp_cutoff=d.f1_emp_cutoff,
+            all_empc=a.empc,
+            all_emp_h=a.emp_h,
         )
 
 
@@ -76,7 +94,8 @@ class CorrectionOutcome:
     """A corrector's metrics plus its lever-specific extras.
 
     ``info`` is ``{}`` (naive/reweight), ``{"train_seed": int}`` (retrain), or
-    ``{"n_explored": int, "explored_defaults": int}`` (explore).
+    ``{"n_explored": int, "explored_defaults": int, "exploration_cost": float}``
+    (explore).
     """
 
     metrics: LeverMetrics
@@ -160,7 +179,13 @@ class ExplorationCorrector(Corrector):
 
         X = cohort["features"].to_numpy(dtype=float)
         y = cohort["true_default"]
-        weights = np.where(explored, 1.0 / ctx.exploration_rate, 1.0)
+        # ctx.exploration_rate == 0.0 guarantees explored is all-False (the
+        # comparison above never fires), so the inverse-propensity value here
+        # is never actually selected by np.where -- but it is still evaluated
+        # eagerly, so guard the division to avoid a ZeroDivisionError on that
+        # placeholder. No effect on any exploration_rate > 0 run.
+        inv_propensity = (1.0 / ctx.exploration_rate) if ctx.exploration_rate > 0.0 else 1.0
+        weights = np.where(explored, inv_propensity, 1.0)
         model = model_pd.train_pd_model(
             X[funded], y[funded], sample_weight=weights[funded], random_state=ctx.seed
         )
@@ -168,5 +193,40 @@ class ExplorationCorrector(Corrector):
         metrics = LeverMetrics.from_scored(scored)
         return CorrectionOutcome(
             metrics=metrics,
-            info={"n_explored": int(explored.sum()), "explored_defaults": int(y[explored].sum())},
+            info={
+                "n_explored": int(explored.sum()),
+                "explored_defaults": int(y[explored].sum()),
+                "exploration_cost": float(_exploration_cost(cohort, y, explored)),
+            },
+        )
+
+
+def _exploration_cost(cohort: dict, y: np.ndarray, explored: np.ndarray) -> float:
+    """Net dollar cost of the exploration budget: -(sum of realized profits on
+    the EXPLORED rows), priced by ``cldd.emp`` economics (spec sec. 4).
+
+    SCM cohorts (``'days_to_default'`` present) price each explored row by its
+    own planted default timing via :func:`cldd.emp.realized_profits` (which
+    also carries the decision-5 day-90 mean-body-lambda imputation). Flat
+    cohorts plant no timing, so explored rows are priced by the stated
+    conservative flat convention: defaulters at lambda = 1 (profit = -A),
+    goods at the full-term collections (profit = A*(f + r_term)) — computed
+    inline here since ``realized_profits`` has no ``default_day=None`` path.
+
+    Positive = the exploration budget cost money net. 0.0 when nothing was
+    explored (in particular, whenever exploration is off).
+    """
+    if not explored.any():
+        return 0.0
+    requested_amount = cohort["features"]["requested_amount"].to_numpy(dtype=float)
+    if "days_to_default" in cohort:
+        profits = emp.realized_profits(y, cohort["days_to_default"], requested_amount)
+    else:
+        economics = emp.LoanEconomics()
+        is_default = y.astype(bool)
+        profits = np.where(
+            is_default,
+            -requested_amount,
+            requested_amount * (economics.origination_fee_rate + economics.r_term),
         )
+    return -float(profits[explored].sum())

{closed_loop_default_detection-0.1.0 → closed_loop_default_detection-0.2.0}/src/cldd/diagnostics.py

@@ -1,6 +1,6 @@
 """Deployable positivity/overlap diagnostics — the observable frontier signal.
 
-Lesson from the hackathon harness (FABLE.md): the operating frontier was measured
+Lesson from the accompanying article (docs/assessment.md): the operating frontier was measured
 against *planted* truth (declined-cohort ECE), which a real lender can never
 compute — so the frontier, as measured, is not deployable knowledge. What a
 lender CAN compute is how pathological the selection itself looks: