Round 3 publication checkpoint for the revised analytical outputs, generated artifacts, and claims consuming them.
Round 3 publication review: pass with documented conditions
Round 3 confirmed the Round 2 remediation is intact, found and fixed a reproducibility blocker (non-deterministic breadth bins and threaded-BLAS model jitter that broke artifact hashing), added interval and state-count display to correlation extremes, and recorded the geography-conflict evidence in the denominator audit. Rebuilds are now byte-identical. Remaining limits are documented exploratory conditions: text-derived service mappings, four-state small cells, descriptive overlap nulls, and post-selection proxy premiums.
03 / 03
Review rounds completed
3
Independent reviewer lenses
Closed
Reproducibility blocker
Pass
Publication disposition
Review contract
Round 3 verified the Round 2 remediation under the publication lens: artifact reproducibility, residual limitation presentation, and final claim-language alignment. Deployment and server behavior remain out of scope. The statistical-methods, data-integrity, and interpretation lenses all reach the same disposition: pass with documented conditions.
Statistical methodsEstimands, formulas, uncertainty, validation, leakage, and aggregation.
Data integritySuppression, denominators, joins, mappings, periods, artifacts, and reproducibility.
InterpretationClaim strength, prior-work comparison, causal language, stale labels, and limitations.
Round 3 findings
Reproducibility blocker fixed: the breadth-matched null iterated a Python set, so quintile bins and the expected overlap drifted between runs (observed values 806.40, 805.53, 804.45 across rebuilds). Provider iteration is now sorted, and two consecutive full builds produce byte-identical artifact hashes.
Model-output determinism fixed: threaded BLAS and joblib produced last-ULP jitter in fitted coefficients and RF fold scores that broke audit-file hashes. BLAS threads are pinned at package import and the random forest trains single-threaded; linear and RF outputs are now bit-stable across processes.
Correlation extremes now display the 95% Fisher-z interval and the number of states used, satisfying the small-cell publication condition without changing the analysis.
The geography-conflict claim is now evidence-backed: denominator_audit.json records exactly 1 of 2,833 shared providers with conflicting geography, confined to urban-rural, with inpatient precedence documented.
Ten contract tests pass, and the deterministic rebuild means the committed audit hashes are reproducible by a reviewer rerunning the pipeline.
Residual publication conditions
The Round 3 disposition is conditional on keeping these limits visible; none requires further analysis for publication.
Condition · 01
Exploratory mappings remain exploratory
No official CMS crosswalk is available locally; seven known-invalid DRG codes remain quarantined from predictor modeling and the validation artifact records the quarantine policy. Service-family coefficients must stay labeled exploratory.
Small state cells remain unstable despite intervals
Nine inpatient and one outpatient service still use the four-state minimum. The extremes tables now show the Fisher-z interval and state count with every extreme, and the scatter hover carries the state count; extremes remain descriptive rankings.
Overlap nulls remain descriptive, not selection-free
The 1,134-provider overlap compares with the 755.0028 shared-universe independence baseline and the deterministic 804.4485 breadth-matched baseline (observed-to-expected 1.4097). Provider-level flag overlap cannot be tested with within-provider service permutations because a provider's flagged status is invariant under them; the reported nulls are the correct descriptive benchmarks and must not be read as institutional behavior.
The charge-to-payment proxy reuses submitted charge as its numerator after the cohort is selected on submitted charge. Proxy premiums remain denominator-specific and descriptive; the artifact carries an explicit proxy-selection note.
The project-defined high-charge cohort remains separated from the OIG cohort; the ratio remains distinct from CCR, actual cost, and markup.
FY2023 inpatient and CY2023 outpatient periods, full versus cost-observed outpatient denominators, suppression, Maryland, and non-equivalent historical comparisons remain disclosed.
Primary inpatient maps and correlations exclude Maryland; maryland_sensitivity.json records the included comparison.
Explicit categorical references, provider-disjoint GroupKFold splits, held-out fold importance, provider-preserving permutations, and exact payment-field metadata remain in place.
The build continues to execute the ten contract tests and records source, code, test, dependency, artifact, and test provenance in audit_manifest.json.
Validated strengths
Raw physical row counts and source hashes reconcile to the current manifest.
Provider-service keys, suppression handling, core denominators, and headline flag/proxy counts reconcile to generated outputs.
Full provider union is 3,199, with 3,126 full outpatient-presence providers and 2,993 cost-observed providers tracked separately.
DRGs and APCs remain separate service systems; provider joins use CCN; mapping quarantine covers 1,371 inpatient rows.
OIG historical context stays distinguished from the project's charge-based cohort, CCR proxy, and flag-pair concentration.
Rebuilds are byte-identical, so audit hashes and contract-test results are reproducible from a clean checkout.
Next review rounds
01
Analytical auditCompleted · baseline findings recorded
02
Revision review and remediationCompleted · substantial findings addressed
03
Publication reviewCompleted · pass with documented conditions
Future work may add an official CMS crosswalk, a higher-minimum-state sensitivity, or a provider-level geography-reversal sensitivity, but none blocks the current descriptive publication.