Results and Discussion

Generated by just snapshot from artifacts/full_with_peer at 2026-08-06T13:30:42+00:00.

Connection to Paper Plan

This page is the artifact-backed Results and Discussion companion to docs/paper_plan.md. The paper plan defines the research question, Materials and Methods, metric choices, and expected experiments; this snapshot reports the realized outcomes from those experiments using the current study artifacts.

The structure below follows the planned experiment sequence: retrospective benchmark stability, concept drift, opacity, public-cascade construction, public-cascade prediction, and benchmark-public construct overlap. Each package table and figure is rendered once in the experiment that owns it.

When using this page for manuscript prose, read it as an interpretation guide rather than a model leaderboard. Headline claims should describe filing-origin measurement, prevalence-aware ranking, and construct overlap within the stated bridge tier. Alternative windows and exploratory maxima remain diagnostics; the sole headline public specification is all + expanding.

Results Overview

• Research question. Can filing-origin SEC/PCAOB information predict whether an issuer later records submission-dated correspondence or public correction outcomes, and how does this review-and-correction construct relate to, but differ from, the detected-misstatement benchmark?

• Data. The workflow combines the gvkey x data_year detected-misstatement benchmark, the public SEC/PCAOB lake, the gold issuer_origin_panel and filing_origin_panel, and a raw-only gvkey-CIK-year bridge for cross-source comparison.

• Models. The core public cascade uses XGBoost over metadata, XBRL, auditor, Prior-filing history (legacy artifact key: oversight), visibility/history, and all-feature sets. Note/disclosure-breadth variables enter all without a standalone ablation. Peer-compatible Dechow, Perols, Bao, and Bertomeu-style suites are included when the peer-enabled study directory is present.

• Metrics. The common metric vocabulary is PR-AUC relative to prevalence, ROC-AUC, Brier, Brier Skill Score, ECE, top-k precision, top-decile lift, and Bao-style top-fraction precision, sensitivity, specificity, BAC, and NDCG.

• Bridge boundary. Construct overlap is wrds_validated using the confirmed WRDS SEC Analytics Suite CIK-GVKEY bridge; shared non-reliance mechanisms make Item 4.02 alignment non-independent.

• Bounded contribution. The study develops a filing-origin, submission-dated measure of review-and-correction outcomes and evaluates prevalence-aware ranking. It does not identify hidden misconduct, causal regulatory effects, or same-estimand superiority over detected-misstatement studies.

Results for Experiment 1: Retrospective Reference Benchmark

This experiment evaluates a retrospective detected-misstatement reference label under annual out-of-time folds. The supplied panel does not contain validated target-linked detection dates, so it cannot support a label-maturation or detection-lag experiment.

table_05_benchmark_timing_metrics

• Claim. The retrospective detected-misstatement reference label is rankable under annual out-of-time folds.
• Evidence. The final reference label is evaluated across expanding and rolling historical windows; Table 5 reports the strongest descriptive window.
• Boundary. The supplied panel lacks target-linked detection dates, so this is not label-maturation or hidden-misconduct evidence.

Table: Retrospective detected-misstatement reference benchmark

Reference label	Best window	Folds	PR-AUC	PR-AUC dispersion	ROC-AUC	Top-100 precision
Retrospective final label	Rolling 5y	14	0.0702	[0.0465, 0.0939]	0.7261	0.0829

Note: PR-AUC dispersion entries are descriptive fold-dispersion intervals over annual out-of-time test folds after excluding sparse folds with fewer than 10 positives. Rolling and expanding training windows overlap, so the intervals describe evaluation-period dispersion rather than independent sampling uncertainty, superpopulation confidence intervals, or causal inference uncertainty. The reference outcome is a retrospective material non-reliance-restatement firm-year label. The supplied panel does not contain validated target-linked detection dates; current-or-prior res_an* announcement-history fields are excluded from predictors and are not used to reconstruct them. Best-window entries are descriptive stability checks, not headline model-selection claims.

table_06_detected_misstatement_peer_metrics

• Claim. Peer-compatible model families provide benchmark-side metric-language alignment.
• Evidence. Detected-misstatement peer-family PR-AUC and ROC-AUC are reported across valid folds.
• Boundary. These are model-family transfer checks, not exact replications of prior samples or private data settings.

Table: Detected-misstatement peer-compatible model-family metrics

Model	Folds	Valid folds	PR-AUC	PR-AUC dispersion	ROC-AUC
Bertomeu-style XGBoost	14	14	0.0714	[0.0566, 0.0862]	0.7425
Perols linear SVM	14	14	0.0378	[0.0316, 0.0439]	0.6661
Perols Stacking	14	14	0.0374	[0.0317, 0.0432]	0.6633
Perols Bagged	14	14	0.0368	[0.0331, 0.0406]	0.6748
Perols Logit	14	14	0.0367	[0.0300, 0.0433]	0.6577
Perols MLP	14	14	0.0357	[0.0288, 0.0426]	0.6450
Bao-inspired tree ensemble	14	14	0.0334	[0.0281, 0.0387]	0.6697
Dechow variable logit	14	14	0.0247	[0.0170, 0.0325]	0.5310
Perols Entropy Tree	14	14	0.0247	[0.0211, 0.0283]	0.6131

Note: PR-AUC dispersion entries are descriptive fold-dispersion intervals over annual out-of-time test folds after excluding sparse folds with fewer than 10 positives. Rolling and expanding training windows overlap, so the intervals describe evaluation-period dispersion rather than independent sampling uncertainty, superpopulation confidence intervals, or causal inference uncertainty. Peer-compatible families are ranking checks under transferred model vocabularies, not calibrated probability comparisons or original-paper replications. Peer-model folds and public-cascade folds can cover different historical sequences, so dispersion widths should not be compared across evidence layers.

Detected-misstatement peer-family PR-AUC

• Claim. Detected-misstatement peer-compatible model families provide benchmark-side metric-language context.
• Evidence. The figure reports Dechow-, Perols-, Bao-, and Bertomeu-style families on the detected-misstatement benchmark task.
• Boundary. These rows are transferred model-family diagnostics, not original-paper numeric replications.

Figure note. Colored bars or dots encode mean PR-AUC; grey points encode valid annual test folds; capped black lines encode descriptive fold-dispersion intervals.

Detected-Misstatement Peer Fit and Skip Status

Status	Reason	Rows
fit	fit	504
skipped	missing_required_mapping	56

Detected-Misstatement Benchmark Panel

Field	Value
Rows	82,908
Firms	9,156
Years	2001-2019
Positive rate	0.0297
Reference outcome	retrospective material non-reliance-restatement firm-year
Validated target-linked detection timing	not available in the supplied panel

Full Window Summary

Label mode	Window	PR-AUC	ROC-AUC	Brier Skill Score	ECE	Top-100 precision	Top-decile precision
naive	rolling_5y	0.0702	0.7261	-4.1971	0.2048	0.0829	0.0529
naive	rolling_7y	0.0652	0.7322	-5.1230	0.2341	0.0779	0.0520
naive	rolling_10y	0.0588	0.7348	-5.9424	0.2563	0.0771	0.0514
naive	expanding	0.0523	0.7357	-5.4695	0.2503	0.0686	0.0527

Interpretation

These outcomes describe ranking and window stability for the final reference label. They provide historical model-family context, not point-in-time evidence about when a particular restatement became observable and not an estimate of hidden misconduct.

Results for Experiment 2: Concept Drift and Model Shelf-Life

This experiment compares rolling and expanding windows and checks whether feature-family importance shifts around candidate regime breaks.

Strongest Structural-Break Diagnostics

Window	Label mode	Feature family	Break year	F-stat	p-value
expanding	naive	accounting	2005	3985.9367	<0.001
expanding	naive	market	2005	3810.7723	<0.001
expanding	naive	audit	2005	3337.9848	<0.001
rolling_10y	naive	market	2005	2417.0742	<0.001
rolling_5y	naive	market	2005	1944.2735	<0.001
expanding	naive	governance	2005	1882.7072	<0.001
rolling_10y	naive	industry	2005	1844.1642	<0.001
rolling_10y	naive	governance	2005	1818.0057	<0.001
expanding	naive	industry	2005	1676.8575	<0.001
rolling_5y	naive	audit	2005	1432.5967	<0.001
rolling_10y	naive	audit	2005	1108.2615	<0.001
rolling_10y	naive	accounting	2005	1087.2621	<0.001

Mean Feature-Family Importance

Label mode	Feature family	Mean importance share
naive	accounting	0.2931
naive	industry	0.1782
naive	missingness	0.1722
naive	audit	0.1678
naive	market	0.1266
naive	governance	0.0622

Interpretation

These tables translate the paper-plan shelf-life question into realized window-level and feature-family evidence. Large window differences or breakpoint rows should be read as model-maintenance evidence rather than causal regime-shift proof.

Results for Experiment 3: Opacity and Public Review/Correction Risk

The opacity analysis reports adjusted associations from DML partially linear regressions. These estimates are not causal effects.

table_12_public_opacity_dml

• Claim. Opacity/missingness has at most diagnostic adjusted-association evidence for the current outcomes.
• Evidence. DML-style coefficients, issuer-clustered standard errors, intervals, and p-values are reported by outcome.
• Boundary. These are adjusted associations and do not identify causal selection or strategic silence.

Table: Public opacity DML-style adjusted associations

Outcome	Status	N_Obs	Prevalence	Raw_Controls	Encoded_Controls	Opacity_Components	Issuer_Clusters	Coef	Clustered_SE	Clustered_CI_95	P_Value
comment_thread	fit	96,734	0.2689	57	60	17	16026	-0.0328	0.1048	[-0.2382, 0.1725]	0.7539
amendment	fit	96,734	0.1853	57	60	17	16026	-0.1985	0.0933	[-0.3814, -0.0155]	0.0335
8k_402	fit	96,734	0.0215	57	60	17	16026	-0.0199	0.0332	[-0.0849, 0.0451]	0.5485

Note: Raw controls are source variables before encoding; encoded controls are nuisance-model columns reported at the maximum fold-local width after training-fold categorical expansion and imputation; opacity components form the missingness-density treatment. Cross-fitting keeps all rows for an issuer in the same nuisance fold. Intervals and p-values use issuer-clustered OLS covariance with a finite-cluster correction after residualization. The estimates are adjusted associations, not identified structural effects.

Interpretation

The DML rows report adjusted association between filing-origin opacity and later public review/correction outcomes. They distinguish source-availability and missingness diagnostics from silent-imputation claims. Opacity DML is an adjusted-association diagnostic only when at least one required outcome is fitted. Required-outcome statuses are comment_thread=fit, amendment=fit, 8k_402=fit; fitted outcomes are comment_thread, amendment, 8k_402; aggregate maturity is diagnostic. With at least one fitted required outcome, opacity DML is an adjusted-association diagnostic only.

Results for Experiment 4: Public Cascade Construction

This experiment validates whether public SEC/PCAOB data can support the filing-origin review-and-correction measurement surface.

table_02_public_lake_scale

• Claim. The public SEC/PCAOB lake supports the filing-origin measurement surface at scale.
• Evidence. Silver and gold row counts show filing, issuer, XBRL, notes, comment-thread, correction, and annual origin coverage.
• Boundary. Scale and coverage establish feasibility, not causal interpretation or complete regulatory review coverage.

Table: Public data architecture and analytical panel scale

Layer	Artifact	Artifact_Rows	Description
Normalized	filing_dim	21,832,838	Public filing index
Normalized	issuer_dim	970,155	Issuer dimension
Normalized	xbrl_core_fact	21,839,851	Controlled XBRL core facts
Normalized	xbrl_fact_summary	421,160	Accession-level XBRL coverage
Normalized	note_summary	597,598	Notes summary mode
Normalized	comment_thread	125,493	Eventually disclosed SEC correspondence archive
Normalized	correction_event	90,135	Amendment/correction signal
Analytical	issuer_origin_panel	205,651	Annual issuer-year modeling table
Analytical	filing_origin_panel	21,832,838	Filing-origin provenance table

table_18_public_sample_attrition

• Claim. The public sample has an explicit sequential construction path and task-specific eligibility branches.
• Evidence. Generated Table 18 reports source, year, domestic-status, observability, and task-eligibility row counts with parent-relative attrition.
• Boundary. Task rows branch from the shared observable horizon and must not be read as sequential losses from one another.

Table: Public sample attrition and task eligibility

Scope	Stage	Task	Rows	Dropped_From_Parent
sequential	source_issuer_origin	all	205651	0
sequential	fiscal_year_2011_2024	all	97028	108623
sequential	10-K/10-K/A with no observed same-year FPI-form proxy	all	96828	200
sequential	observable_365_day_horizon	all	96734	94
task	eligible_comment_thread	comment_thread	96734	0
task	eligible_amendment	amendment	96734	0
task	eligible_8k_402	8k_402	96734	0

Note: Sequential rows compare with the preceding sample-construction stage. Task rows branch from observable_365_day_horizon and therefore compare with that common observable parent rather than with one another. The 10-K/10-K/A same-year FPI-form proxy validates neither FPI status, domicile, nor US GAAP.

Public Cascade Readiness

Field	Value
Main sample rows	96,734
Fiscal-year span	2011-2024
Sample proxy	10-K/10-K/A with no observed same-year FPI-form proxy; validates neither FPI status, domicile, nor US GAAP
Task positive counts	{"8k_402": 2076, "amendment": 17927, "comment_thread": 26015}
Task exclusion counts	{"8k_402": 0, "amendment": 0, "comment_thread": 0}
Zero-positive tasks	none
Task status counts	{"fit": 576}
Readiness level	xbrl_ratio_baseline

Public Cascade Fit and Skip Status

Status	Reason	Rows
fit		576

Interpretation

Construction results document whether the public lake has enough coverage, nonzero positive labels, and feature-family availability to support the planned public-cascade experiments.

Results for Experiment 5: Public Cascade Prediction

This experiment estimates the filing-origin, submission-dated review-and-correction outcome state and compares the revision-frozen all + expanding headline with feature-family, window, calibration, selection, and peer-transfer sensitivity evidence.

table_03_public_task_metrics

• Claim. The revision-frozen primary specification produces above-prevalence ranking evidence for the three review-and-correction outcomes.
• Evidence. Generated Table 3 owns the all + expanding primary rows and reports PR-AUC, ROC-AUC, fold support, calibration diagnostics, and prevalence by task.
• Boundary. Weak Brier Skill Score and ECE keep the claim to ranking/prioritization rather than calibrated probability rules.

Table: Public cascade task metrics

Task	Panel_Positives	n_folds	valid_folds	Mean_Prevalence	Mean_PR_AUC	PR_AUC_Dispersion	Mean_ROC_AUC	Mean_Brier	Mean_Brier_Skill	Mean_ECE
comment_thread	14,327	8	8	0.2610	0.4276	[0.3847, 0.4706]	0.6973	0.2088	-0.0882	0.1793
amendment	8,415	8	8	0.1540	0.3385	[0.3060, 0.3711]	0.7183	0.1630	-0.2646	0.1993
8k_402	1,219	8	8	0.0219	0.0603	[0.0271, 0.0934]	0.7532	0.1007	-5.5151	0.2192

Note: PR-AUC dispersion entries are descriptive fold-dispersion intervals over annual out-of-time test folds after excluding sparse folds with fewer than 10 positives. Rolling and expanding training windows overlap, so the intervals describe evaluation-period dispersion rather than independent sampling uncertainty, superpopulation confidence intervals, or causal inference uncertainty. Table 3 and Figure 1 report only the revision-frozen all + expanding primary specification. Public-side holdouts are complete origin-calendar years. Training rows come only from origin years whose 365-day outcomes are fully mature before the holdout year begins. The evaluation unit is a public-cascade task summarized over these holdouts at the issuer-CIK filing-origin grain. Panel positives sum exact positive_test support over the one-to-one fit-owner rows, while mean prevalence is averaged over the reported task-window-feature evaluations; it should not be read as positives divided by a single manuscript-wide denominator. Brier Skill Score is measured relative to the corresponding prevalence-only Brier baseline. ECE is a 10-bin uniform-width calibration diagnostic from raw probability scores. Weak calibration should be read as evidence against using the scores as calibrated decision rules, not against the paper's ranking estimand.

table_04_feature_family_metrics

• Claim. Feature-family comparisons provide sensitivity evidence around the frozen primary public specification.
• Evidence. Feature counts and PR-AUC dispersion are shown across the configured public feature-family grid.
• Boundary. Feature-family summaries are aggregation evidence and should not be read as causal source dominance.

Table: Public cascade feature-family metrics

Feature_Set	Features	XBRL_Ratios	XBRL_Coverage	Best_Window	n_folds	valid_folds	Mean_PR_AUC	PR_AUC_Dispersion	Mean_ROC_AUC
All	74	11	15	expanding	8	8	0.2742	[0.2585, 0.2899]	0.7197
Metadata	23	0	0	expanding	8	8	0.2551	[0.2399, 0.2702]	0.6885
Visibility history	21	0	0	rolling_5y	8	8	0.2513	[0.2358, 0.2667]	0.6810
XBRL	42	11	15	rolling_5y	8	8	0.2195	[0.2004, 0.2386]	0.6483
Prior-filing history	1	0	0	rolling_7y	8	8	0.1589	[0.1509, 0.1670]	0.5307
Auditor	6	0	0	rolling_5y	8	8	0.1513	[0.1449, 0.1578]	0.5063

Note: PR-AUC dispersion entries are descriptive fold-dispersion intervals over annual out-of-time test folds after excluding sparse folds with fewer than 10 positives. Rolling and expanding training windows overlap, so the intervals describe evaluation-period dispersion rather than independent sampling uncertainty, superpopulation confidence intervals, or causal inference uncertainty. Entries are feature-family summaries over public-cascade task-window evaluations, not issuer-year sample sizes. Task-specific base-rate context is reported in the task tables. Note/disclosure-breadth variables enter the all-feature information set but are not reported as a standalone family row. Best-window entries are descriptive configuration summaries, not headline model-selection claims.

table_07_public_peer_metrics

• Claim. Peer-compatible families also rank review-and-correction outcomes under the submission-dated estimand.
• Evidence. Peer-family PR-AUC and ROC-AUC are reported under the same outcome design.
• Boundary. These values are within-outcome diagnostics, not cross-estimand superiority claims.

Table: Review-and-correction peer-compatible model-family metrics

Model	n_folds	valid_folds	Mean_PR_AUC	PR_AUC_Dispersion	Mean_ROC_AUC
bao_inspired_tree_ensemble	8	8	0.2163	[0.2053, 0.2273]	0.6319
bertomeu_style_xgb	8	8	0.2162	[0.2050, 0.2274]	0.6317
perols_bagged	8	8	0.2061	[0.1947, 0.2175]	0.6200
perols_stacking	8	8	0.1985	[0.1857, 0.2114]	0.6034
perols_mlp	8	8	0.1976	[0.1821, 0.2130]	0.6043
perols_entropy_tree	8	8	0.1932	[0.1833, 0.2031]	0.6016
perols_linear_svm	8	8	0.1928	[0.1776, 0.2080]	0.6024
perols_logit	8	8	0.1915	[0.1782, 0.2049]	0.5979
dechow_public_xbrl_proxy_logit	8	8	0.1622	[0.1523, 0.1721]	0.5169

Note: PR-AUC dispersion entries are descriptive fold-dispersion intervals over annual out-of-time test folds after excluding sparse folds with fewer than 10 positives. Rolling and expanding training windows overlap, so the intervals describe evaluation-period dispersion rather than independent sampling uncertainty, superpopulation confidence intervals, or causal inference uncertainty. Peer-compatible families are ranking checks under transferred model vocabularies, not calibrated probability comparisons or original-paper replications. Peer-model folds and public-cascade folds can cover different historical sequences, so dispersion widths should not be compared across evidence layers.

table_13_public_fold_support

• Claim. Annual outcome test folds have sufficient positive support for reported dispersion summaries.
• Evidence. Task-year rows, positives, prevalence, and sparse-fold flags are reported.
• Boundary. Fold support makes dispersion auditable; it does not remove class-imbalance or calibration concerns.

Table: Origin-calendar-year outcome fold support

Task	Test_Year	Configs	Test_Rows	Positives	Prevalence	Sparse_Excluded
8k_402	2017	24	6,502	96	0.0148	No
8k_402	2018	24	6,651	86	0.0129	No
8k_402	2019	24	6,615	53	0.0080	No
8k_402	2020	24	6,531	70	0.0107	No
8k_402	2021	24	7,018	439	0.0626	No
8k_402	2022	24	7,642	154	0.0202	No
8k_402	2023	24	7,187	178	0.0248	No
8k_402	2024	24	6,741	143	0.0212	No
amendment	2017	24	6,502	1,061	0.1632	No
amendment	2018	24	6,651	1,147	0.1725	No
amendment	2019	24	6,615	1,141	0.1725	No
amendment	2020	24	6,531	1,069	0.1637	No
amendment	2021	24	7,018	1,199	0.1708	No
amendment	2022	24	7,642	950	0.1243	No
amendment	2023	24	7,187	959	0.1334	No
amendment	2024	24	6,741	889	0.1319	No
comment_thread	2017	24	6,502	1,942	0.2987	No
comment_thread	2018	24	6,651	1,647	0.2476	No
comment_thread	2019	24	6,615	1,537	0.2324	No
comment_thread	2020	24	6,531	1,691	0.2589	No
comment_thread	2021	24	7,018	1,653	0.2355	No
comment_thread	2022	24	7,642	2,091	0.2736	No
comment_thread	2023	24	7,187	1,895	0.2637	No
comment_thread	2024	24	6,741	1,871	0.2776	No

Note: Entries report complete origin-calendar-year test support collapsed across configurations because test rows and positive counts are task-year properties. Sparse folds are those with fewer than 10 positives; such folds are excluded from formal fold-dispersion intervals.

table_14_task_feature_family_metrics

• Claim. Task-level feature-family rankings provide sensitivity evidence around the frozen primary rows.
• Evidence. Task-by-feature-family PR-AUC, ROC-AUC, calibration diagnostics, and fold support are reported across the configured grid.
• Boundary. Use this table for label-specific prose rather than a single global feature-family ranking.

Table: Task-by-feature-family public-cascade metrics

Task	Feature_Set	n_folds	valid_folds	Mean_Prevalence	Mean_PR_AUC	PR_AUC_Dispersion	Mean_ROC_AUC	Mean_Brier_Skill	Mean_ECE
8k_402	All	8	8	0.0219	0.0554	[0.0311, 0.0796]	0.7458	-5.2284	0.2107
8k_402	Visibility history	8	8	0.0219	0.0500	[0.0230, 0.0770]	0.6898	-7.3510	0.2713
8k_402	Metadata	8	8	0.0219	0.0456	[0.0284, 0.0627]	0.7038	-6.4330	0.2577
8k_402	XBRL	8	8	0.0219	0.0422	[0.0256, 0.0589]	0.6537	-8.0766	0.2920
8k_402	Auditor	8	8	0.0219	0.0279	[0.0106, 0.0452]	0.5076	-15.9790	0.4741
8k_402	Prior-filing history	8	8	0.0219	0.0238	[0.0090, 0.0386]	0.5193	-13.8918	0.4318
amendment	All	8	8	0.1540	0.3383	[0.3067, 0.3699]	0.7159	-0.3199	0.2124
amendment	Metadata	8	8	0.1540	0.3341	[0.3090, 0.3593]	0.7012	-0.3208	0.2152
amendment	Visibility history	8	8	0.1540	0.3337	[0.3114, 0.3561]	0.7044	-0.4478	0.2502
amendment	XBRL	8	8	0.1540	0.2274	[0.2006, 0.2542]	0.6307	-0.6914	0.2872
amendment	Prior-filing history	8	8	0.1540	0.1726	[0.1504, 0.1948]	0.5305	-0.9316	0.3408
amendment	Auditor	8	8	0.1540	0.1641	[0.1595, 0.1687]	0.5146	-0.9756	0.3526
comment_thread	All	8	8	0.2610	0.4289	[0.3868, 0.4710]	0.6974	-0.0929	0.1813
comment_thread	XBRL	8	8	0.2610	0.3888	[0.3380, 0.4396]	0.6604	-0.2583	0.2420
comment_thread	Metadata	8	8	0.2610	0.3855	[0.3506, 0.4204]	0.6605	-0.1117	0.1749
comment_thread	Visibility history	8	8	0.2610	0.3701	[0.3327, 0.4074]	0.6489	-0.1796	0.2049
comment_thread	Prior-filing history	8	8	0.2610	0.2804	[0.2581, 0.3028]	0.5422	-0.3229	0.2456
comment_thread	Auditor	8	8	0.2610	0.2620	[0.2417, 0.2822]	0.4966	-0.3037	0.2374

Note: PR-AUC dispersion entries are descriptive fold-dispersion intervals over annual out-of-time test folds after excluding sparse folds with fewer than 10 positives. Rolling and expanding training windows overlap, so the intervals describe evaluation-period dispersion rather than independent sampling uncertainty, superpopulation confidence intervals, or causal inference uncertainty. Entries are task-by-feature-family averages over the configured public-cascade training windows. They clarify the aggregation behind the feature-family summary and should be read as information-set evidence, not causal decomposition.

table_17_selection_profile

• Claim. Recorded outcomes partly reflect selected scrutiny and issuer visibility states.
• Evidence. Outcome rates are profiled across filing size, XBRL assets, filing history, prior correspondence, form type, and FPI proxy strata.
• Boundary. This descriptive profile is not a causal SEC-selection correction.

Table: Selection-aware outcome profile

Stratum	Group	Issuer_Years	Comment_Rate	Amendment_Rate	Item_4_02_Rate
Filing size	Below median (1,855,960)	99,609	0.1117	0.2538	0.0169
Filing size	At/above median (1,855,960)	99,610	0.2951	0.1914	0.0251
XBRL log assets	Below median (19.6571)	40,281	0.2939	0.2388	0.0318
XBRL log assets	At/above median (19.6571)	40,282	0.2961	0.1192	0.0134
Prior filing count	Below median (113)	99,547	0.1117	0.2356	0.0169
Prior filing count	At/above median (113)	99,672	0.2950	0.2096	0.0251
Days since prior filing	Below median (13.0000)	97,682	0.2701	0.2164	0.0247
Days since prior filing	At/above median (13.0000)	100,642	0.1402	0.2283	0.0176
Prior correspondence-thread history	No prior 3y correspondence thread	113,956	0.1034	0.2147	0.0136
Prior correspondence-thread history	Any prior 3y correspondence thread	85,263	0.3370	0.2331	0.0309
Annual form	10-K	196,919	0.2051	0.2206	0.0211
Annual form	10-K/A	2,300	0.0587	0.3943	0.0126
Observed same-year FPI-form indicator	No observed 20-F/40-F/6-K	198,910	0.2031	0.2225	0.0210
Observed same-year FPI-form indicator	Observed 20-F/40-F/6-K	309	0.3722	0.3139	0.0324

Note: Issuer_Years are descriptive strata from the existing public issuer-origin panel. They show how outcome rates vary with filing visibility, history, and issuer profile variables. Parenthetical values in group labels are split thresholds, not sample sizes; XBRL log-asset strata are limited to observations with available XBRL asset values; days since prior filing refers to any prior EDGAR filing, not only a prior annual report. The observed same-year FPI-form indicator records whether a 20-F, 40-F, or 6-K is observed for the issuer fiscal year; it does not validate FPI status or domicile. The table is selection-aware evidence, not a causal adjustment model or proof that SEC scrutiny selection has been solved. Prior comment-thread strata use eventually disclosed correspondence submission dates; those history fields are excluded from predictive models because release-time availability at the filing origin is not verified.

Review-and-correction outcome PR-AUC

• Claim. The revision-frozen primary specification ranks later review-and-correction outcomes above task prevalence.
• Evidence. Figure 1 receives only the declared all + expanding rows also owned by generated Table 3.
• Boundary. This is ranking evidence, not calibrated deployment evidence; Brier Skill Score and ECE remain the calibration gate.

Figure note. Colored bars or dots encode mean PR-AUC; grey points encode valid annual test folds; capped black lines encode descriptive fold-dispersion intervals.

Feature-family PR-AUC

• Claim. Feature-family comparisons provide sensitivity evidence around the frozen outcome specification.
• Evidence. The figure compares all-feature, metadata, XBRL, auditor, Prior-filing history (legacy artifact key: oversight), and visibility/history information sets under the same outcome evaluation frame.
• Boundary. Interpret as information-set evidence rather than a structural source-importance or mechanism claim.

Figure note. Colored bars or dots encode mean PR-AUC; grey points encode valid annual test folds; capped black lines encode descriptive fold-dispersion intervals.

Review-and-correction peer-family PR-AUC

• Claim. Familiar accounting ML model-family vocabularies can be evaluated on the review-and-correction outcomes.
• Evidence. The figure moves the peer-compatible families to the outcome task and keeps the metric vocabulary comparable.
• Boundary. Do not compare these values as same-estimand superiority over detected-misstatement studies.

Figure note. Colored bars or dots encode mean PR-AUC; grey points encode valid annual test folds; capped black lines encode descriptive fold-dispersion intervals.

Public Peer Task Summary

Task	Metric rows	Mean prevalence	Mean PR-AUC	Mean ROC-AUC
comment_thread	1,248	0.2604	0.3248	0.5905
amendment	1,248	0.1533	0.2331	0.6105
8k_402	1,248	0.0223	0.0422	0.6266

Public Peer Fit and Skip Status

Status	Reason	Rows
fit	fit	3,744
skipped	missing_required_mapping	480
skipped	one_class_train_or_empty_features	192

Interpretation

Prediction results should be read within each label, prevalence, feature family, training window, and model-family mapping. Public peer rows provide model-language transfer evidence, not same-estimand superiority over the detected-misstatement literature. Alternative configurations remain sensitivity evidence rather than headline selections.

Results for Experiment 6: Detected-Misstatement Benchmark and Public Cascade Overlap

This experiment is the integrated-paper gate. The current bridge is the confirmed WRDS SEC Analytics Suite CIK-GVKEY link export, used as a raw-only gvkey-CIK-year bridge.

table_08_bridge_coverage

• Claim. The bridge covers most benchmark rows and firms before overlap claims are made.
• Evidence. Row, firm, and positive-row coverage are reported for the raw-only WRDS gvkey-CIK-year bridge.
• Boundary. Construct-overlap claims remain bounded to matched bridge rows.

Table: Bridge coverage

Metric	Value
raw_rows	82,908
raw_firms	9,156
matched_raw_rows	79,273
matched_raw_firms	8,758
row_coverage_rate	0.9562
firm_coverage_rate	0.9565
raw_positive_rows	2,460
matched_positive_rows	2,337

Note: These rates describe raw CIK-GVKEY bridge availability. Construct-overlap claims use the narrower high-confidence sample reported in the bridge-overlap sample-boundaries appendix table.

table_09_construct_alignment

• Claim. Public scores and detected-misstatement scores show reciprocal severe-tail enrichment under the bridge gate.
• Evidence. Generated Table 9 is the sole owner of the two declared primary alignment rows and their top-decile lift, precision, FDR, and bootstrap intervals.
• Boundary. Lift above one supports enrichment, while low absolute precision and high FDR rule out event-identification claims.

Table: Construct-overlap ranking alignment

Direction	Model	Target	Feature_Set	Window	N	Positives	Top_10pct_K	Top_10pct_Hits	PR_AUC	Top_Decile_Lift	Top_10pct_Precision	Top_10pct_FDR	Lift_Bootstrap_Interval	Bootstrap_Issuers	Confidence_Tier
Public score to benchmark positives	public_cascade	Item 4.02	all	expanding	10,828	127	1,083	48	0.0423	3.7788	0.0443	0.9557	[2.8664, 4.7039]	3,410	high_confidence
Detected-misstatement score to recorded outcomes	benchmark_xgb	Item 4.02	benchmark_all	expanding	31,909	507	3,191	134	0.0354	2.6429	0.0420	0.9580	[2.2663, 3.0655]	5,870	high_confidence

Note: Lift bootstrap intervals are issuer-cluster percentile bootstrap intervals from the bridge-gated artifacts, not annual fold-dispersion intervals. Bridge tier is wrds_validated, and displayed rows are restricted to high-confidence bridge rows. Top-10% precision and FDR report the absolute base-rate burden behind lift; the implicit bridge-sample base rate can differ from the full public-cascade task prevalence because of the high-confidence bridge restriction. The external reference and public Item 4.02 outcome share a non-reliance mechanism but use different units, clocks, and sources; the severe-tail rows therefore show cross-source consistency rather than independent construct validation or label equivalence.

table_15_bridge_overlap_matrix

• Claim. Recorded outcomes and detected-misstatement labels are related but not identical across bridge tiers.
• Evidence. The matrix reports benchmark/outcome rates, co-occurrence, and lifts by outcome and bridge tier.
• Boundary. The typed pattern is cross-source reference alignment, not independent construct validation or label equivalence.

Table: Bridge-gated outcome overlap matrix

Public_Label	Bridge_Tier	Bridge_Rows	Benchmark_Positives	Public_Positives	Both_Positive	Benchmark_Rate	Public_Rate	Public_Rate_If_Benchmark_Pos	Benchmark_Rate_If_Public_Pos	Public_Lift_If_Benchmark_Pos	Benchmark_Lift_If_Public_Pos
comment_thread	high_confidence	44,183	1,303	13,394	407	0.0295	0.3031	0.3124	0.0304	1.0304	1.0304
comment_thread	ambiguous	2,158	65	760	21	0.0301	0.3522	0.3231	0.0276	0.9174	0.9174
comment_thread	all_matched	46,341	1,368	14,154	428	0.0295	0.3054	0.3129	0.0302	1.0243	1.0243
amendment	high_confidence	44,183	1,303	9,373	532	0.0295	0.2121	0.4083	0.0568	1.9246	1.9246
amendment	ambiguous	2,158	65	470	28	0.0301	0.2178	0.4308	0.0596	1.9779	1.9779
amendment	all_matched	46,341	1,368	9,843	560	0.0295	0.2124	0.4094	0.0569	1.9273	1.9273
8k_402	high_confidence	44,183	1,303	1,025	258	0.0295	0.0232	0.1980	0.2517	8.5351	8.5351
8k_402	ambiguous	2,158	65	47	14	0.0301	0.0218	0.2154	0.2979	9.8894	9.8894
8k_402	all_matched	46,341	1,368	1,072	272	0.0295	0.0231	0.1988	0.2537	8.5951	8.5951

Note: Bridge_Rows are bridge-gated label-overlap diagnostics. The table reports absolute outcome and benchmark rates and lift for each recorded outcome by bridge tier. These descriptive rates broaden the cross-source reference comparison beyond the sparse Item 4.02 severe-tail ranking rows; they do not imply independent construct validation or label equivalence.

table_16_bridge_sample_boundaries

• Claim. The bridge exercise has explicit covered, ambiguous, dropped, and unmatched sample boundaries.
• Evidence. Benchmark rows and positives are shown by bridge-overlap boundary.
• Boundary. Generalization beyond high-confidence mapped rows should be qualified.

Table: Bridge-overlap sample boundaries

Boundary	Benchmark_Rows	Row_Share	Benchmark_Positives	Positive_Share	Interpretation
full_raw	82,908	1.0000	2,460	1.0000	Benchmark rows entering the bridge-overlap accounting screen
ambiguous	2,158	0.0260	65	0.0264	Mapped rows retained for sensitivity diagnostics, not headline overlap
dropped	36,567	0.4411	1,092	0.4439	Rows without a usable high-confidence public-side overlap match
high_confidence	44,183	0.5329	1,303	0.5297	Rows used for headline bridge-gated construct-alignment statistics

Note: This table reports the construct-overlap sample boundary after bridge accounting, not the raw CIK-GVKEY coverage rate reported in the bridge-coverage table. Construct-overlap claims are bounded to high-confidence rows. Dropped rows define the generalizability boundary of the overlap exercise. An additional 3,635 raw benchmark rows lack a usable public-side identifier and are outside all overlap statistics.

Cross-source alignment lift

• Claim. The WRDS-validated bridge documents related-but-non-identical alignment between the recorded outcomes and detected-misstatement labels.
• Evidence. Figure 5 receives exactly the two declared primary alignment rows owned by generated Table 9, alongside precision/FDR context in the table.
• Boundary. Item 4.02 lift reflects a shared non-reliance mechanism across different units, clocks, and sources; it is not independent construct validation or event-identification proof.

Figure note. Blue bars encode top-decile lift; capped black lines encode issuer-cluster percentile-bootstrap intervals; the dashed vertical line marks lift = 1; annotations report top-decile precision and FDR.

The declared primary ranking-alignment rows are severe-tail diagnostics. Lift above one shows enrichment, while low absolute precision and high FDR keep the interpretation bounded to construct overlap rather than event identification.

Exploratory maxima (post-hoc)

These exploratory, post-hoc maxima summarize the searched raw alignment grids. They disclose search-universe size and model-selection spread but never replace the generated Table 9 primary rows.

Direction	Search-universe rows	Maximum-row keys	Maximum lift	Primary lift	max_minus_primary
public_to_benchmark	72	{"bridge_tier": "high_confidence", "feature_set": "all", "label_mode": "benchmark_naive", "model_id": "public_cascade", "score_aggregation": "mean", "task": "8k_402", "train_window": "expanding"}	3.7788	3.7788	0.0000
benchmark_to_public	720	{"bridge_tier": "high_confidence", "feature_set": "peer_compatible", "label_mode": "naive", "model_id": "bertomeu_style_xgb", "score_aggregation": "benchmark_score", "target_public_label": "label_8k_402_365", "train_window": "rolling_7y"}	3.1360	2.6429	0.4931

The contingency matrix is the broader cross-source reference comparison. Comment threads are broad submission-dated scrutiny outcomes, amendments show stronger correction/friction alignment, and Item 4.02 is a rare severe-tail state. Because the external reference and public Item 4.02 outcome share a non-reliance mechanism, the Item 4.02 lift is consistency evidence rather than independent construct validation.

Aggregation Sensitivity

Outcome label	Bridge tier	Aggregation rule	Rows	Pre-agg rate	Post-agg rate	Rate delta	Sensitive
label_comment_thread_365	ambiguous	label_max	2,158	0.1896	0.3522	0.1626	True
label_comment_thread_365	high_confidence	label_max	44,183	0.3031	0.3031	0.0000	False
label_amendment_365	ambiguous	label_max	2,158	0.1152	0.2178	0.1026	True
label_amendment_365	high_confidence	label_max	44,183	0.2121	0.2121	0.0000	False
label_8k_402_365	ambiguous	label_max	2,158	0.0110	0.0218	0.0108	True
label_8k_402_365	high_confidence	label_max	44,183	0.0232	0.0232	0.0000	False

Benchmark-Positive Outcome Co-occurrence

Pattern	Comment	Amendment	8-K 4.02	Benchmark positives	Share	Display count
none	0	0	0	513	0.3937	513
8k_402_365	0	0	1	41	0.0315	41
amendment_365	0	1	0	271	0.2080	271
amendment_365+8k_402_365	0	1	1	71	0.0545	71
comment_thread_365	1	0	0	178	0.1366	178
comment_thread_365+8k_402_365	1	0	1	39	0.0299	39
comment_thread_365+amendment_365	1	1	0	83	0.0637	83
comment_thread_365+amendment_365+8k_402_365	1	1	1	107	0.0821	107

Event-Time Concentration

Relative year	Outcome label	Benchmark pos rows	Benchmark neg rows	Rate if benchmark pos	Rate if benchmark neg	Difference	Balanced window
-3	label_comment_thread_365	610	20,518	0.1738	0.3014	-0.1277	True
-3	label_amendment_365	610	20,518	0.2902	0.2028	0.0873	True
-3	label_8k_402_365	610	20,518	0.0213	0.0199	0.0014	True
-2	label_comment_thread_365	610	20,518	0.2066	0.3076	-0.1010	True
-2	label_amendment_365	610	20,518	0.2984	0.1966	0.1018	True
-2	label_8k_402_365	610	20,518	0.0328	0.0213	0.0114	True
-1	label_comment_thread_365	610	20,518	0.2557	0.3066	-0.0508	True
-1	label_amendment_365	610	20,518	0.3033	0.1859	0.1173	True
-1	label_8k_402_365	610	20,518	0.1049	0.0182	0.0867	True
0	label_comment_thread_365	610	20,518	0.3344	0.3089	0.0256	True
0	label_amendment_365	610	20,518	0.4082	0.1788	0.2294	True
0	label_8k_402_365	610	20,518	0.2131	0.0166	0.1965	True
1	label_comment_thread_365	610	20,518	0.3770	0.3064	0.0707	True
1	label_amendment_365	610	20,518	0.3689	0.1701	0.1987	True
1	label_8k_402_365	610	20,518	0.1934	0.0173	0.1761	True
2	label_comment_thread_365	610	20,518	0.3902	0.3089	0.0813	True
2	label_amendment_365	610	20,518	0.3098	0.1663	0.1435	True
2	label_8k_402_365	610	20,518	0.1607	0.0180	0.1426	True
3	label_comment_thread_365	610	20,518	0.3721	0.3038	0.0683	True
3	label_amendment_365	610	20,518	0.2721	0.1801	0.0920	True
3	label_8k_402_365	610	20,518	0.1164	0.0170	0.0994	True

Interpretation

Overlap results determine whether the benchmark and public cascade are related enough for an integrated comparison. The evidence can support a related-but-non-identical interpretation only when bridge coverage, multiplicity, reciprocal alignment, and event-time concentration are all reported.

Discussion

The evidence supports a filing-origin measurement and ranking contribution, not a hidden-misconduct detector or calibrated deployment rule. The frozen public specification is interpreted against prevalence and calibration, while WRDS-validated overlap documents related but non-identical recorded outcomes.

Answers to the research questions

• Outcome-model results support a prevalence-aware ranking claim for three review-and-correction outcomes, but calibration diagnostics keep the interpretation to ranking and prioritization.
• Recorded outcomes and detected-misstatement benchmark labels are related but non-identical constructs.
• Public-cascade scores can rank benchmark positives in the matched overlap; detected-misstatement scores can also rank severe correction outcomes.
• Selection-profile rows show that submission-dated correspondence partly reflects the scrutiny process, not a clean issuer-risk-only label.
• wrds_validated bridge evidence supports the integrated benchmark-to-public cross-source alignment interpretation.

Comparison with prior literature

The detected-misstatement and outcome peer suites align model-family and metric language with prior accounting prediction work. They are transferred family diagnostics, not original-sample replications or same-estimand leaderboard comparisons.

Accounting and institutional interpretation

The outcomes record submission-dated correspondence, correction, and filing-friction states. The accounting contribution is therefore a filing-origin measurement and prioritization design, with machine learning serving as measurement infrastructure.

Selection and visibility

Correspondence and correction outcomes are partly selected by issuer visibility, filing history, source availability, and scrutiny. Selection-profile and opacity rows describe those boundaries without claiming a causal selection correction.

Prior-filing history means prior_filing_count, not PCAOB inspection. is_domestic_us_gaap_proxy means 10-K/10-K/A with no observed same-year FPI-form proxy and validates neither FPI status, domicile, nor US GAAP. PCAOB inspection archives are provenance inputs; inspection events are not joined to Gold, and there are no model-eligible inspection features.

Partner nonadministrative-amendment facts use the post-year-proxy uncensored public-model panel: 96734 rows evaluated, 96734 nonmissing, 0 nonzero, range [0, 0], is_constant_zero=true, and equality to Item 4.02 is false across 89181 rows.

Generalizability

Outcome-model findings generalize to the documented fiscal-year, filing-origin, and submission-date frame. Cross-source alignment findings generalize only to the covered wrds_validated bridge sample.

Limitations and future work

• The evidence supports measurement and decision-useful ranking claims, not causal proof of hidden misconduct.
• Correspondence is submission-dated and disclosed later; it does not represent the complete SEC review universe.
• Negative Brier Skill Score or large ECE remains evidence against deployment-ready probability rules.
• External temporal validation and identified selection corrections remain future work rather than current findings.

Claim ledger

The controlled categories are reportable, supporting, diagnostic, and deferred; no headline is promoted from a raw maximum.

Claim	Evidence	Category	Boundary
Filing-origin information ranks later submission-dated correspondence and public correction outcomes.	Generated Table 3, annual fold support, and Figure 1.	reportable	Ranking evidence relative to prevalence, not calibrated deployment.
Feature and training-window patterns qualify the frozen public specification.	Generated Tables 4 and 14 plus Figure 2.	supporting	Information-set evidence, not mechanism or XBRL dominance.
Recorded-outcome and detected-misstatement constructs are related but non-identical.	WRDS bridge coverage, generated Table 9, Figure 5, and contingency matrix.	supporting	Conditional on bridge tier and covered sample.
Opacity/missingness has adjusted-association evidence for the recorded outcomes.	DML adjusted-association rows.	diagnostic	Null or weak rows cannot support strategic-silence claims.
Exploratory maximum lifts describe the searched alignment grid.	Separately labeled post-hoc maximum table and search-universe counts.	diagnostic	Never substitutes for the declared primary alignment rows.
The models identify hidden misconduct or causal regulatory effects.	No current artifact identifies either estimand.	deferred	Requires a different design and evidence base.

Reproducibility and Provenance

table_01_component_status

• Claim. All paper-facing study components are available for the current artifact-backed run.
• Evidence. Component statuses are read from the peer-enabled study manifest.
• Boundary. Component completion is a reproducibility status, not by itself a substantive empirical claim.

Table: Study component status for manuscript evidence

Component	Status	Tier	Manuscript role
benchmark	complete		Detected-misstatement timing and model-family diagnostics
public_cascade	complete		Filing-origin outcome-ranking evidence
bridge_probe	crosswalk_available		CIK-GVKEY coverage and multiplicity checks
peer_comparison	complete		Benchmark model-family transfer checks
public_peer_comparison	complete		Review-and-correction model-family transfer checks
construct_overlap	complete	wrds_validated	Bridge-gated cross-source reference alignment

Run and package identity

Field	Value
Artifact generation time	2026-08-03T18:46:28+00:00
Snapshot generation time	2026-08-06T13:30:42+00:00
Snapshot mode	full
Study commit	c968c1458c088467044b45db838c7fd288781558
Git dirty	False
Package schema	manuscript-package-v2
Package study commit	c968c1458c088467044b45db838c7fd288781558
Package study-manifest hash	c649e8fb10ed44474f6b8669099005f8c475796497a21c12b580e97bd2991daa
Config hash	e4302457a2c7935e161339a0a353ca131332be7870f9e13e55b3dc8c90f12401
Input hash	00dd79e0ca18d8d2b7b7da214129c4c421eb17206827d917e99400e77623e5fb
uv.lock hash	c9a3f55546c9ca7cbe812f0b2414c6eb97f749bd5dbd934099efc7275361e0f6
Public-data as-of date	2026-07-06
Public-lake fresh build	True
Public-lake commit	c968c1458c088467044b45db838c7fd288781558
Public-lake Git dirty	False
Public-lake config hash	c6da3c3a83dd57d1e4e84d7848f25e4c0de439c5abb9ad52082b347cc9b6065b
Public-lake input hash	b7408e8696d6937a362ff4f9722f4ea1ce82ca03e598f7c3d062a0c71b6ab3ab
Public-lake uv.lock hash	c9a3f55546c9ca7cbe812f0b2414c6eb97f749bd5dbd934099efc7275361e0f6
Form AP source kind	verified_zip_member
Form AP archive hash	8358afa8393084673b80e1efc5535d21c389ec1ff7dbb500c75f681945322d12
Form AP member	FirmFilings.csv
Form AP member hash	a05223c41169d4a2e80dd8e48e7367b6e0c55f2dafad3e987ec3cef4a34de5ee
WRDS source	15 observed combinations across 4 WRDS source families in the current run; full values remain hash-bound in the study manifest
WRDS version	WRDS SEC Analytics Suite / CIK-GVKEY Link Table.csv
WRDS extraction time	2026-08-03T18:05:27+00:00
WRDS hash	1a1f214b526ea661a817f91d3988b607dd3ff77884a4d1c8d54197a3ada02798
Canonical status	CANONICAL
Canonical predicate failures	none
Public-lake final-report schema	public-lake-final-report-v1
Peer comparison mode	full
Bridge status	crosswalk_available
Cross-source alignment bridge tier	wrds_validated

Source roles and hashes

Local input paths are omitted; stable roles, availability, and content hashes identify the evidence inputs.

Source role	Available	SHA-256
Detected-misstatement benchmark input	True	468609b1879c6820618c01f051367eaea4094451087e9c51d3cd17b649bd05d2
Public issuer dimension input	True	103b4a95a443a1cec3ba2a371eb03c4ff1fd54fa9ecddc3e3412f7b6f596d695
Public issuer-origin panel input	True	51a1a713392e6f4f7ed8dafc06ae6b5860975753d1b36064157f6bbf0429a004
Public-lake final report input	True	e40b9c1cd0e432595cd7bee858bab37ba3411f7394eb26a5941db39881974663
CIK-GVKEY bridge input	True	1a1f214b526ea661a817f91d3988b607dd3ff77884a4d1c8d54197a3ada02798
Public-lake run metadata input	True	1970761dcbd4d181d195953d02adea488b1ce28d667b65c284b7f79cfff9f04b
Form AP source metadata input	True	8c899fc2dc627ddbe6d4a507e840e3efdc7f74e3e6f885970a32b69aca6c28b6

Claim maturity

Claim	Status
public_prediction	reportable
feature_and_window_sensitivity	supporting
construct_alignment	supporting
opacity_dml	diagnostic