Manuscript Audit and Polish Brief

Use this editorial brief when reviewing or polishing the reporting-risk-cascade paper in ../reporting-risk-cascade-manuscript for a target submission to The British Accounting Review (BAR).

This is an internal quality-control brief for academic review, claim discipline, and prose polish. It is not manuscript prose. Its purpose is to keep the paper's accounting contribution, evidence boundary, and terminology stable while making the argument readable to both accounting/finance and machine-learning readers.

This brief was last reconciled for BAR targeting on 2026-05-28 against the binding design contract in reporting-risk-cascade/docs/paper_plan.md and the current artifact-backed results in reporting-risk-cascade/docs/results_snapshot.md. The manuscript source lives in reporting-risk-cascade-manuscript; live submission mechanics should be checked against the current official BAR and Elsevier author instructions before upload.

Mode Invocation Rule

When executing this brief, run exactly one Review Mode per invocation unless the user explicitly asks for a combined audit. If the user asks for "full audit", execute the Full audit mode only; do not also polish sections or rewrite the Abstract unless separately requested. This keeps outputs deep enough to be useful and prevents shallow, mixed-mode responses.

This is an author-side internal quality-control brief. It is not a journal peer-review process and should not be represented as use of AI by a BAR editor or reviewer. If AI-assisted tools are used in manuscript preparation, the manuscript must include the required Elsevier-style disclosure and the authors remain responsible for all citations, claims, tables, and interpretations.

Editorial Stance

Review the manuscript as a skeptical but fair BAR referee or associate editor. The strongest version of the paper is an accounting measurement contribution about label observability in financial reporting, with machine learning used as measurement infrastructure. The paper is not a generic fraud detection article, not an algorithm leaderboard, not a network cascade paper, and not a causal enforcement study.

The likely rejection path is clear: BAR reviewers may decide that the accounting contribution is obscured by workflow detail, that selected submission-dated correspondence is oversold as reporting failure or contemporaneously public scrutiny, that sparse Item 4.02 evidence carries too much weight, or that the study is too close to a comment-letter prediction paper. Every review pass should therefore protect the construct first, then the evidence, then the style.

Do not invent empirical findings. Empirical claims must be traceable to the manuscript, the Paper Plan, the Results Snapshot, generated manuscript tables/figures, or current artifacts.

Research Object

The central construct is the filing-origin, submission-dated review-and-correction outcome state: whether information verifiably available when an annual filing becomes public can rank issuers that later enter typed recorded outcome channels. SEC correspondence is eventually disclosed but is dated here by the underlying EDGAR submission date, not by an unverified public- release date.

For continuity with older drafts, the phrase pre-disclosure reporting-risk state may appear in notes, but article prose should prefer filing-origin, submission-dated review-and-correction outcome state. This reduces terminology drift and ties the construct to the actual predictor and outcome clocks used in the design.

The paper has three evidence layers:

Benchmark layer. The detected-misstatement benchmark is at the gvkey x data_year grain. It is used for timing sensitivity, drift, missingness, and peer-compatible model-family checks. It is not the definition of the public reporting-risk estimand.
Public cascade layer. The public-cascade panel is at the issuer_cik x fiscal_year x origin_date grain. The origin date is the selected annual filing date. The headline public tasks are comment_thread, amendment, and 8k_402.
Bridge gate. The gvkey-CIK-year bridge is used only for construct-overlap validation. Bridge status must be read from the current study manifest, construct-overlap summary, and bridge/overlap artifacts before any integrated benchmark-to-public overlap claim is allowed. Current artifacts may report wrds_validated, but that is a checked artifact state, not a permanent prompt fact. If the bridge tier is stale, missing, or candidate-only, integrated overlap claims must be downgraded. The standalone public-cascade measurement claim does not require bridge validation.

The public-cascade labels are typed recorded outcome states:

Artifact label	Stable manuscript term	Use when	Claim boundary
`comment_thread_365`	submission-dated, eventually disclosed SEC correspondence	Describing the correspondence outcome.	Dated by the underlying EDGAR submission, not public release; selected and not the full SEC review universe.
`amendment_365`	public amendment/friction/correction channel	Describing the broad amendment or filing-friction outcome.	A heterogeneous filing-friction and correction state, not automatically a material misstatement.
`8k_402_365`	Item 4.02 non-reliance event	Describing the underlying 8-K event.	A severe public correction signal, rare but rankable, not fraud occurrence.
`8k_402_365`	public Item 4.02 material-correction proxy	Describing the label's construct role.	A proxy for severe public correction, not a complete material-misstatement or fraud label.

BAR Positioning

The paper should read as accounting research before it reads as a modeling exercise.

BAR-facing contribution:

A measurement redesign for reporting-risk research under partial observability.
A public-data-first, filing-origin outcome system for public review-and-correction risk.
A reproducible SEC/PCAOB public-data architecture with dated sources, origin-date discipline, and provenance.
A bridge-gated comparison to detected-misstatement benchmarks that supports related-but-non-identical construct validation.
Metric-compatible model-family transfer to familiar Dechow, Perols, Bao, and Bertomeu-style vocabularies without claiming same-estimand superiority.

Not the contribution:

Unobserved fraud occurrence detection.
Causal identification of SEC, issuer, auditor, or enforcement behavior.
A direct performance ranking over prior fraud-prediction papers.
A deployment-ready regulatory or audit automation system.
A claim that public comment letters are the complete SEC review universe.

BAR-specific positioning:

Route the paper primarily as financial-reporting measurement and disclosure review, with audit/regulatory oversight as a secondary fit because PCAOB Form AP can supply auditor and engagement-partner features. PCAOB inspection archives are Bronze/provenance inputs only, not Gold or model features.
Let the UK/EU discussion support cautious portability. The paper can say that the measurement design is portable to settings with dated filings and public review or correction channels; it cannot claim out-of-U.S. validation.
Policy and practice implications should remain tied to measurement, screening, and public-data reproducibility rather than investor-return or deployment claims.
BAR's closed special-issue activity on AI in accounting and finance may be used as a contextual scope signal in cover-letter thinking only after source verification; do not present a closed call as a live submission target.

Stable Terminology

Use one term for one concept whenever possible.

Preferred terms:

Term	Use when	Do not confuse with
filing-origin, submission-dated review-and-correction outcome state	Naming the paper's central construct.	Detected misstatement, latent fraud occurrence, or a release-time backtest.
submission-dated review-and-correction outcomes	Describing the downstream outcome family.	Full SEC review, enforcement, or hidden misconduct.
submission-dated, eventually disclosed SEC correspondence	Describing the comment-thread outcome.	Correspondence public availability on the submission date or the complete SEC review universe.
public amendment/friction/correction channel	Describing the heterogeneous amendment outcome.	A material-misstatement label.
Item 4.02 non-reliance event	Describing the 8-K event itself.	Fraud occurrence.
public Item 4.02 material-correction proxy	Describing the label's role as a severe public-correction proxy.	A complete public-material-misstatement universe.
detected-misstatement benchmark	Describing the `gvkey x data_year` benchmark layer.	The public-cascade estimand.
benchmark layer	Referring to detected-misstatement diagnostics and peer-model checks.	Public-cascade prediction.
public cascade layer	Referring to filing-origin public labels and public features.	Benchmark model-family comparison.
bridge gate	Referring to the gvkey-CIK-year linkage condition before overlap claims.	A silent join or target merger.
WRDS-validated construct-overlap	Referring to integrated overlap only when the current artifacts report a WRDS-validated tier.	Any stale, candidate, or unverifiable crosswalk.
bridge-gated construct validation	Referring to the validation design that uses the bridge gate.	Proof of equivalence.
related-but-non-identical constructs	Interpreting benchmark/public alignment.	Same-estimand superiority.
model-family transfer	Describing Dechow, Perols, Bao, or Bertomeu-style families in this repo.	Original-paper numeric replication.
metric-compatible ranking evidence	Describing shared metric language across different targets.	Cross-estimand model superiority.
adjusted association	Describing DML or regression-style estimates after controls.	Causal effect.
predictive association	Describing score or feature relationships in prediction tasks.	Structural mechanism.
annual fold dispersion	Describing variation across annual out-of-time test folds.	Population confidence interval.
uncertainty caveat	Flagging interval, sparse-fold, or diagnostic-status limits.	Formal independent-sample inference.
limited screening interpretation	Translating ranking evidence into prioritization language.	A calibrated decision rule or deployment protocol.

Avoid or replace:

Avoid	Replace with
fraud detection, true fraud, latent fraud occurrence	detected-misstatement benchmark; unobserved fraud occurrence only when stating a boundary
SEC review	selected, eventually disclosed SEC correspondence, unless explicitly discussing the broader institutional process
complete enforcement universe	public review-and-correction channels
causal effect	adjusted association or predictive association
outperforms prior studies	metric-compatible evidence within this estimand
state-of-the-art model	implemented model family or peer-compatible model family
model-driven insight	measurement-and-ranking evidence
cascade as contagion, transmission, diffusion, or spillover	typed recorded outcome states
feature importance as mechanism	feature-family or information-set evidence
population confidence interval	annual fold dispersion or out-of-time evaluation dispersion

Cross-Discipline Terminology Bridge

Use short first-use definitions for terms that are familiar in one audience but not the other. Do not rely on acronyms alone.

Term	Accounting-reader-friendly definition	ML-reader-friendly accounting bridge
PR-AUC	Rare-event ranking quality read against the outcome base rate.	Precision-recall area is more informative than ROC-AUC when positives are sparse.
ROC-AUC	General ranking separation between positives and negatives.	Useful but can look high when rare-event precision is still weak.
ECE	Expected calibration error: whether predicted probabilities match observed rates.	A calibration diagnostic, not a ranking metric.
Brier Skill Score	Probability-forecast improvement relative to a baseline.	A calibration and probability-quality check.
Bao-style metrics	Top-fraction precision, sensitivity, specificity, BAC, and NDCG used for literature comparability.	A metric-language bridge to Bao-style fraud-prediction evidence, not replication.
DML-style high-dimensional adjustment	Flexible adjustment for observed controls before estimating an association.	Double/debiased ML here supports adjusted association, not causal identification.
scale_pos_weight	Class-imbalance weight used in tree models.	A modeling device for rare positives, not an economic cost function.
annual fold dispersion	Descriptive variation across annual out-of-time folds.	It is not a population confidence interval or superpopulation claim.
Item 4.02 non-reliance	A public 8-K signal that prior financial statements should not be relied on.	A severe public-correction proxy, not a direct fraud label.
PCAOB Form AP	Public audit-participant disclosure for issuer audits.	A public audit-market feature source, not proof of audit failure.
AAER	SEC Accounting and Auditing Enforcement Release.	Enforcement-tail context only; not a headline public-cascade label.

Claim Boundaries

These boundaries are non-negotiable:

Public-cascade labels are not alternative names for fraud.
The comment-thread label is selected, eventually disclosed SEC correspondence dated by its underlying EDGAR submission. It is not a public-release-time outcome and not the full SEC review universe.
Amendment evidence is heterogeneous and must not be collapsed into material misstatement evidence.
Item 4.02 evidence is sparse severe-tail evidence; it cannot alone carry the headline construct-validity claim.
DML-style high-dimensional adjustment estimates are adjusted associations, not causal effects.
Peer-compatible Dechow, Perols, Bao, and Bertomeu-style rows are model-family transfer and metric-language alignment, not original-paper numeric replication.
PR-AUC must be interpreted relative to prevalence/base-rate.
Top-k precision and lift are screening evidence and ranking evidence, not calibrated decision rules.
Point-in-time discipline is mandatory: financial and public predictors must be point-in-time, as-first-reported, or explicitly available at or before the filing-origin date. Any restated, retroactive, or post-origin predictor ambiguity is a look-ahead-bias risk.
Submission-dated comment-thread history must not enter filing-origin predictors unless public-release timing is independently verified. A submission before origin_date does not establish public availability before origin_date.
Public-side folds must use complete origin-calendar-year holdouts. Training rows may enter a holdout only when their 365-day outcome windows mature before the holdout year begins. This establishes temporal ordering under the stated submission-date outcome clock, not a live correspondence-release backtest.
Uncertainty intervals are annual fold dispersion summaries only when valid folds >= 5. Sparse folds with fewer than 10 positives must be excluded from formal interval displays and listed in provenance.
Asymmetric costs and false negatives may motivate limited screening interpretation, but the current paper does not estimate a utility model.
Reciprocal risk-score alignment and event-time concentration are diagnostic validation evidence, not equivalence or causality.
Treat retired enforcement-tail outputs as outside the current paper-facing design. This means deprecated AAER or enforcement-tail artifacts, labels, or tables are archival/status-only unless the paper plan explicitly reactivates them. They cannot enter headline public-task claims, feature-family rankings, peer comparisons, robustness claims, or manuscript tables.

Public-data discipline:

Treat public-data-first as a scientific information-set design, not a budget compromise.
Treat the public-data-only design as reproducible accounting measurement, not a weaker substitute for paid data.
Do not make absence of Audit Analytics or other unprovided commercial datasets sound like a fatal manuscript flaw.
Do not recommend paid data as required for the current v1 paper.
Do not browse for live pricing, data-budget pages, or commercial-data availability unless the user explicitly asks for a separate data-budget task.

Claim-Strength Ladder

Use this Claim-strength ladder in every audit:

reportable finding: directly supported by current generated artifacts and safe for main-text use.
candidate evidence: artifact-backed but limited by sparse labels, coverage, bridge tier, or another explicit validation condition.
diagnostic only: useful for motivation, design checks, robustness framing, or appendix context, but not a standalone headline claim.
not supported: absent, contradicted, stale, or too speculative for the current paper.

Evidence Order

Read current evidence in this order:

docs/paper_plan.md for the research-design contract.
docs/results_snapshot.md for artifact-backed outcomes and current evidence boundaries.
reporting-risk-cascade-manuscript/main.tex, tables, figures, statements, submission files, provenance files, and audit notes.
Current generated artifacts when a claim depends on them.
If artifacts/full_with_peer exists, inspect the study manifest, public cascade outputs, peer-comparison outputs, bridge probe, and construct-overlap ledgers before allowing an integrated claim.

Important artifact families include:

study_summary.md
study_run_manifest.json
detected_misstatement_model_family_metrics.csv
public_model_family_metrics.csv
public_model_family_task_status.csv
construct_overlap_summary.md
label_contingency_lift.csv
public_score_benchmark_ranking.csv
reciprocal_alignment.csv
event_time_concentration.csv

P0 and P1 Checks

Complete these checks before style polish.

Before stylistic polish, three claim-level checks dominate: the bridge tier must support any integrated overlap claim; submission-dated, eventually disclosed SEC correspondence must not be treated as release-time data or the full SEC review universe; and sparse Item 4.02 evidence must remain rare-tail support rather than the whole validation story.

P0 checks:

Bridge freshness check. Verify the current bridge tier from the study manifest and construct-overlap artifacts before allowing integrated construct-overlap claims. If the tier is missing, stale, candidate-only, or inconsistent across artifacts, downgrade overlap language to candidate evidence or diagnostic only. If the current tier is wrds_validated, WRDS-validated construct-overlap language may be reviewed as reportable subject to the stated caveats.
Comment-letter date and selection check. Confirm that the manuscript says comment_thread_365 is dated by the underlying EDGAR correspondence submission, not by public release. The outcome is eventually disclosed and selected; a model that ranks it may rank both issuer reporting-risk signals and the SEC process that selects correspondence for later disclosure.
Primary-source citation check. Require primary-source citations for SEC, PCAOB, BAR, and Elsevier policy claims. Public-release timing, double-anonymization mechanics, AI disclosure, Form AP, inspection data, Item 4.02, and AAER coverage claims should be supported by SEC, PCAOB, BAR, or Elsevier primary sources; otherwise flag P1.
Sparse Item 4.02 check. Check whether any headline claim relies on sparse 8k_402_365 evidence. If so, require rare-but-rankable and severe-tail caveats.
Retired enforcement-tail check. Confirm that deprecated AAER or enforcement-tail artifacts are absent from the current paper-facing claims unless explicitly reactivated by the Paper Plan.
Article-prose contamination check. Check for article prose that still sounds like a response memo or design memo, especially phrases where "the manuscript" is the subject.

P1 checks:

Comment-letter literature differentiation check. Explain why the paper is not merely another comment-letter prediction paper. Predicting comment threads is not novel by itself; the novelty is the filing-origin estimand, typed public cascade, public-data architecture, and bridge-gated validation.
Keep the accounting problem before the method in the Abstract, Introduction, and Discussion.
Ensure predictor text and outcome text are separated. Outcome-side amendment, Item 4.02, or comment-letter response text filed after origin_date must not be used as a predictor for the event it defines.
Point-in-time predictor check. Verify point-in-time predictor discipline. Financial and public predictors should be as-first-reported or available at or before the filing-origin date; retroactively restated or post-origin fields require explicit exclusion or a look-ahead-bias caveat.
Public-side fold check. Verify that public-side holdouts use complete origin-calendar years, that every training outcome window matures before the holdout starts, and that the prose does not misstate this ordering as a release-time availability backtest. Submission-dated correspondence histories must be excluded from predictors unless release dates are verified.
Issuer-dependence check. Verify that public-opacity nuisance folds keep repeated rows for an issuer together and that reported DML-style intervals use issuer-clustered covariance. Construct-overlap bootstrap intervals must resample issuers rather than rows. If either artifact remains row based, downgrade its inferential language and require an explicit dependence caveat.
Keep feature-family results as feature fusion, not XBRL dominance. The bounded phrase is: feature fusion helps, metadata remains strong.
Apply a common-sample / coverage caveat to XBRL, auditor, prior-filing history, and other narrower feature-family comparisons. If the comparison is not on a common evaluation sample, the text must state that differences may reflect both source coverage and signal content.
Missingness and opacity check. Distinguish missingness cases before interpreting opacity: economically informative missingness, public-source coverage limits, parser/source unavailability, and model imputation are not the same. Do not interpret a missingness pattern as strategic silence unless the artifact supports that claim. If public-opacity DML p-values are null, weaken strategic-silence language.
Table/figure interpretation check. Require table and figure captions for performance evidence to state the evaluation unit, task, base-rate context, and bounded measurement interpretation. Captions should translate the number into a measurement claim, not a model-performance slogan.
Model-selection optimism check. If the manuscript highlights the best model, best window, best feature family, or max PR-AUC, require descriptive language unless there is a pre-specified selection rule or validation correction. A single best configuration should not become the headline claim.
Table-unit clarity check. Audit every column named "Rows". If it counts metric rows, fold-task rows, task-window-feature rows, or model-family rows rather than issuer-years or firm-years, rename it or require a table note explaining the evaluation unit.
Avoid claims about unpriced market risk unless market variables and current artifacts directly support the claim.
Do not let any single-year 8k_402_365 extreme dominate the evidence narrative.
Ensure uncertainty intervals are only reported when valid folds >= 5, and sparse folds with fewer than 10 positives are properly excluded from formal interval displays.
Scan for AI-flavored prose patterns: triple-hedge sentences, over-compressed abstract sentences, formulaic "this paper contributes by" claims without accounting motivation, and balanced-but-empty transitions.

Elegance and Readability

Use structure to lower cognitive burden. Results paragraphs should usually follow the same four-step rhythm:

What was tested, in accounting language.
What was found, with numbers and prevalence context.
What it means for the accounting construct.
What boundary or caveat prevents overclaiming.

When explaining machine-learning concepts to accounting readers, move from intuition to name to role: explain the practical idea first, name the method second, and state why it is needed in this paper third. For international relevance, embed UK/EU examples as portability conditions: the U.S. setting is the first test, while settings such as the UK FRC Corporate Reporting Review require jurisdiction-specific label mapping and validation.

Voice and style examples:

Weak or formulaic	Better academic version
This paper contributes to the literature in three ways.	The contribution is to make the observability system part of the reporting-risk construct.
The model robustly outperforms alternatives.	The specification has the highest reported mean PR-AUC in this table, subject to fold-dispersion and coverage caveats.
The results show that AI can detect fraud.	The evidence shows that filing-origin information ranks later submission-dated correspondence and public correction outcomes above their base rates.
The table reports 180 rows.	The table reports 180 metric rows across task-window evaluations, not 180 issuer-years.

Audit Workflow

Drafting stage	Primary mode
Empty or placeholder manuscript only	Skeleton Mode
Writing or revising one section	Section-Level Polish
After a section is drafted	Triage audit
After substantive body completion	Claim-To-Evidence Audit
When the full paper reads coherently	BAR Readiness Audit
Before submission	Full audit plus Referee Report
After revisions or reviewer comments	Delta audit

Review Mode: BAR Readiness Audit

Purpose: whole-manuscript review for BAR fit, construct integrity, and likely referee objections.

Referee robustness check:

What is the accounting insight beyond another classifier?
Is the estimand clearly distinct from ex post detected misstatement?
Why is submission-dated, eventually disclosed SEC correspondence a defensible selected outcome?
Does the paper separate reporting-risk prediction from predicting which filings enter an eventually disclosed correspondence record?
Why is public-data-first a scientific information-set design?
Does the benchmark layer remain diagnostic rather than a same-target contest?
Does the paper remain credible for a BAR reader who is not interested in model implementation details?

Review dimensions:

Core contribution and "so what".
Construct and label integrity.
Comment-letter selection and endogenous scrutiny.
Timing, censoring, and leakage.
Rare-event evaluation and Item 4.02 discipline.
Bridge-gated construct validation.
Literature positioning, especially the comment-letter prediction literature.
Economic mechanism and information-set interpretation.
Formulaic prose, response-memo residue, and terminology drift.
BAR fit, international relevance, and submission mechanics.

Deliverable:

Manuscript readiness: not started, outline only, partial draft, submission draft, or not assessable.
BAR fit judgment: one sentence.
Main paper risk and likely rejection path: one sentence each.
Best current positioning: one sentence.
Findings ordered by P0, P1, and P2.
For each finding: title, severity, location, evidence, why it matters, and concrete revision instruction.
Required ledgers: terminology ledger, claim-to-evidence + claim-strength table, bridge-gate assessment, economic-mechanism assessment, citation audit, public-data support ledger, and no more than 10 revision priorities.

Review Mode: Claim-To-Evidence Audit

Purpose: claim extraction and evidence classification before substantive prose polish.

Audit every empirical claim and classify it with the Claim-strength ladder.

Benchmark layer checks:

Are timing diagnostics described as label-observability sensitivity?
Are financial predictors explicitly point-in-time, as-first-reported, or available at or before the filing-origin date?
Are res_an* outputs kept out of predictors and outside paper-grade label maturation unless externally validated dates exist?
Are naive, proxy_drop_observed, and proxy_imputed_lag modes distinguished?
Are drift and missingness results diagnostic unless stronger evidence exists?

Public cascade layer checks:

Are comment_thread_365, amendment_365, and 8k_402_365 reported separately?
Are headline public tasks are comment_thread, amendment, and 8k_402?
Are public-cascade labels described as typed recorded outcomes, not fraud or real-time public-release states?
Are PR-AUC, ROC-AUC, Brier, Brier Skill Score, ECE, top-k precision, top-decile lift, and Bao-style metrics interpreted under rare-event prevalence/base-rate?
Are top-k and lift framed as screening evidence, not calibrated decision rules?

Bridge and overlap checks:

Is bridge status read from the current manifest and construct-overlap artifacts, rather than assumed from stale prompt text?
If the current artifacts state wrds_validated, is WRDS-validated construct-overlap language used only for the integrated overlap layer?
Are high-confidence, ambiguous, and dropped bridge cases distinguished when overlap claims depend on them?
Is construct overlap described as related-but-non-identical?
Is reciprocal risk-score alignment presented as diagnostic validation rather than equivalence?
Is event-time concentration used as validation context rather than causal proof?

Sparse Item 4.02 checks:

Does any headline claim rely too heavily on sparse 8k_402_365 evidence?
Does the text describe Item 4.02 as rare but rankable?
Does it avoid promoting a single-year 8k_402_365 extreme over averages or validation ledgers?
If top-decile lift is strong but PR-AUC is close to a low base rate, require a sparse severe-label caveat.

Peer-compatible model-family checks:

Are Dechow, Perols, Bao, and Bertomeu-style models described as model-family transfer and metric-language alignment?
Does the manuscript avoid original-paper replication claims unless mapping and sample gates support that claim?
Does it avoid outperformance or superiority language across estimands?

DML and opacity checks:

Are DML estimates described as DML-style high-dimensional adjustment or adjusted association?
Are machine-learning scores described as predictive association rather than structural mechanism?
If p-values are null, does the manuscript avoid a strong strategic-silence claim?
Does the manuscript distinguish disclosure breadth and tag entropy from mechanical opacity?
Does it distinguish economically informative missingness, public-source coverage limitations, parser/source unavailability, and model imputation?
Does the manuscript avoid claims about unpriced market risk unless market variables and artifact support are present?

Deliverable:

A table with: claim, manuscript location, supporting artifact, classification, problem if any, and required revision.
The three most dangerous unsupported claims.
The three claims strong enough to lead the paper.
The three claims that should move to limitations, robustness, or future work.

Review Mode: Section-Level Polish

Purpose: one-section revision after the claim audit has cleared the section's empirical statements.

Section objectives:

State the accounting problem before the method.
Define the estimand before discussing models.
Keep the benchmark layer, public cascade layer, and bridge gate distinct.
Use mechanism-aware accounting prose without converting prediction evidence into causal mechanism.
Remove response-memo and formulaic prose.
Remove promotional model language.
Convert overclaiming into disciplined contribution language.
Explain why each table or figure matters immediately after the callout.
Ensure performance captions state evaluation unit, task, base-rate context, and bounded measurement interpretation.

Style constraints:

Prefer short, high-information sentences.
Avoid "the manuscript should", "this draft", and similar internal-process residue in article prose.
Avoid repetitive contribution lists without economic logic.
Do not call the cascade contagion, transmission, network diffusion, or spillover.
Do not call the task fraud detection.
Do not use feature importance as mechanism.
Do not use "outperforms", "superior", "state-of-the-art", "model-driven", "robust insights", "comprehensive framework", or "to the best of our knowledge" to inflate claims.

Results and Discussion polish:

Translate precision or lift into practical screening economics only when the table values directly support the translation. If prevalence is 0.0208, the baseline screen implies roughly one positive per 48 issuer-years; do not invent utility or cost savings.
Treat best-model, best-window, and max-PR-AUC language as descriptive ranking evidence unless selection was pre-specified or separately validated.
Keep feature-family results as feature fusion, not XBRL dominance.
Use common-sample / coverage caveat language for XBRL, auditor, and prior-filing history feature-family comparisons.
Treat public-data reproducibility as open-science and accounting-measurement value. Do not oversell deployment-ready RegTech.

Subsection template:

What was tested, in one accounting-framed sentence.
What was found, with numbers and prevalence/base-rate context.
What the result means for the accounting construct.
Boundary, caveat, or diagnostic status.

Deliverable:

Section diagnosis.
Structural revision plan.
Five sentence-level revisions: original, problem, revised.
Polished section text.
Residual caveats needing artifact verification or citation support.

Review Mode: Abstract And Introduction

Purpose: storyline revision before submission or after a major audit.

The Abstract and Introduction should answer the BAR referee's "so what" question without overstating the evidence. The sequence should be:

Measurement problem: observed restatement, correction, comment-letter, and enforcement labels mix occurrence, detection, timing, institutional selection, and public observability.
Estimand: filing-origin, submission-dated review-and-correction outcome state.
Public-data design: dated filing-origin predictors, correspondence submission dates, and public correction dates give a sharper event clock than many year-level detected-misstatement labels without implying correspondence was public on submission.
Public cascade: submission-dated eventually disclosed correspondence, amendment/friction/correction, and Item 4.02 non-reliance as typed recorded outcomes.
Evidence architecture: benchmark layer, public cascade layer, bridge gate.
Contribution: measurement redesign under partial observability, reproducible public-data architecture, typed cascade ontology, and bridge-gated validation.
Main evidence: filing-origin information ranks future submission-dated correspondence and public correction events; correspondence is broad and ranked above base rate across folds; amendments are intermediate; Item 4.02 is rare but rankable; feature fusion helps, metadata remains strong; bridge-gated overlap supports related-but-non-identical constructs.
Claim boundary: not unobserved fraud occurrence, not causal enforcement effects, not full SEC review universe, not full enforcement universe, and not same-estimand superiority over prior fraud-prediction papers.

Deliverable:

Revised Abstract, normally 150-250 words unless live BAR guidance states otherwise.
Revised Introduction opening, 5-7 paragraphs.
Contribution paragraph.
Claim-boundary paragraph.
Sentences to delete because they sound like a design memo or response memo.

Review Mode: Referee Report

Purpose: skeptical but fair referee simulation after the polished draft compiles.

Write a skeptical but fair referee report. Do not rewrite the paper. Identify the main reasons the paper would receive a major revision.

Focus on:

whether the estimand is distinct from prior detected-misstatement prediction;
whether submission-dated correspondence is a defensible selected outcome;
whether the paper is too close to the comment-letter prediction literature;
whether the amendment label is too heterogeneous;
whether Item 4.02 is too sparse for headline inference;
whether WRDS-validated bridge evidence is sufficient for overlap claims;
whether public-data-first is positioned as scientific design rather than constraint;
whether model evaluation is appropriate for rare events;
whether peer-compatible model-family results are overinterpreted;
whether feature-family evidence has economic content;
whether the paper honestly handles partial observability, endogenous scrutiny, detection delay, right censoring, and public-source incompleteness.

Deliverable:

Summary assessment, one paragraph.
Major concerns, 5-8 items.
Minor concerns, 5-10 items.
Required revisions for credible resubmission.
Best possible contribution if revised correctly.
What the paper must not claim.

Audit Output Types

Use a triage audit for fast P0/P1 prevention while drafting.

Triage audit deliverable:

Audit date, repo SHA if available, manuscript readiness, audit mode, BAR fit judgment, main paper risk, and likely rejection path.
P0/P1 findings only.
Claim-to-evidence + claim-strength table for at most five main claims.
Bridge-gate assessment.
No more than five revision actions.

Use a full audit when the paper is close to submission.

Full audit deliverable:

All P0/P1/P2 findings.
Claim-to-evidence + claim-strength table.
Terminology ledger.
Bridge-gate assessment.
Economic-mechanism assessment.
Literature and citation audit.
Public-data support ledger.
Table/figure architecture review.
Abstract audit.
Conclusion audit.
Revision priorities.

Use a delta audit after revisions.

Delta audit deliverable:

Changed sections or files reviewed.
Earlier P0/P1 issues marked resolved, partially resolved, or open.
New or changed claims only, unless they affect the global story.

Skeleton Mode

Fallback only. The current manuscript is drafted; use skeleton mode only if the manuscript folder is empty or has no real body. Do not simulate a review of nonexistent text.

Skeleton deliverable:

Working title and thesis.
Abstract target of 150-250 words unless live BAR guidance states otherwise.
Introduction: measurement problem, contribution, main evidence, boundaries.
Literature and construct development: detected-misstatement labels, partial observability, public disclosure review, audit/oversight signals, and machine learning as measurement infrastructure.
Data and label construction: public SEC/PCAOB/EDGAR sources, origin date, public labels, censoring, public/private status, and bridge gate.
Research design: filing-origin information set, model families, metrics, benchmark layer, public cascade layer, and construct-overlap design.
Results: public cascade performance, feature-family evidence, peer model-family alignment, and construct-overlap evidence.
Additional analyses and robustness: common-sample checks, sparse Item 4.02 interpretation, DML-style high-dimensional adjustment, and sensitivity checks.
Discussion: limited screening interpretation, asymmetric costs, false negatives, public-data reproducibility, and international relevance.
Conclusion: accounting insight, strongest evidence, limitations, and honest generalizability.

BAR Mechanics and Citation Discipline

Use the official BAR Guide for Authors and Elsevier instructions as the live source for submission mechanics. Verify mechanics against primary sources before upload.
Keep manuscript files double-anonymized when current BAR/Elsevier instructions require it, with title page and anonymized manuscript separated as instructed.
Keep the abstract concise and factual.
Provide no more than 6 keywords unless live instructions say otherwise.
If highlights are required or submitted, keep them short, specific, factual, and within the current Elsevier length rule.
Tables should be editable text with concise titles and self-contained notes.
Include data availability, competing interest, CRediT, and tool-use disclosures as required by Elsevier and coauthor policy.
Use primary sources for SEC/PCAOB/BAR/Elsevier rules.
If live web verification is needed, use primary sources only and state exactly which claims were verified.
Elsevier-style AI policy should be kept separate by role. Author-side AI tools may be used only with human oversight, disclosure, and author responsibility under current instructions. Reviewers and editors should not upload submitted manuscripts or confidential peer-review material to generative-AI tools. This brief is author-side QC, not journal peer review.
A closed BAR special issue or article collection on AI in accounting and finance can support scope alignment only as verified editorial context; it is not a live special-issue target and should not make the cover letter sound AI-led. Verify through primary sources such as the BAR special issues page or the AI in Accounting and Finance Research and Practice collection.

Core literature coverage should include:

Dechow, Ge, Larson, and Sloan on the F-score / misstatement-prediction tradition.
Perols on classifier comparison.
Bao, Ke, Li, Yu, and Zhang on machine-learning fraud detection.
Bertomeu, Cheynel, Floyd, and Pan on machine-learning misstatement detection.
Barton, Burnett, Gunny, and Miller on occurrence versus detection.
Comment-letter and disclosure-review studies, including the comment-letter prediction literature.
Restatement, amendment, and correction studies.
Auditing and PCAOB/Form AP/inspection oversight work.
Disclosure opacity, XBRL, structured reporting, and public-data processing.
Partial observability, selection, drift, rare-event ranking, and cost-sensitive evaluation.
International accounting and enforcement portability when discussing UK/EU relevance.

Table And Figure Architecture

Keep numbering generic in audit language. Refer to functions, not fixed numbers, because numbering can change during revision.

The conceptual figure should show the reporting-risk cascade as typed recorded outcomes, not contagion or a release-time sequence.
The label ontology table should define construct, horizon, source, and caveat.
Public-lake scale tables should separate public-source construction from benchmark compatibility.
Performance tables should report PR-AUC with prevalence/base-rate and an uncertainty caveat or fold dispersion where available.
Performance tables and figures should state evaluation unit, task, base-rate context, and bounded measurement interpretation in the caption or note.
Feature-family tables should be information-set evidence with a common-sample / coverage caveat.
Feature-family tables should not let model-selection optimism or max metrics become the headline claim.
Peer-model tables should state model-family transfer, not original-study replication.
Construct-overlap tables should show related-but-non-identical evidence, not equivalence.
Opacity/DML tables should state adjusted association and avoid causal language.
Captions must tell the bounded economic or measurement interpretation without requiring repository knowledge.