Dataset Verification

Corpus v3.0 — Publication Snapshot

Designed to Detect, Not to Correct is computed from a frozen, immutable snapshot of the Inside the Work oversight corpus: publication version 3.0, frozen July 6, 2026. This page publishes the verification record referenced in Appendix A — the snapshot's SHA-256 hashes, the 95-column schema, the two recurrence measures, the denominator register, the status-currency disclosure, and the Gate G1 corrections log — the hashes fix the identity of the frozen files, the schema and the register fix what every figure was computed against, and any later copy of the dataset can be checked against that freeze. Reproducing the figures themselves requires the coded records.

The Corpus

Every empirical claim in the book is computed from a fixed corpus: 1,219 publicly available reports from federal inspectors general and the Government Accountability Office, spanning fiscal years 1999 through 2026, drawn from ten issuing offices. Of those, 1,006 document substantive findings; 690 of them carry at least one recommendation, 3,437 in all. The 1,006 are the default denominator for the book’s report-level rates; the denominator register below lists every narrower population and its exclusion rule.

The corpus is a structured evidence base, not a probability sample of federal oversight. It is weighted toward healthcare and financial management — HHS OIG, GAO, and VA OIG are its largest contributors. A corpus rate describes these 1,006 reports, not the federal universe.

The unit of analysis is the report-level record — one per report. Where a rate is stated for findings or failures, the unit is that record; a report may contain several distinct findings, which the book does not count separately. Rates per recommendation say so.

Frozen Files

oig-dataset-PUBLICATION-FROZEN-v3.0-20260706.tsv — 1,219 report records × 95 columns

SHA-256: 8bb50bbdb23802feb4632028b36ec6378177e4187f282285a485684519bcbbbd

oig-recommendations-PUBLICATION-FROZEN-v3.0-20260706.tsv — 3,437 recommendation records × 14 columns

SHA-256: 8d7df5a675be9d6897b2f648ffb58ab3e161f328cdae51525fc2693a9bef7801

Frozen July 6, 2026. The frozen files are never edited. Corrections create working versions with changelogs and a new manifest at the next freeze. Any distributed copy can be authenticated by computing its SHA-256 hash and comparing it to the values above; any byte-level change produces a different hash.

The 95-Column Schema

Roughly fifty core classification fields across five domains, extended by provenance, source-verification, and derived fields that bind every row to its source document. Locked in fixed order:

Report_ID, OIG_Office, Report_Number, Report_Title, Publication_Date, Fiscal_Year_Issued, Audit_Type, Entity_Level, Entity_Name, Scope_Type, Facility_or_VISN, Program_Area_Primary, Program_Area_Secondary, Recommendation_Count, Repeat_Finding, Prior_Report_Referenced, Contract_Involved, Contractor_Name, IT_System_Involved, Primary_System_Name, Secondary_System_Name, Dollar_Impact_Stated, Dollar_Impact_Value, Financial_Impact, Improper_Payment_Explicit, Financial_Reporting_Impact, Appropriations_Issue_Type, GreenBook_Relevant, GreenBook_Component, Failure_Mode, Governance_Failure, Governance_Failure_Type, Monitoring_Breakdown, Accountability_Failure, Control_Design_Failure, Control_Execution_Failure, Data_Reliability_Issue, Documentation_Gap, Staffing_or_Resource_Factor, Training_Gap, Contractor_Oversight_Failure, Systemic_Risk, Escalation_Level, CFO_Relevance, CFO_Function_Affected, Executive_Relevance_Level, Root_Cause_Primary, Governance_Pattern_Primary, Confidence_Level, Content_Value_Score, Final_Decision, Content_Citations, Period_Audited_Start, Period_Audited_End, Detection_Lag_Years, Report_Is_FollowUp, Prior_Report_Linked, Prior_Report_IDs, Recurrence_Confirmed, Time_Since_Prior_Years, Condition_Summary, Criteria_Type, Criteria_Citation, Cause_Narrative, Effect_Narrative, Sample_Size, Error_Rate, Projected_Population, Dollar_Context_Type, Dollar_Recoverable, Consequence_Type, Population_Affected, Context_Flags, Acting_Official_Present, Mgmt_Response_Present, Overall_Concurrence, OIG_Response_Adequacy, Governance_Pattern_Confirmed, Pattern_Tier, Failure_Layer, Maturity_Stage, Pass1_Corrections, Pass2_Confidence, Pass2_Source_Quotes, Disposition_Tag, Source_File_ID, Source_File_Name, Source_Match_Method, Source_File_Hash, Concurrence_Note, Response_Adequacy_Note, Provenance_Note, Dollar_Recoverable_Note, Root_Cause_Note, Pattern_Note

Recommendation record (14 columns): Rec_ID, Report_ID, Rec_Number, Rec_Text, Rec_Addressee, Rec_Action_Type, Rec_Criteria_Type, Auditee_Concurrence, OIG_Response_Adequacy, Target_Completion, Rec_Status, Rec_Dollar, Prior_Rec_Linked, Closability

Field definitions and controlled vocabularies are documented in Appendix A of the book.

The Two Recurrence Measures

Flagged repeat rate — 22.5 percent (226 of 1,006). The report itself says the condition repeats a prior finding and cites the prior report.

Documented recurrence — 50 percent (500 of 1,006). The union of recurrence confirmed through a linked prior report (491) and the repeat-finding flag (226), overlap counted once.

Both are coded from the report's own statement that the finding returned from a prior audit. They look backward to a prior report. They do not look forward from a closure.

The Denominator Register

Every analytic subpopulation in the book, with its exclusion rule.

1,219 — all reports in the frozen corpus.

1,006 — core report records: reports whose confirmed-pattern classification is anything other than No Adverse Finding; the 213 no-adverse-finding reports are analyzed separately (Chapter 10).

1,004 — core report records with a recorded acting-official value (Chapter 5).

1,001 — the index calibration set: core report records with a mapped control family and a classified pattern tier; five core reports lack a mapped family.

990 — failure-mode-coded report records: sixteen core reports lacked the detail to classify; unclassifiable is distinct from the coded value Other (five reports), which denotes a determinable mode outside the seven named categories.

947 — the two-tier comparison set: 609 structural plus 338 transactional; fifty-nine core report records are hybrid (55) or untiered (4) and are excluded from structural-versus-transactional contrasts.

700 — core report records with a numeric detection lag (Chapter 8).

695 — non-follow-up core report records. Follow-up here is the coded flag for a report commissioned against a linked prior report — a different thing from the Follow-Up Review report type in the corpus profile (16 reports); the 94 percent recurrence figure for follow-ups rides on the flag, not the type.

690 — report records carrying at least one extracted recommendation.

576 — report records with at least one concurred recommendation (recommendations file).

549 — report records whose report-level management response was full concurrence (Concur or Concur All).

500 — report records with documented recurrence: the union of recurrence confirmed through linked prior reports (491) and the repeat-finding flag (226), overlap counted once; the 50 percent documented recurrence rate is 500 of 1,006.

409 — report records with the dollar-impact indicator flagged; 406 of them carry a parsable positive dollar value, which is why two figures appear in different places.

273 — report records with a recoverable-dollar figure.

258 — report records with a curated, finding-attributable dollar value under the finding-attributable classification.

69 — report records with at least one recommendation recorded Closed-Implemented or Closed in the snapshot: the publication closed set. Resolved (agreed, not yet implemented, in OIG usage) is excluded, and generic Closed is not characterized as implemented; status as captured at coding, not refreshed — see Era Composition and Status Currency. 28 of the 69 carry no documented recurrence; 41 do.

38 and 31 — those 69 report records scoring at or above and below 60 on the index.

19 — closed-set report records carrying documented recurrence and an attributable dollar value at the $10 billion per-finding cap.

Era Composition and Status Currency

The corpus is weighted toward recent reports — 486 of the 1,006 core report records (48 percent) were issued in fiscal years 2025 and 2026, 442 in fiscal years 2020 through 2024, and 78 in fiscal year 2019 or earlier — and two of the book's reference measures move with that composition. The flagged repeat rate is 6 percent in the oldest stratum, 12 percent in the middle, and 34 percent in the newest. The rise appears inside every major issuing office, and it may reflect the extraction engine's more complete capture of prior-report references in later batches as much as any change in oversight practice; the corpus cannot separate the two. The corpus-wide 22.5 percent baseline is therefore a pooled figure. Every elevated repeat rate in the chapters was re-tested within the fiscal year 2025–2026 stratum alone, and every contrast holds in the same direction. Several narrow — Design Deficiency repeats at 37 percent against 34 percent for the rest of that stratum; Training Gap with Governance Failure at 38 against 33 — and the reader should weigh the direction of those contrasts more heavily than their pooled magnitude.

Recommendation status is the second composition effect. Recommendation status in the snapshot is the status captured when each report was coded, and was not refreshed to a single date before the freeze; the chapter narratives cite later closures taken from the issuing offices' trackers that the snapshot does not hold. Closure rises with report age: 10 percent of core report records issued in fiscal year 2019 or earlier carry a recorded closed recommendation, against 5 percent of those issued in fiscal years 2025–2026. Status currency also differs by issuing office. The GAO-issued recommendations in this snapshot are 98 percent open and 1 percent closed, and two of the 247 GAO core report records appear in the 69-record closed set; the refresh of GAO recommendation status against GAO's own tracking post-dates the freeze, and a pilot of that refresh found closures for ten of eleven baseline-open GAO recommendations. Excluding GAO, the report-level closure rate is 9 percent (67 of 759); restricted to reports issued in fiscal year 2022 or earlier it is 10 percent (31 of 304). The 69-record closed set in Chapter 10 is therefore a captured-status set, and the book uses it as one: to ask what the record knew about a finding when it was closed, not to state a conversion rate.

Corrections and Validation Log

Gate G1 — September 24, 2026. Before the final interior build, every closed-set and closure-dependent figure was re-derived from the frozen files by script and checked against regression assertions (226, 500, 491, 690, 576, 549 all reproduce). Four changes followed:

  1. Closed set redefined. The earlier aggregation counted a recommendation as closed if its status was Closed-Implemented, Closed, or Resolved (88 reports). In inspector-general usage, Resolved means agreed but not yet implemented; 19 of the 88 reached the set only through Resolved. The book now uses recorded closure — Closed-Implemented or Closed — 69 reports, 6.9 percent. Generic "Closed" is not characterized as implemented.
  2. Concurrence statistic. Concurred recommendations carry a recorded closure at 6.9 percent (180 of 2,610) against 1.8 percent (6 of 336) for recommendations management did not concur with — a recommendation-level measure. Among reports at least two years old: 9.4 against 2.9 percent. An earlier age-controlled figure (13.9 against 3.5 percent) had no documented specification and was withdrawn.
  3. Dollar median. A $0 value had been counted as dollar-bearing in one small-cell median; it is now excluded.
  4. Denominator. The FY 2022-or-earlier closure rate excluding GAO is 31 of 304, not of 343 (343 included GAO).

Forward citation — a later corpus report citing an earlier report as prior — is now computed by a documented rule: a record's report number, exactly or with a leading GAO-, DODIG-, OIG-, or A- prefix removed, appearing as a whole token in another core record's prior-report field. Under that rule, 190 of 1,006 core records are forward-cited, 10 of them among the 69 closed-set records.

Method

The corpus was assembled from the public report libraries of the issuing offices within the inclusion criteria, seeded from VA OIG and expanded across agencies — a structured corpus, not a probability sample of federal oversight output and not a complete enumeration of those libraries; the corpus profile and the era composition above describe what it holds. Each report was processed through a five-pass extraction engine into the 95-column record with controlled vocabularies, then frozen; the publication snapshot is immutable and hashed, and every figure in the book computes from the frozen files. Full methodology, coding disclosures, and inter-rater agreement coefficients appear in Appendix A of the book.

The reproduction scripts and the Gate G1 validation record are available to reviewers on request.

How to Cite

Inside the Work federal-oversight governance dataset, frozen publication snapshot v3.0 (July 6, 2026). Haverstick Advisory and Media LLC. insidethework.com/dataset.

Verification questions: joseph@insidethework.com