Analysis from the Oversight Corpus
Each analysis published here works one verified figure from the frozen Inside the Work oversight corpus through the same structure: the pattern the data shows, the mechanism that produces it, what it implies for execution, and the takeaway a practitioner can act on.
Every statistic traces to the frozen v3.0 publication snapshot — 1,006 substantive-finding records, 3,437 tracked recommendations — documented and verifiable at the dataset verification page. New pieces land here first, then go out through the Inside the Work newsletter.
Why Findings Come Back
Half of the 1,006 substantive-finding records in the frozen Inside the Work oversight corpus document a failure the government had already documented.
The flagged repeat rate
22.5%
226 of 1,006
A report announcing its own repetition — the number the report admits about itself.
The documented recurrence rate
50%
500 of 1,006
The condition confirmed through linked prior reports, or flagged by the report itself.
Two measures of recurrence — separate constructs, one frozen corpus. Overlap counted once; the flagged set is a component of the documented measure.
The pattern
Read one of the 1,006 substantive-finding records in the frozen Inside the Work oversight corpus and there is roughly a one-in-two chance you are not reading something new. You are reading a rerun. Across those 1,006 records, 22.5 percent — 226 records — carry the Repeat Finding flag: the report itself announces that this condition has been cited before. That is the number the system admits about itself. The fuller measure adds recurrence confirmed through linked prior reports — the same condition, documented in an earlier report, counted once even where a record carries both markers — and documented recurrence reaches 49.7 percent: 500 of the 1,006 records. Half of these records are the government finding what it already found.
Two strata inside the corpus bracket that headline from both directions. Among the 311 records coded as follow-up work against a linked prior condition, the prior condition was confirmed 94 percent of the time — when oversight goes back to check, the failure is almost always still there. Among the 695 records not coded as follow-up work, documented recurrence still runs 30 percent. The two strata reconstruct the headline exactly: 292 plus 208 is 500. Across both strata, recurrence remains substantial: a large share of what federal oversight finds is not new.
The mechanism
Findings do not come back because nobody noticed them. Detection is the stage the oversight loop performs best — the finding exists, the evidence is cited, and the condition is named. Findings can come back even after the machinery that follows detection has produced everything it is designed to produce: a recommendation is drafted, a concurrence is signed, a corrective-action plan is filed — and the underlying condition, the reporting line that splits authority from accountability, the monitoring function wired to no consequence, the control that runs and changes nothing, can still be standing when the paperwork closes.
That distinction — between closing a recommendation and correcting the condition — is its own problem, with its own numbers, and it is where this series goes next.
The implication
For an executive, the practical meaning is that recurrence is not an edge case. In this corpus, it is roughly half the record. A new finding on your desk is as likely to be an old finding wearing a new report number as it is to be genuinely new information. That changes the first question worth asking. Not: what does the finding say? But: when did this organization see this condition before, what did it certify about the fix, and what in the structure changed afterward? A finding with a lineage is no longer only a detection problem. It is evidence that prior detection did not prevent the condition from returning.
The execution takeaway
Before accepting any corrective-action plan, pull the lineage. If the condition has been documented before, pull the prior response as well. Where a corrective-action plan existed, the new plan must answer four questions the prior response did not resolve: who owns the outcome by name, what specific action closes the gap between detection and consequence, on what interval the owner verifies the condition — not the paperwork — and what happens to whom if the condition returns. A plan that cannot answer all four is still a document about the failure. And this corpus shows how often documented failures return.
Method note. Computed from the frozen Inside the Work oversight corpus (v3.0, July 2026) — 1,006 substantive-finding records drawn from 1,219 federal oversight reports. Verification record at the dataset page. Two recurrence measures appear here by design, per the corpus claim register: the flagged repeat rate (226/1,006 = 22.5 percent) is a component of the documented recurrence rate (500/1,006 = 49.7 percent, published as 50 percent); overlap is counted once. Strata of the same measure: follow-up 292/311 = 93.9 percent; non-follow-up 208/695 = 29.9 percent; 292 + 208 = 500. Rates describe this corpus, not the federal universe.