No federal penalty in eleven fiscal years — a state may omit 25 percent of its death in custody records and still be found compliant

What is happening?

The Death in Custody Reporting Act of 2013 requires every state to report each death in law enforcement custody to the Department of Justice every quarter, and it gives the Attorney General exactly one lever — a reduction of up to 10 percent in the Byrne JAG award of a state that does not comply. As of 2026-08-08 that lever has never been applied to any state. The Department published the fact itself in its implementation report of September 2022, and the Marshall Project reported on 2025-08-06 that it still held, eleven fiscal years after the reporting duty took effect. Over the same period 70 percent of the records states submitted in fiscal year 2021 were missing at least one element the statute requires, a finding GAO delivered in testimony on 2022-09-20.

What the Department adopted instead is a compliance standard that writes the shortfall into the rule. Under the BJA compliance guidelines of December 2023, a state is compliant if at least 75 percent of reportable death records arrive without missing or unknown information, and if more than 75 percent of the deaths BJA itself finds in an open-source review are reported. A state may therefore omit a quarter of the required content and a quarter of the deaths the government already located in the press, and still be found compliant. That document is the one GAO credits with closing its recommendation.

The official statistic disappeared alongside the enforcement. The Bureau of Justice Statistics closed its Mortality in Correctional Institutions collection on 2021-03-31 with 2019 as the last reference year, and its arrest-related component had already been suspended in 2014 over data quality. The duty moved in 2016 from that statistical agency to the Bureau of Justice Assistance, a grants agency, because a statistical agency collects for statistical purposes only and cannot use what it collects for enforcement. The tables BJA publishes carry a disclaimer that they are not official statistics of the federal government. Deaths in federal custody are unaffected and remain an official federal statistic, which makes the asymmetry exact — the smaller half of the national count is official and the larger half is not.

Whose problem is this?

The body that would apply the penalty is the same body that published a reasoned argument for why applying it would make the data worse. The body that would fix the design is Congress, which was asked to do so in September 2022 and has not.

Where does this problem end?

The boundary matters because the collection exists and runs on schedule while the thing it was built to guarantee does not follow from it. Measurement is not what is absent here. A statute assigns the duty, fifty-six jurisdictions file quarterly, a federal system flags records with missing fields and returns them, and an annual cross-check against outside databases runs. What is absent is a consequence.

What is the state now, and what should it be?

Now

Needs a new measurementthe target state: no source opened here names a completeness level the collection is meant to reach, a date by which full compliance is expected, or a number of states expected to reach it. GAO asked for exactly those targets in 2022. The guidelines of December 2023 answered with per-record thresholds instead, set them at 75 rather than 100 against a statute that requires every listed element for every death, and no opened source explains where the figure of 75 came from.

How big is it?

Between 681 and about 32,000 records, across fiscal years 2020 through 2024. The two bounds rest on different kinds of evidence, and the distance between them is the finding rather than the noise.

The low bound is 681. That is the number of deaths an independent audit documented one case at a time as absent from the federal data, matched against news reports, an outside university database and reader submissions, and published on 2025-08-06. Over 450 of them were in a single state. It is a hard floor because each entry is a specific person. It is also short of the stated span: the underlying dataset ends 2023-09-30, and no comparable outside audit covers fiscal year 2024.

The high bound is about 32,000 and it is arithmetic laid on top of comparisons the government published about itself. Two terms. First, deaths never reported at all — the federal report of March 2023 shows a shortfall of 845 arrest-related deaths, 247 jail deaths and over 900 prison deaths against its own chosen comparators, which is on the order of 2,000 a year and about 10,000 across five fiscal years. Second, deaths that were reported with a required element missing — GAO measured that at 70 percent of records in fiscal year 2021, and applied to roughly 6,300 records a year, a rate derived from the 25,393 records the audit obtained for four fiscal years, that is about 22,000.

Every step in the upper bound extends a single-year rate across five years, and the three components of the annual shortfall rest on three different time bases. It is an order of magnitude and not a measurement, and the subtraction of 845 and 247 is arithmetic performed here rather than a figure any agency published.

What neither bound can count: deaths in facilities that have never reported at all, because no roster of the facilities that owe reports exists and therefore no ceiling can be derived; jail and prison deaths missed by every outside list, because the only external cross-check covers police-caused deaths; prison deaths outside the National Prisoner Statistics frame, which the federal report states excludes people unsentenced or sentenced to one year or less and therefore falls hardest on the six states with combined jail and prison systems; deaths in federal custody, immigration detention and tribal facilities, which sit outside this collection entirely; and the families, whom no federal source counts and for whom no sourced multiplier exists.

Under what conditions does it arise?

1. Most states cannot compel the agencies that hold the records. The Department reported in September 2022 that most states have no law requiring local agencies to report deaths in custody to the state government, and that state administering agencies name that lack of leverage as their top concern about compliance. The data the statute demands sits with county sheriffs, municipal police departments, jails and medical examiners. The grant that would be cut belongs to the state.

2. The federal collection has no roster and therefore no denominator. The reporting guidance of April 2025 states that BJA does not hold a list of jail points of contact to share, and that each state determines which facilities are required to participate. Without a roster there is no way to notice a facility that has never reported at all. BJA also records that it presumes state submissions to be complete and accurate.

3. The only external cross-check covers one category of death. The federal completeness review compares state submissions against two private databases that track police-caused deaths, and the federal report notes that one of them includes only deaths resulting from police shootings. A death inside a jail or a prison that nobody reports is invisible to that check by construction, and the same report states plainly that no comparable outside source exists for jail deaths.

4. The number has no consumer. The statute does not require publication. The study the Attorney General owed Congress within two years of enactment, due December 2016, has not been delivered. The published tables are masked wherever a value is five or less, and BJA discourages both cross-state comparison and year-to-year trend use, which removes the two uses that would create demand for accuracy.

What has been tried?

Two directions were tried in sequence and neither reached the consequence. The collection was rebuilt and then audited, and every audit found the same shortfall. The redesign that would give the audit teeth was requested by the agency that would wield it and has not been enacted.

What was found?

Why is it still unsolved?

Enforcement absent — the duty, the collection, the verification and the audit all exist, and the one consequence attached to them has never been applied in eleven fiscal years.

The lever points at a party that does not hold the records. Congress conditioned state grant money on data held by county sheriffs, municipal jails, police departments and medical examiners, and most states have no law compelling those agencies to report to the state. So the entity that would lose money is frequently not the entity that withheld the data, and the entity that withheld it often receives no JAG passthrough at all and is therefore beyond the reach of the lever entirely. The Department published that analysis in its own voice and asked Congress to redesign the penalty. Congress has not.

Because using the sanction would make the data worse, not using it is a reasoned position rather than neglect. Cutting a grant removes money from the reporting infrastructure the statute wants built, and it falls on states that may be passing along everything they received. Eleven fiscal years of non-use is a decision that has been explained in a report to Congress, and that is the difference between a rule being ignored and a rule being administered exactly as its designers left it. A participant summary from the federal convening of December 2022 records the same judgment from the other side of the table, that the 10 percent reduction would not incentivize improved participation.

With the sanction dormant, compliance collapsed into a test of paperwork, and then the paperwork standard absorbed the shortfall. What is actually checked is whether a state filed a plan and filed its quarterly reports, and every jurisdiction passes both. The guidelines of December 2023 then set the record-completeness bar at 75 percent and made continuous quality improvement a matter of movement rather than of level, with a worked example in which a state flagged on about 40 percent of records should aim for 25 percent. Meanwhile the two halves of the government that could object were separated by design in 2016 — the statistical agency is barred from enforcement and the grants agency disclaims its own output as not official statistics. The product is a number that no part of the government will vouch for and no part of the government will compel.

What observation would mean it is solved?

Candidates — (a) BJA publishes an annual compliance determination naming each state and a consequence lands on at least one that fails (b) the gap between the federal count and independent counts of the same deaths closes (c) an official federal statistic for deaths in state and local custody returns.

(a) alone is weaker than it looks. A determination can be published while the threshold behind it stays at 75 percent, so a state can be named compliant with a quarter of its records missing. The softer action the guidelines authorize is an administrative hold, and no source states whether it has ever been used, so an announced consequence may already have occurred without anything changing on the reporting side.

(b) alone is worse. Every completeness comparison in this record is against another list that is itself acknowledged to be incomplete. The private databases used for the federal cross-check cover police-caused deaths, so a closing gap would say nothing about jails and prisons, which is where the larger share of deaths sits and where the check does not reach. A gap can also close because the comparator got worse.

(c) alone restores a label without restoring leverage. BJS can produce a number and is barred from using it for enforcement, which is precisely why the duty moved in 2016. Counting well and compelling reporting are different powers, and the design placed them in different agencies. The three have to be read together, and (a) only means anything if the threshold behind it is stated alongside the determination.

What is it connected to?

Fills with researchdeaths in immigration detention and in tribal facilities, medical examiner capacity and how a cause of death is determined and recorded, the wider pattern of federal grant conditions used as data levers across other statutes, and custodial death reporting regimes in other countries. Relation type and evidence grade were not confirmed in this round.

What these sources do not say

• No denominator exists anywhere. No source gives the number of deaths that should have been reported in any fiscal year. Every completeness figure in the record is a comparison against another admittedly incomplete list, which is why the GAO figure and the DOJ figure cannot be reconciled and why an upper bound on the shortfall cannot be derived at all.
• No source names a single non-compliant state. The guidelines of December 2023 promise that each state will be notified of its compliance determination by January of each year. No determination for any fiscal year appears on any page opened here, and no page lists compliant or non-compliant jurisdictions.
• Whether the administrative hold has ever been applied. It is the weaker action the guidelines authorize and the one most likely to have been used, and its use or non-use is stated nowhere.
• Where the figure of 75 came from, and why the study due in December 2016 was never delivered. For the threshold there is no rationale, no citation, no comparison to any other federal collection, and no explanation of why it is not 100 when the statute requires every listed element for every death. For the study the Attorney General owed Congress within two years of enactment, nothing opened here says why it was not carried out or whether the government still considers it owed.
• The flag rate BJA computes for itself. The quarterly verification process determines how many records failed and returns that to each state. The published tables carry no count of it. The one routine completeness measure the government produces is the one it does not release.
• Why the path to the underlying records was removed after November 2024. The independent audit reached individual-level records through the source data behind the public dashboard on 2024-11-20, the path was later removed, and the office declined to say whether it intended to leave that access in place.
• Why the BJS series was allowed to end at 2019 rather than run alongside the new collection, which would have preserved a comparable trend line across the handover. No opened source addresses the decision.
• How the three federal counts fit together. The state tables published by BJA, the prison deaths counted by the National Prisoner Statistics program and the federal deaths counted by BJS are three separate records of custodial death held inside one department. No source opened here reconciles them into a single national figure, and none claims that their sum would be one.

See the evidence

Every substantive figure here comes from a government document read directly, and the strongest of them are the government measuring its own shortfall. The statute, the two GAO products, the September 2022 implementation report, the March 2023 report to appropriators, the December 2023 compliance guidelines, the April 2025 reporting guidance and four BJA and BJS program pages were all opened at the URLs listed. The PDFs among them were text-extracted locally because the inline reader could not decode them. The one independent audit is a two-part investigation by a news organization, and it is the only source in this file that examined individual records. Where sources overlap they agree: the non-use of the penalty appears in a DOJ report and in an independent investigation three years apart, and the 75 percent threshold appears in the guidelines and is consistent with the good-faith-effort language GAO found in the earlier internal guidance. Where they disagree the disagreement is left visible rather than resolved — the GAO count of nearly 1,000 unreported deaths across all three categories in FY2021 is smaller than the DOJ count of over 900 unreported prison deaths alone in each of CY2020 and CY2021, and the two are not averaged here because they use different comparators, different periods and different universes; and the two pages of the same investigation give the earlier match rate as 92 percent and as 96 percent without either acknowledging the other. Two rows carry no URL. The Senate staff report and the bill text were both refused with HTTP 403, the rows are kept rather than deleted, and no claim in this dossier depends on either. This is a Path A output, a research-based definition, so observation_refs is empty and provenance_mode: press-derived.

People affected

Estimated range 681–32,000 As of FY2020-FY2024

Sensitivity The width of this interval is not sampling noise. It is the distance between what outsiders could prove one death at a time without access to the underlying federal records, and what the government implies about itself when it compares its own collection to outside lists. The low bound counts only individually documented absences over four of the five fiscal years. The high bound counts every record that is either missing or incomplete, and every step of it extends a single-year rate across five years. Several things the count cannot capture would push it upward: deaths in facilities that have never reported at all cannot be counted because no roster of the facilities that owe reports exists, so no ceiling is derivable; jail and prison deaths missed by every outside list are invisible because the only external cross-check covers police-caused deaths and one of the two databases used includes only deaths from police shootings; and the prison comparator itself excludes people unsentenced or sentenced to one year or less, which falls hardest on the six states with combined jail and prison systems. The limit in the opposite direction is that the upper bound double-counts nothing but assumes stability the sources do not establish, since the number of reporting jurisdictions filing in all three categories rose from 22 in FY2020 to 31 in FY2022 and the completeness rate may have moved with it. Deaths in federal custody, immigration detention and tribal facilities are outside this collection and outside both bounds, and no federal source counts the next of kin, so no multiplier for families is applied here.

Regional breakdown State-level values cannot be recovered from the published tables. BJA masks every value of five or less to protect the confidentiality of individual decedents, so state totals cannot be summed from the public data, and BJA states on the same page that the data cannot necessarily be used to make accurate comparisons among states and that year-to-year comparison is discouraged. The one state-level figure available from any source opened here is that over 450 of the 681 individually documented absences were in a single state, which is a floor for that state and gives no basis for the other 55 reporting jurisdictions. Splitting the national bounds by state population would be proportional allocation, and the shortfall is concentrated in jurisdictions whose local agencies do not report rather than distributed with population, so that split would be wrong in both directions at once.