Governance evidence
What agencies reported about the OMB M-25-21 minimum practices for high-impact AI. This is a disclosure analysis, not an audit, a compliance finding or an official assessment. An unanswered question is missing evidence, not evidence that a safeguard is absent.
Practice-by-practice evidence
Denominator: 19 records · deployed high-impact use casesReported in place: 16 (84.2%)AI impact assessment
Reported in progress: 3 (15.8%)
| Practice | In scope | In place | In progress | Not in place | N/A | Precluded or waived | Other | No answer | Answered | In place of applicable |
|---|---|---|---|---|---|---|---|---|---|---|
| Pre-deployment testing | 19 | 19 | 0 | 0 | 0 | 0 | 0 | 0 | 100% | 100% of 19 |
| AI impact assessment | 19 | 16 | 3 | 0 | 0 | 0 | 0 | 0 | 100% | 84% of 19 |
| Independent review | 19 | 16 | 3 | 0 | 0 | 0 | 0 | 0 | 100% | 84% of 19 |
| Ongoing monitoring | 19 | 19 | 0 | 0 | 0 | 0 | 0 | 0 | 100% | 100% of 19 |
| Operator training | 19 | 18 | 1 | 0 | 0 | 0 | 0 | 0 | 100% | 95% of 19 |
| Fail-safe | 19 | 16 | 0 | 0 | 3 | 0 | 0 | 0 | 100% | 100% of 16 |
| Appeal process | 19 | 7 | 0 | 0 | 9 | 3 | 0 | 0 | 100% | 100% of 7 |
| User and public consultation | 19 | 6 | 2 | 0 | 0 | 0 | 11 | 0 | 100% | 32% of 19 |
Answered = records with any answer ÷ records in scope. In place of applicable = reported in place ÷ (in scope minus not applicable, precluded and waived). Unanswered records stay in that denominator, so the rate is a lower bound on reported evidence, not an estimate of real practice.
| Testing | Impact assess. | Indep. review | Monitoring | Training | Fail-safe | Appeal | Consultation | n | |
|---|---|---|---|---|---|---|---|---|---|
| DHS | 100% | 91% | 91% | 100% | 91% | 100% | 100% | 45% | 11 |
| SSA | 100% | 100% | 100% | 100% | 100% | 100% | 100% | 0% | 6 |
| EPA | 100% | 0% | 0% | 100% | 100% | 0% | 0% | 0% | 1 |
| STATE | 100% | 0% | 0% | 100% | 100% | 100% | 100% | 100% | 1 |
Hatched cells: the agency answered nothing for that practice, so there is no evidence either way. Compare agencies with care: each agency is one reporting unit, and portfolios differ in size and type.
Records requiring further investigation
3 records with at least one practice unanswered or in progressOpen in explorer| Use case | Stage | Test | AIA | Review | Monitor | Train | Fail-safe | Appeal | Consult |
|---|---|---|---|---|---|---|---|---|---|
| Verification Match Model DHS · DHS-2384 | Deployed | Reported in place | Reported in progress | Reported in progress | Reported in place | Reported in progress | Reported in place | Reported in place | Reported in progress |
| RCRA PA EPA · 78 | Deployed | Reported in place | Reported in progress | Reported in progress | Reported in place | Reported in place | Not applicable | Not applicable | Reported in progress |
| AI-Augmented Declassification Review STATE · R3213 | Deployed | Reported in place | Reported in progress | Reported in progress | Reported in place | Reported in place | Reported in place | Reported in place | Reported in place |
Unanswered analytical questions
Generated from the data; for follow-up with agencies- 3 high-impact use cases are reported as deployed while their AI impact assessment is reported in progress. Was deployment authorised before the assessment was complete?
- 9 high-impact use cases mark an appeal process as not applicable. On what basis was each individual outcome judged not contestable?
- 6 high-impact use cases report using PII but give no link to a Privacy Impact Assessment. Is a PIA published elsewhere?
- 4 governance answers are ambiguous (for example, an answer option from a different question). What did the agency intend?
These are questions the inventory raises but cannot answer. They are not findings of non-compliance.