Technical reference

Raptor as a Compliance-Evidence Substrate for OMB M-25-21

Agencies are in continuous compliance posture: high-impact AI use cases must demonstrate minimum risk management practices when reviewed, inventoried, or challenged. This paper maps Raptor's architecture to each of the seven practices — what it produces, what it partially addresses, and what remains your organization's responsibility.

Executive summary

The compliance evidence gap is architectural, not procedural.

OMB M-25-21 establishes seven minimum risk management practices for high-impact AI. The practices presume a governance layer — something that produces evidence that the process was controlled, the output was classified, and the action was authorized. Probabilistic-everywhere AI architectures do not provide this layer.

Raptor is not a compliance product. It is a compliance-evidence substrate. Its architecture separates the AI's proposal (probabilistic, upstream, unrestricted) from the system's commit (deterministic, downstream, governed). This separation produces the artifacts M-25-21 requires as a byproduct of every governed execution, not as a separate compliance workflow.

Practice-by-practice mapping

Four of the seven practices, addressed at the substrate.

01 Pre-deployment testing

The cross-model evaluation harness produces immutable, replayable test results across providers. Testing evidence is recoverable on demand, not reconstructed.

Status: Implemented

02 AI impact assessment

Trust-boundary classification and provenance chains supply the structural inputs an impact assessment is built from — intent lineage, evidential basis per segment, decision-time risk records.

Status: Implemented

03 Ongoing monitoring

Every governed response stored immutably with trust-boundary tags. The execution history is the monitoring record — queryable by time, trust distribution, or correlation ID.

Status: Implemented (dashboards in development)

05 Human oversight

The confirmation gate halts every action with side effects until an authorized human approves. Each decision recorded immutably with actor, reason, and a policy snapshot.

Status: Implemented

Practices 4, 6, and 7 (training, remedies/appeals, public feedback) are organizational processes. Raptor provides supporting infrastructure — the trust-boundary taxonomy as a training framework, deterministic replay for remediation, and the correction feedback loop as a technical foundation — but does not substitute for the agency process itself. The full paper states this plainly.
What the paper covers

Nine sections. Audit-grade throughout.

Section What it covers
1. The compliance evidence gap What M-25-21 requires, what probabilistic architectures can't produce, why observability and evals don't close the gap
2. Raptor's architecture Proposes/disposes model, execution events, trust boundaries, confirmation gate, multi-provider substrate
3. Practice-by-practice mapping All seven M-25-21 practices mapped to architecture, evidence produced, and gaps stated
4. M-25-22 & M-26-04 Vendor lock-in, data rights, US-produced AI, unbiased AI principles
5. What Raptor does not solve Probabilistic proposal layer, organizational practices, replay vs. re-inference, what's not built yet
6. Verification matrix Claim-by-claim status: implemented, validated, queryable, roadmap, requires agency process
7. Buyer risk assessment FedRAMP, data exposure, integration completeness, single-founder continuity — with mitigations
8. Evidence artifacts Named deliverables agencies receive from governed execution
9. Action layer Specific next steps for CAIOs, contracting officers, and prime compliance leads
Cross-references

Also addresses M-25-22 and M-26-04.

Raptor's multi-provider substrate and open-weight bridge address M-25-22 vendor lock-in protections. Execution data stays in agency-controlled infrastructure, addressing government data-rights requirements. The platform is US-built and SDVOSB-certified.

For M-26-04, Raptor does not certify neutrality. It provides provenance, replay, and audit evidence that can support neutrality reviews.

Next step

Start with the paper. Then start a conversation.

The full paper is 18 pages — 5 minutes for the thesis, 30 minutes for the complete mapping with verification matrix and risk assessment. The first conversation is technical, not sales — you'll talk to Paul James III, a 20-year DISA veteran who built the substrate.

SDVOSB · CAGE 9M2N1 · UEI UJ5VCLDTKK87 · NAICS 541512 / 541511 / 511210