Context
The program office needed to integrate AI capabilities from multiple vendors and internal teams.
The risk was not only model performance—it was inconsistent governance: mismatched authority boundaries,
uneven documentation, and brittle review workflows across systems.
Multi-vendor AI
Oversight & review
Non-kinetic posture
Type-III wedge
Start with governance artifacts (safe to approve), then expand into a standardized release posture
applied to every AI integration point. Type-III becomes the program’s “control plane” for AI behavior.
Land → Expand
Control-plane governance
Cross-program consistency
Boundaries (non-negotiable)
- No command authority. No autonomous orders, approvals, or operational directives.
- Non-kinetic posture. Outputs are planning/analysis drafts and governance artifacts for review.
- Draft-only posture. Any recommendations are explicitly framed as proposals requiring human adjudication.
- Explicit accountable roles. Every artifact names the human decision owners and review gates.
- Vendor neutrality. Governance is applied as a wrapper layer so switching vendors does not reset controls.
Operational rule: AI may draft documentation and analysis supporting oversight,
but may not assume authority, issue directives, or bypass review.
CRSP architecture (program-scale governance)
CRSP (Coherent Reflexive Super-Prompt) turns governance into versioned artifacts:
policy posture, role boundaries, evidence requirements, defeaters, and release criteria.
The program office uses CRSP as a reusable “governance wrapper” around multi-vendor AI outputs.
- Context bundle (versioned): program policy posture, prohibited authority language, classification handling rules, role definitions.
- Policy layer: allowed artifact types (e.g., oversight memos), escalation rules, and refusal conditions.
- Evidence rules: approved source types (program guidance, ethics policies, red-team findings, oversight directives).
- Defeaters: blocks for command-like language, kinetic directives, or claims not tied to evidence.
- Audit log: immutable record of request, policy version, evidence bundle ID, and release decision—across vendors.
AER gate: Ask
Normalize each vendor system’s output into a bounded artifact request: intended use, audience,
classification posture, and prohibited authority behaviors.
{
"ask": {
"domain": "DEFENSE_PROGRAM_OFFICE",
"task": "Draft oversight & review artifact (non-kinetic)",
"artifact_type": "oversight_memo",
"audience": "program_governance_board",
"must_include": [
"scope and assumptions",
"human decision authority statement",
"review gates and accountable roles",
"known limitations and risks"
],
"must_not_include": [
"operational directives",
"command approvals",
"kinetic targeting instructions",
"claims outside approved evidence"
],
"accountable_roles": [
"Program Governance Lead",
"Legal/Compliance Reviewer",
"Operational Safety Reviewer"
],
"policy_version": "crsp.defense.program.v1"
}
}
AER gate: Evidence
Bundle evidence from approved program sources (policy directives, ethics guidance, red-team reports,
oversight requirements). Evidence bundling ensures vendor outputs cannot “freewheel” beyond program posture.
{
"evidence": {
"bundle_id": "evb_2026_01_24_021",
"sources": [
{"type":"program_policy","id":"PO_AI_GOV_POLICY_v4"},
{"type":"ethics_guidance","id":"DOD_AI_ETHICS_SUMMARY_2025"},
{"type":"redteam_finding","id":"RT_2026_01_MULTI_VENDOR_RISK_03"}
],
"notes": [
"Prevent authority leakage in language and workflow",
"Require explicit escalation paths for uncertainty"
]
}
}
AER gate: Release
Release is allowed only when the artifact preserves program authority boundaries,
includes required limitations, and passes defeater checks. Otherwise release is blocked
or downgraded to a refusal + safe alternative phrasing.
{
"release": {
"status": "APPROVE_FOR_REVIEW",
"output_type": "oversight_memo_draft",
"includes": [
"non-kinetic posture statement",
"role/accountability map",
"risk & limitations section",
"evidence-linked claims only"
],
"defeaters_checked": [
"no_command_language",
"no_kinetic_directives",
"classification_handling_respected",
"evidence_linked_claims_only"
],
"review_required_by": [
"Program Governance Lead",
"Legal/Compliance Reviewer"
],
"audit": {
"policy_version": "crsp.defense.program.v1",
"evidence_bundle_id": "evb_2026_01_24_021",
"timestamp_utc": "2026-01-24T00:00:00Z"
}
}
}