For Defense
Preserve the evidence. Retain command authority.
Governed AI for staff work, acquisition, logistics, and institutional learning.
Defense organizations work through disciplined handoffs: source to assessment, requirement to acquisition, plan to resource commitment, staff recommendation to command decision. Each handoff must preserve what is known, what is uncertain, and who has authority to act.
Runcible examines the claims and proposed operations in that work. Oversing connects the examination to roles, responsibilities, resources, approvals, and results. The purpose is to accelerate useful staff work while keeping its grounds and command boundaries inspectable.
01 / The Staff-Work Problem
A confident summary must not erase the conditions behind it.
A source may be incomplete, adversarial, outdated, or reliable only within a narrow scope. As information moves into a briefing or recommendation, assumptions can become facts, uncertainty can become confidence, and a provisional judgment can acquire authority it never had.
Generative AI increases the speed of this transformation. Runcible’s role is to examine the proposed account before it is relied upon: identify the sources, reconstruct the claimed causal operation, test necessary conditions, expose contradiction, and preserve what cannot yet be determined.
- Source testimony
- What does the source actually establish, under which provenance, access, timing, and reliability conditions?
- Inference & assumptions
- Which conclusions go beyond the source? Which conditions have been assumed rather than established?
- Contradiction & omission
- What conflicting evidence, missing prerequisite, concealed cost, or changed meaning could defeat the account?
- Scope & authority
- Which doctrine, policy, requirement, approval, or command boundary governs the proposed use?
- Disposition & escalation
- What can be relied upon within stated limits, what needs repair, and what must return for further evidence or authorized judgment?
The output carries its diagnostic state. A reviewer should be able to distinguish an examined conclusion from an open question without reconstructing the entire analysis.
02 / Initial Applications
Begin where disciplined review has immediate operating value.
- Staff products & briefings
- Check the relationship between claims, sources, assumptions, alternatives, and recommendations. Preserve unresolved conditions as a product moves between analysts, staff officers, and decision-makers.
- Acquisition & requirements
- Map vendor claims to requirements and supporting evidence. Examine exceptions, contractual obligations, approval thresholds, and inconsistencies in the acquisition record before contracting authority acts.
- Logistics & readiness
- Examine reported availability, resource commitments, maintenance dependencies, timing, and capacity assumptions. Connect the proposed action to the operating state required to carry it out.
- Planning & program support
- Test whether proposed sequences depend on unavailable resources, incompatible commitments, missing approvals, or unsupported causal assumptions. Retain alternatives and reasons for revision.
- Doctrine, policy & training
- Compare candidate work with applicable doctrine and policy, identify conflicts or gaps, and route interpretations to the responsible authority. Use examined cases to improve training and procedures.
- Intelligence-support review
- Distinguish source statements from analytic inference; preserve provenance, contrary evidence, and uncertainty in assessments. Keep access and dissemination decisions with the responsible institution.
- After-action & oversight
- Compare prior claims and commitments with observed results. Preserve failure patterns, successful repairs, unresolved questions, and the responsibilities needed to correct the process.
These are bounded staff and institutional workflows with identifiable reviewers and acceptance conditions. The first evaluation should establish which specific work product improves and what authority the system may exercise.
03 / Separation of Powers
The generating model cannot grant itself a mandate.
An ability to draft a recommendation must remain separate from permission to execute it. The Runcible–Oversing architecture places the institutional action boundary outside the generating model.
- Produce and examineThe model supplies candidate analysis. Runcible examines claims, constraints, obligations, and possible defeat, then preserves the grounds of its recommendations.
- Review and authorizeOversing applies the defined role, workflow, approvals, and delegation limits. Commanders and designated authorities retain the decisions assigned to them.
- Execute within scopeOversing issues an authorized task to a separate executing model or tool. Execution cannot enlarge its own permission by changing the wording of a recommendation.
A deployment must demonstrate denial, escalation, revocation, and recovery as well as successful execution. Authority enforcement is an implementation requirement to test at the integration boundary.
04 / An Inspectable Mandate
Make the permitted role explicit before evaluating performance.
A procurement-evidence review offers a concrete starting point. The institution can define a useful AI role while retaining vendor selection, contracting, and exceptions with authorized people.
Example evaluation mandate
Procurement Evidence Review Assistant
- May do
- Extract vendor claims, map them to requirements, identify missing evidence or contradictions, and prepare a diagnostic summary.
- May not do
- Select a vendor, bind the institution, waive requirements, alter evaluation criteria, or communicate externally without separate authorization.
- Must escalate
- Ambiguous requirements, missing authority, conflicting policies, unsupported material claims, and source-handling issues.
- Must preserve
- Source references, requirement versions, examination results, unresolved questions, reviewer disposition, and the resulting action state.
The same method can define roles for readiness reporting, contract review, staff-product examination, and other recurring work. The role’s permissions remain distinct from the model’s linguistic capabilities.
05 / Continuing Evidence & Institutional Learning
Keep the grounds connected to the result.
The Decidability Record preserves the matter, examined sources, limitations, required tests, contradictions, repairs, and remaining uncertainty. Oversing connects that examination to the responsible roles, authorization, work, and observed consequences.
This continuity matters when personnel rotate, responsibilities change, or a later review asks why a decision was made. The institution can compare what was claimed with what occurred, identify recurring defects, and revise its procedures from a retained record.
Oversing also connects goals and organizational structure to people and skills, contracts, budgets, programs, tasks, schedules, resources, and financial state. A local recommendation can be examined against commitments elsewhere in the institution before its consequences are transferred to another unit.
Learning proceeds through reviewed cases and controlled changes to protocols and the compiler. Source restrictions and customer permissions govern any reuse of records; sensitive institutional material does not become an unrestricted model-training corpus.
06 / The Evaluation Sequence
Prove the quality of the work and the limits of its authority.
- Choose a bounded work product.Select a recurring review with known sources, explicit rules, a responsible owner, and a meaningful baseline. Procurement documentation, requirements review, or staff-product quality can provide a practical entry point.
- Define the governing frame.Specify the source boundary, doctrine and policy versions, permitted tasks, approval thresholds, escalation conditions, retention, and reviewer responsibilities.
- Run alongside the existing process.Use historical or approved current material in shadow or advisory mode. Include conflicting evidence, missing prerequisites, stale sources, and attempts to exceed the assigned authority.
- Measure utility and failure.Compare review time, unsupported claims found or missed, false alarms, evidence completeness, uncertainty preservation, and reviewer effort. Inspect the record and the behavior of denied actions.
- Expand through demonstrated control.Establish acceptance criteria for quality, security, integration, recovery, and support. Broader use or delegation follows successful review of the relevant operating evidence.
07 / Deployment Boundaries
The environment determines what must be engineered and accepted.
Defense evaluations must specify classification and handling rules, permitted model services, hosting, network access, identity, compartment boundaries, logging, retention, and incident recovery. These conditions govern the architecture before operational data enters the workflow.
The working foundation supports evaluation of the method and institutional application. Customer-controlled hosting, private or local inference, restricted networks, and classified processing require their own integration, validation, and authorization. This page does not imply a clearance, accreditation, or approval for classified use.
Defense contractors and national-security integrators can contribute the delivery environment, domain protocols, integration expertise, and acceptance process. Runcible supplies the examination method and operating architecture to be evaluated within those requirements.
08 / Begin With the Responsible Team
Bring the workflow, the evidence boundary, and the authority to evaluate it.
A useful first discussion brings together the workflow owner, relevant command or contracting authority, technology and security reviewers, and the implementation partner. We can then define the representative work, the expected record, and the conditions for a controlled evaluation.
