Separate the proposal
from the evidence.

Four kinds of statement

Policy commitment

A proposed choice about authority, rights, public purpose, or spending. It should be assessed for legitimacy, feasibility, trade-offs, and practical implementation.

Hypothesis

A conditional explanation of how a benefit might occur. Every source connection in this library remains at this stage.

Context and interpretation

A description of public records or a comparison with other ideas. It should preserve dates, sources, and the boundary between documented statements and editorial inference.

Empirical finding

A result from an appropriate evaluation, with methods, comparison, uncertainty, and limitations. The supplied MAR material contains no completed findings.

These categories should remain distinct in public communication. A policy can reflect a legitimate preference while its expected technical benefit remains uncertain.

Design a comparison that can change the decision

Begin with the decision that the evidence is meant to inform. Define the service, the disruption, the participants, the information, and the proposed mechanism. Explain what would happen under current arrangements.

Compare the proposed process with current practice and, where feasible, a simpler non-AI improvement. Identify differences in staff, training, communications, and governance so that a favorable result is not automatically credited to a model.

Specify outcomes, interpretation criteria, failure conditions, and the measurement period before testing. Publish a versioned plan and record justified deviations. Reviewers should be able to identify a result that would substantially weaken the claim.

Measure benefits and harms together

• Civilian service availability or recovery, appropriate to the particular use case.
• False warnings, missed warnings, verification burden, and unnecessary action.
• Privacy, civilian protection, unequal access, and limits on meaningful challenge.
• Local fallback, revocation, provider dependence, and the practical cost of withdrawal.
• Full lifecycle cost, staff burden, maintenance, and the cost of independent scrutiny.

Use the right strength of claim

A favorable exercise result supports a claim about the conditions studied. It does not automatically establish real-world readiness, effects in other countries, strategic deterrence, or a reduction in conflict.

Report uncertainty, missing information, and alternative explanations. An inconclusive test should remain inconclusive. Simulated behavior and tabletop choices should not be presented as predictions about particular governments.

A forecasting claim would require an appropriate model, historical validation, and documented calibration. The current site supplies none of these and does not assign probabilities or produce country forecasts.

What a reproducible lab selection records

The lab export includes the software and library edition, selected concept IDs, exact pairwise source records, editorial safeguards and test proposals, context assumptions, export time, and the selection URL.

The three context modes describe assumptions to inspect: verified cooperation, fragmented baseline, and low trust. They do not alter the source’s evidence status or generate conditional outcome scores.

A selection of several strategies produces separate pairwise records and an explicit null combined outcome. The available source cannot establish whether their effects reinforce, offset, or interact with one another.

Independent review would also need access to the evaluation methods, appropriately protected inputs, changes, and results of any future exercise. A lab selection is an aid to framing that work, not a substitute for it.

Requirements for future event-based analysis

This edition has no live-event feature. An event-based extension would need user-selected, separately sourced and dated snapshots, provenance, revision history, and independent verification.

The record would need to separate observations, disputed accounts, assumptions, and interpretation. Changes to the evidence should lead to a visible revision rather than an unexplained change in a result.

Current-event relevance would not by itself justify new information access, country scoring, or operational advice. Any new feature would require its own purpose, evidence standard, and safeguards.

Reference frameworks and their limits

These primary sources provide general background. They do not validate MAR, endorse Safe Nations, or establish compliance by a proposed implementation.

The original MAR matrix is described on the source-provenance page. Dated public records informing the policy principles are cited separately on the policy principles and sources page.

Reference pages reviewed 5 November 2026. No reference supplies an empirical MAR effect estimate.