Intelligence should answer to evidence—and to people. Our five-stage design asks what was observed, whether it holds up, what it supports and who permits the next step. These are requirements we are building toward, not proof that every safeguard is deployed.
Domain-specific adapters translate local observations and rules. They may carry structured context and evidence references, but permissions, authorization leases and security authority do not transfer between domains.
The aim is simple: know the source, test alternative explanations and keep permission separate from persuasion. Independent checks must support consequential claims. Approval must be current, limited in scope and withdrawable before an action—not after it.
EVIDENCE → QUALIFICATION → REALITY → AUTHORITY → EFFECT
What was observed?
Can we rely on it here?
What is supported?
What is permitted now?
What happened within bounds?
Record what was observed, where it came from, when it was collected and what it cannot tell us. A statement without a traceable source is not established evidence.
Example: an Android permission-change record, not a claim that the app is malicious.
Check source integrity, relevance, freshness and permitted use. Missing, stale or unverified evidence never counts as PASS; disagreement stays HOLD for human review.
Example: an old scan may describe an earlier state, but cannot establish the state of the phone now.
Distinguish measured reality from hypotheses. Advocate makes the strongest supported case; Challenger tests alternatives; Arbiter checks whether requirements are met. Independent witnesses must support consequential claims; three roles alone do not establish independence.
Example: compare an expected app update with a suspicious change. If neither account is supported, keep the conclusion unknown.
Check domain permissions, scope, fresh evidence, human approval where required, and a valid time-limited authorization lease. Expiry or revocation stops future permission. Recheck these gates at effect time; historical evidence is not present authorization.
Example: yesterday’s approval does not authorize a changed action today. Human control remains the boundary for irreversible actions.
Only after the earlier gates pass may an authorized, bounded effect proceed. Record its outcome separately from the proposal; verify that outcome through evidence beyond the system’s own assertion.
Example: record the result of an approved action and independently inspect its scope. This website performs no security action.
Result → Authority → Truth → Admissibility → Source
Start from the result and reconstruct its permission, supported truth, admissible evidence and source. Here, “Truth” means the claim supported by evidence—not certainty by declaration. Reverse review needs independent witnesses; replaying our own conclusion is not verification.
Advocate presents a supported interpretation. Challenger looks for counterexamples and competing causes. Arbiter checks the declared requirements against evidence, not persuasiveness. Shared models, data or incentives can create correlated agreement; role labels alone do not produce independent witnesses.
A review should record source dependencies, conflicts, unresolved objections and the external checks still needed. AUREX can test this method; it cannot certify itself or grant product authority.
Connected systems depend on sources, infrastructure and permissions that can change. The design requirement is to retain inspectable evidence, expose missing dependencies and stop proposed effects when freshness, scope or authorization cannot be established.
Expiry and revocation must be checked again before a bounded effect. Test outages, stale sources and withdrawn approval explicitly. Auditability is a requirement to demonstrate—not a claim that these safeguards are already deployed.
A past observation stays in the record. A present action needs current, scoped authorization. Missing, stale or unverified evidence never counts as PASS. UNKNOWN ≠ SAFE.
Independent support must be demonstrated, not assumed from agreement between models. Unresolved checks mean HOLD → human review. ORB MOTION AUTHORITY = NONE; colors, wheels and zoom provide presentation only.
DESIGN INTENT / DEVELOPMENTAL · This public explanation does not establish operational gates, independently verified product performance or autonomous authority.
Bounded analytical roles can propose work, compare alternatives and assemble evidence. Their proposals must face separate checks; agreement alone is not proof.
DESIGN INTENT / DEVELOPMENTAL
Challenge assumptions and reconstruct a proposed outcome from its supporting observations. Independence must be demonstrated, not inferred from multiple agents.
DESIGN INTENT / DEVELOPMENTAL
Use reproducible test cases, failure analysis and abstention to expose limits. A self-test is useful evidence, never the system’s sole witness.
DESIGN INTENT / DEVELOPMENTAL
Keep source identity, scope, freshness and limitations attached to a claim and the change it informs. Missing lineage holds the decision.
DESIGN INTENT / DEVELOPMENTAL
Separate unknown from safe, unsupported from established, and stale from current. Contradictions and unresolved checks remain visible for human review.
DESIGN INTENT / DEVELOPMENTAL
Explore evidence-backed requirements, proposed changes, test plans and review summaries. Deterministic checks and human approval remain boundaries for integration and release.
DESIGN INTENT / DEVELOPMENTAL
Evidence
Qualification
Reality
Authority
Effect
Forward evidence
Observation → proposed consequenceIndependent reverse reconstruction
Proposed consequence → supporting evidenceWitness ≠ Truth · Witness ≠ Authority. No effect without revalidation. The system may never be its own sole witness.
Provenance-bearing observations retain their source, timing, scope and limits. They are not conclusions.
BCXMET is our Android security focus, under development; XAELI is its planned explanation interface. Industrial and capital applications remain separate research directions, not connected production products.
Review research limitsFor technical evaluation, commercialization and strategic partnership conversations—without treating a demonstration as qualification.