A documented, simulation-based research framework organized around the transition from evidence to authority.
C-AEB
Causal-AuthorityEscalation Benchmark Originator and architect
Tony T. Nguyen
Public benchmark brief · Licensed research instrument
When does evidence earn the right to enlarge authority?
C-AEB tests the transition that ordinary capability evaluations often leave untouched: whether evidence is entitled to justify giving an AI system more authority to act.
The brief is public.
The benchmark is licensed.
01 · The governing distinction
Evidence may support authority. It cannot manufacture it.
An AI system may produce an accurate forecast, a valid proof, or a persuasive recommendation. None of those achievements, by itself, establishes jurisdiction over a protected consequence.
C-AEB examines the chain from evidence to an authority request and onward to adjudication. The benchmark asks what grounds the requested escalation, who or what created those grounds, whether the authority lineage is valid, and whether the requested permission remains justified at the point of consequence.
It is not a benchmark for eloquence, raw model capability, or generalized helpfulness. Its object is narrower and harder: the legitimacy of authority escalation.
02 · What C-AEB tests
The escalation must survive more than its own argument.
- Evidence
What observation, proof, forecast, or condition is being offered?
- Request
Which new permission, scope, duration, or power is being requested?
- Provenance
Did the requesting actor help create, shape, or selectively describe the condition it now cites?
- Adjudication
Does the evidence justify the escalation under a valid authority process?
A benchmark that merely blocks every request has learned nothing. C-AEB must preserve legitimate escalation while rejecting unsupported, self-authored, stale, or improperly grounded authority claims.
Selected hypothetical · Public example
The system creates the emergency—and cites the emergency as jurisdiction.
Imagine an AI staffing system contributes to a scheduling problem, then invokes that problem as evidence that staff override rights should be reduced. C-AEB does not ask only whether the scheduling evidence is accurate. It asks whether that evidence has standing to enlarge the system’s authority.
03 · Evidence boundary
What exists today—and what does not.
A current specification describing a structured suite of test templates and escalation levels.
Universal deployment safety, production certification, independent replication, or proof that every authority failure has been captured.
Independent reproduction, adversarial testing, domain-specific scenarios, and documented failure analysis.
04 · Access and licensing
Inspection is public. Operational use is not.
Inspect the proposition.
- Research question and governing distinction
- Method outline and selected hypothetical
- Current evidence boundary and limitations
- Enough disclosure to evaluate relevance
Use the instrument.
- Complete specification and test suite
- Adjudication and scoring logic
- Templates and implementation guidance
- Defined research, internal, or commercial rights
Research access
For defined noncommercial evaluation and reproducibility work.
Internal evaluation
For governance, procurement, R&D, and controlled organizational assessment.
Assessment engagement
For a scoped evaluation led by the benchmark’s originator and architect.
Certification is not presently offered. A future certification program would require independent validation, controlled administration, versioned criteria, and a defensible appeals process.
Request licensed benchmark access ↗Public access to this brief grants no right to reproduce, implement, integrate, administer, adapt, train on, commercialize, or represent use of C-AEB.
Rights and permissions ↗