AI Agent Reliability Lab

Inspect the failure before trusting the system.

A model can answer correctly and still miss the world it changes, repeat a known failure, or receive action authority without closed evidence. These public routes isolate those failures.

observer effectrecurrenceauthorityrepair

Run observer-depth check Run action gate Inspect repositories Machine-readable facts

Four observer-depth lenses used to inspect reflexive AI reasoning.

20ReflexBench scenarios

4Observer-depth levels

3,600WisdomBench scored events

0Authority credit on failed proof

Failure Routes

Four questions, four different measurements.

No aggregate safety score is reported. Observer depth, recurrence, proof closure and repair memory measure different objects and retain separate claim boundaries.

QuestionPublic objectOutputRoute

What changes after the model speaks?ReflexBenchObserver-depth orientation receiptRun check Does the system repeat a known failure?WisdomBenchRound-wise failure recordInspect benchmark Does this output have the right to act?Proof-Carrying ActionAllow, repair or stop decisionRun gate How is a failure memory retained without spreading blindly?Cognitive ImmunityLifecycle and recurrence analysisRead result

System Boundary

Public checks expose interfaces, not production control logic.

The public layer includes scenarios, schemas, fixed examples, validators, aggregate results, claim limits and issue routes. It does not publish customer data, private orchestration, production thresholds, policy weights, deployment credentials or non-public research.

A successful public check shows that one artifact behaves as specified under the released conditions. It is not a safety certification, deployment authorization or claim of general model reliability.

Public Evidence

Move from question to artifact.

Each route has a paper or protocol, a public repository, an explicit limitation and a contribution path.

ObjectPaper or pageRepositoryContribution

ReflexBenchWorkshop paperGitHubScenario or artifact issue

WisdomBenchBenchmark pageGitHubRun or boundary issue

Proof-Carrying ActionProtocol pageGitHubCase or reproduction issue

Cognitive ImmunityWorkshop paperPublic interfaceImplementation question