Folkbench unveils relay model-degradation detection via current-events and fingerprint probes

Folkbench's starting point is a gap between two existing kinds of monitoring: leaderboards measure a theoretical ceiling in the official environment, and uptime checks only ask whether an endpoint responds. Neither answers whether reasoning depth, constraint adherence and model-specific behavior stay the same as a direct official connection when a gateway sits between the caller and the model lab. The gateway may be doing only forwarding and billing metering, but cost or congestion controls can produce degradation that communication metrics do not capture.

The design responds to two structural holes in manual checks. Asking a model its own name fails because it does not know the proxy in front of it, and repeating famous fixed quizzes fails because those questions are often in training data and repeated patterns are easy to serve from a static cache or route to a clean inspection-only path. Folkbench therefore updates current-events problems over time so old snapshots break on knowledge cutoffs, and it looks at thinking habits under multi-step constraints and format boundaries rather than only final-answer correctness.

The official direct-connect key is run in the same time window not as a fixed passing score but as that moment's baseline, so provider fine-tuning and ordinary fluctuation are not mistaken for relay degradation. Folkbench presents the diagrams as schematic, not measured data, and frames the output as an early-warning signal rather than proof. Separately, the open-source desktop client Folkbench Switch checks configuration differences for local tools before writing them, while continuous integrity monitoring is a Web-side feature.