Policy is explicit, not implicit
Thresholds decide when the kernel is allowed to stay quiet. They are shown here so a reader can disagree with them.
Inference backend
Local ONNX is the default path. The Hugging Face Inference Endpoint is optional and only used once it is configured and answering.
No endpoint is provisioned yet. Latency and error rate stay empty rather than showing a fabricated baseline — the local runtime's own timings are reported in the model version panel.
Requests are queued at most 2 at a time with a 15s wait deadline and a 20s request deadline. Anything past those deadlines, or past the queue limit, is scored locally instead — never dropped, never left hanging.
Kernel thresholds
Preview build: adjustments are local to this session and are not persisted.
below this, a typed signal is treated as ordinary variation
above this, the unknown channel forces an information request
scenarios below this are generated but not surfaced
above this, model validity degrades to fragile
Admission control policy
- Ignore — variation is inside the reference envelope.
- Remember as exception — real but non-structural; stored in surprise memory.
- Open new regime — the reference set no longer describes the system.
- Request more information — the unknown channel dominates the signal.
- Escalate for human review — impact is high and credibility is contested.
Provenance
All numbers in this preview come from deterministic synthetic reference models running in the browser. No trained forecast, live feed or proprietary dataset is used. Domain packs loaded: Earth, Finance, Critical.