Cubie Watch
1 · Choose data
2 · Confirm
Live runtime
Watch the detector work in real time as a stream is replayed, with a Prometheus /metrics endpoint and a live Apache-ECharts dashboard — the permissive, insert-at-runtime replacement for Grafana (which is AGPL, and banned here).
It runs as a self-contained local binary (no third-party deps). Point it at any log:
# then open http://127.0.0.1:8788 (Prometheus at /metrics)
It calibrates on the baseline prefix, replays the log on a clock maintaining a live CUSUM channel per stream, and exposes cubie_cusum_accumulator, cubie_threshold_h, cubie_fires_total, … for any Prometheus scraper. The live page draws each stream's accumulator climbing toward its threshold and firing in real time.
Hosting the live surface on this deployment (streaming a live feed rather than a one-shot upload) is the next step; the one-shot Detect tab already gives you the full charts on any file, in-browser.
About & method
Cubie Watch implements the Page (1954) CUSUM recurrence over numeric streams extracted from a text log. The same Rust implementation runs in the CLI and in browser WebAssembly; the implementation-to-proof relationship is documented, not a formal code-extraction theorem.
k = mean(|z|)+0.25σ. Pass B replays the quantized Page accumulator and selects h = max(ceil(1.5·peak(s)), k+1), strictly above the prefix peak.Honest boundary
A fire means sustained deviation from the calibrated baseline on that stream. The machine localizes which stream deviated and when — it does not assign domain meaning. The caller-designated prefix selects both k and h; it is calibration data, not held-out evaluation or an out-of-sample generalization claim.
Source of truth for the theorems and their honest status: the CUB registry.