W-Twin · local monitor

Drop a CSV. See if the curve is
drifting from what's expected.

W-Twin compares your real loss curve against a scaling-law forecast and flags divergence — before it's visible in the raw numbers. Everything runs right here, in your browser.

Drop your CSV here, or click to browse

Expected columns: step, loss — production W-Twin columns are used directly when present

Your file never leaves this browser — no upload, no server.

Steps loaded

—

First flag (raw)

—

First sustained alert

—

Total flagged points

—

Verdict

—

Loss trajectory vs. scaling-law baseline

Observed loss Expected (scaling law) Flagged point Sustained alert onset

W(t) = Q(t)·D(t) − T(t)

W(t) Threshold (W=0)

How to read this: "First flag (raw)" is the first alert reported by the production W-Twin output when such a value exists. It is kept separate from the per-row flag vector. "First sustained alert" is taken directly from the production output when available; otherwise it is derived from five consecutive flagged points. Production W-Twin columns are authoritative and are not replaced by the browser-side approximation. If no production W-Twin outputs exist, the browser performs a local demonstration analysis. This gives a directional read on your specific log, not a guaranteed diagnosis.