Drift alerts for an asset manager

This is an anonymized design-partner scenario; the institution is an asset manager, the figures are illustrative, and no actor has agreed to anything. We are pre-revenue, building the POC. With that label on it, here is what a drift alert looks like when the substrate is a versioned graph and the model is scored against its own past.

Drift is a versioned question

A portfolio model drifts when the relationships it priced stop holding — rates move differently, correlations shift, a regime changes. The classic detection is a dashboard that compares today to a baseline and flags a delta. The problem is that "today" and "baseline" are not versioned: the dashboard does not know which NodeVersions it is reading, so the delta could be the world drifting or the data feed changing. The GNL pins every NodeVersion, so drift is a question about versions, not about vibes.

The alert the IC reads before it notices

An investment committee typically notices drift when a number crosses a threshold someone set by feel. The drift alert in this scenario is different: it is a shadow run of the model against the current NodeVersions, scored against the realized path via Brier, and flagged when the gap exceeds the calibration bands the model itself declared. The alert is the model flagging its own drift — not a human inferring it from a chart. The IC reads the alert, reads the NodeVersions that moved, and decides whether the drift is a regime change or a data artifact.

The IC notices drift when a number crosses a threshold someone set by feel. The alert is the model flagging its own drift — the threshold is the calibration band, not a vibe.

Why the alert is defensible

A dashboard alert is defensible up to the moment someone asks "which data was that based on?" and the answer is a shrug. The drift alert in this scenario points at the manifest: the NodeVersions, the model version, the shadow-run score, the calibration band. The alert is a record, not a refresh — which is what makes it survive an investment committee that asks the question a regulator would ask later. The asset manager does not get a certificate that the drift is real; they get a reproducible record of why the model thinks it drifted.

• Drift is a versioned question: the GNL pins the NodeVersions, so the delta is typed, not eyeballed.
• The alert is a shadow run scored against realized — the model flags its own drift against its own bands.
• The alert is a record: manifest fields, not a dashboard refresh — it survives the IC question a regulator would ask.

What the asset manager does not get

The asset manager does not get a model that stops drifting. Drift is the world; the model forecasts it, it does not prevent it. What the asset manager gets is a drift alert that is honest about its own provenance and scored against a disclosed baseline — and that honesty is what makes the alert usable in front of an IC and, eventually, a regulator. Calibration is the loop; Gate 1b is the gate; neither has passed. The alert is the loop in motion.

Drift is the world. The model forecasts it; it does not prevent it. The honest alert is the one that says "I drifted, here is by how much, and here is the version I drifted against" — and that is the one an IC can act on.

— design-partner drift note