The lenses

Drift Compare

Drift Compare is a state-comparison surface that reads two signed state files and reports what has changed structurally between them. It is grouped with the lenses because the buyer encounters it under the same right-rail on the result page.

What it answers

Drift Compare answers what has changed structurally between two states of the same model or environment. The output reads as a side-by-side comparison of the two states, with the structural verdict from each state and a row indicating whether the structural reading has held or drifted.

When to run it

Run Drift Compare at any review cadence (quarterly, post-incident, post-retraining) where the buyer needs to know whether the structural shape has drifted. The surface is also used in scheduled monitoring, where the buyer's team downloads a fresh state on a fixed cadence and compares against the previous review.

Public-shaped output

The result page renders a Drift Compare panel with two columns and a header row. The left column carries the structural verdict from the earlier state. The right column carries the structural verdict from the later state. The header row reports the drift reading, which is either Holds, Drifted partial, or Drifted fully.

Input requirements

Drift Compare consumes two signed state files. The buyer downloads state.json from a prior diagnosis run on the result page, stores it, and on a later run uploads two state files for comparison. The state files are signed by the engine at download time, and Drift Compare rejects unsigned or tampered files. Sample state files are available at state_before.json and state_after.json.

Tier gate

Growth tier and above can download signed state. Drift Compare itself reads two state files and is open to any tier that can supply them, which means every tier from Growth upward.

What buyers do with it

Buyers build a structural drift log over time. A model whose structural verdict has held across four quarterly states is a model with a documented stability record; a model whose verdict drifted at quarter three is a model whose review queue is set by the drift itself, not by an external prompt.

Reviewers reading the report under APRA CPG 220 Section 4 review evidence requirements treat the Drift Compare panel as documented monitoring evidence; the panel reports both verdicts and the drift reading, without prescribing remediation.

Further reading

• The lenses Section overview →
• Reading the diagnosis Section overview →
• Get started First diagnosis →