Introduction

Scenario: Leadership will not act on a disruption-risk forecast they cannot interrogate, so before any of these scores reach an operator they ask you to prove the model's error is acceptable and to show, visually, which suppliers and time windows carry the most risk. You will quantify prediction error and segment risk in Contour and build a Quiver dashboard that makes predicted disruption risk legible over time and by supplier segment.

This unit turns a held-out model's raw predictions into evidence a decision-maker can trust. By now you have a trained disruption-risk model that was evaluated under a Modeling Objective, but a single headline metric like AUC or RMSE does not tell an operator where the model is wrong or which suppliers to worry about. Your job here is statistical analysis and visualization: compare predicted disruption risk against actuals, quantify the error and where it concentrates, and present the high-risk segments and time trends in a form the distribution team can read at a glance. The deliverable is a Contour analysis that quantifies error and segment risk, plus a Quiver dashboard that visualizes predicted risk over time and by supplier segment.

Contour is Foundry's point-and-click analysis tool that operates on tabular data at scale through a sequence of 'boards' — filter, join, aggregation/group, regression, and verify-results boards — that compose into a reproducible path and roll up into a dashboard. You will build boards that join predictions to actuals, compute residuals (predicted minus actual), group error by supplier segment, and fit a regression best-fit line (linear, polynomial, or exponential) to see how predicted and actual risk relate. Quiver is the companion tool for interactive, visual analytics; its time-series regression card and other cards let you chart predicted disruption risk across time and break it out by segment so a reviewer can spot drift and seasonal spikes without writing code.

A key honesty point runs through this unit: what an automated check can confirm and what it cannot. The prediction-output dataset is fully introspectable — Get Dataset Schema confirms the predicted-score and actual columns exist and Read Table Dataset confirms there are rows to analyze. The Contour analysis and the Quiver dashboard, by contrast, are confirmed existence-only: Filesystem v2 Get By Path returns each artifact's RID and resource type (CONTOUR_ANALYSIS and QUIVER_DASHBOARD), but the boards, computations, residual math, and chart configurations inside them are not exposed by any read API. That means the analytical correctness of your work — whether the residuals are computed right and the high-risk segments are truly the high-risk segments — is a manual, self-attested judgment, and you should treat it with the rigor that no machine check is watching.

Capability focus: Contour statistical analysis; Quiver dashboards; error/segment-risk interpretation; prediction-output querying. · Artifact: A Contour analysis quantifying error/segment risk plus a Quiver dashboard visualizing predicted risk.

Key concepts

Companion video

1. 1

   Prediction-output dataset is queryable for stats

   Before you analyze anything, confirm there is real model output to analyze. The prediction-output dataset is the scored dataset that pairs the model's predicted disruption-risk values with the corresponding actuals; it is the queryable source of truth for every board and card you will build. Verify the dataset resolves and that it has rows — an automated check uses Read Table Dataset to confirm the prediction-output dataset has at least one row, which is the difference between a usable result set and an empty schema. Open the dataset and sanity-check that both the predicted-score column and the actual/ground-truth column are present and populated, because residual analysis is meaningless if either side of the comparison is missing. This is the one part of the unit that is fully machine-verifiable: the dataset's schema and row count are exposed by read APIs, unlike the analysis artifacts you build on top of it.

   not startedinstance check

   Run check

   Confirms the prediction-output dataset has rows for statistical analysis.

2. 2

   Contour analysis artifact exists

   Build the statistical analysis in Contour, Foundry's point-and-click tool for tabular data at scale. Create a Contour analysis over the prediction-output dataset and compose the boards into a reproducible path: join (or pivot) predictions to actuals if they are not already aligned, add an aggregation/group-by board to compute residuals (predicted minus actual) and roll error up by supplier segment, add a regression board to fit a best-fit line (linear, polynomial, or exponential) between predicted and actual risk to expose bias or non-linearity, and add a verify-results board to inspect the rows behind the numbers so the error figures are defensible rather than taken on faith. An automated check confirms this artifact exists by calling Filesystem v2 Get By Path on the analysis path and verifying it returns type CONTOUR_ANALYSIS with a RID — note that this is existence-only: the check cannot see the boards or computations inside, so the burden of getting the residual math and segment grouping right is entirely on you.

   not startedinstance check

   Run check

   Confirms a Contour analysis artifact exists (type CONTOUR_ANALYSIS).

3. 3

   Quiver dashboard/analysis artifact exists

   Build the visualization in Quiver, Foundry's interactive visual-analytics tool, so the risk story is legible to a non-technical reviewer. Create a Quiver dashboard that charts predicted disruption risk over time — use the time-series regression card to plot the trend and surface drift or seasonal spikes — and add cards that break risk out by supplier segment so the highest-risk segments are visually obvious. Mirror what the Contour analysis quantifies: where the Contour boards prove the error and segment risk numerically, the Quiver dashboard makes them readable at a glance for leadership. An automated check confirms the artifact exists by calling Get By Path on the dashboard path and verifying it returns type QUIVER_DASHBOARD (Quiver artifacts may also resolve as QUIVER_ANALYSIS) with a RID. As with Contour, this is existence-only confirmation: the cards, their configuration, and the correctness of what they show are not introspectable by any read API.

   not startedinstance check

   Run check

   Confirms a Quiver dashboard artifact exists (type QUIVER_DASHBOARD).

4. 4

   Analysis interprets error/segment risk correctly

   Finally, attest to the analytical correctness of your work, because no automated check can. Step through the Contour boards and confirm the residuals are computed as predicted minus actual on correctly joined rows, that the regression best-fit is the right model form for the relationship, and that the verify-results board shows the underlying rows match the aggregates you are reporting. Then confirm the Quiver dashboard's time-series and segment cards genuinely surface the high-risk suppliers and time windows rather than artifacts of a mis-scoped filter or an unsorted axis. This step is honestly manual: Get By Path proved the artifacts exist, and Get Dataset Schema and Read Table Dataset proved there is data underneath, but the substance of the analysis — whether the residuals, segments, and trends are interpreted correctly — lives inside the boards and cards that no read API can inspect. Treat your self-attestation as the only line of defense and document the high-risk segments and error bounds you are willing to stand behind.

   not startedself-attested

   Mark complete

   Self-attested: the Contour/Quiver boards correctly surface residual/segment risk.