AN INTERACTIVE MODEL LITERACY LAB
A model is only
as steady as
its reference.
Change the reference window, shift the data, or adjust the review threshold. See which labels move—and which records the model cannot score.
01 / CHANGE AN ASSUMPTION
The controls
A higher threshold generally flags fewer records. It does not prove the model is better.
Uses the first N examples in each route, preserving input order. Fewer than eight cannot support a route score.
Adds minutes to reference transit values only. Candidates stay fixed. This is a controlled distribution-shift experiment.
02 / INSPECT THE CONSEQUENCE
Same records.
Different assumptions.
| Record | Original | Experiment | Distance / threshold |
|---|
Click a record to inspect its nearest reference examples. “Within reference” means within a fitted distance range, not safe, normal or clinically acceptable.
NEIGHBOR EVIDENCE
Select a record
03 / KEEP THE LIMITS VISIBLE
Experiment, then ask better questions.
What the model measures
Three-nearest-neighbor mean distance over transit minutes, queue minutes and missing scans. Features use robust MAD scaling with a floor. The original threshold is the leave-one-out reference 95th percentile. Every route needs at least eight reference examples.
What this cannot tell you
There are no truth labels, so the lab cannot calculate accuracy, precision, recall or patient outcomes. The selected window is order-dependent and may be biased. No automated decision should follow these teaching labels.
New work and prior work
The kNN model comes from Amrit Lahari's Custody Lens, adapted from MED-BLACKBOX. This new Hyperbloom contribution is the parameter experiment, reference-shift simulation, label comparison, evidence view, reproducible export and teaching interface. OpenAI Codex assisted with code, tests and documentation.
Imports are processed in the current tab and never saved by this app. Use synthetic or non-sensitive data only; IDs and routes are not anonymized. Maximum 200 references, 100 candidates, 2 MB.