Why we audit our own analogies
Every Tensorcraft lesson anchors a new ML concept to a frontend primitive you already know. Analogies are load-bearing here, which means a bad one teaches a wrong mental model. So we classify each bridge by how far it holds:
- identity: the same thing under a different name (npm install vs loading TFJS).
- structural: the shape matches, the internals differ. The lesson must say where it bends.
- intuition: a feel-level pointer only. The lesson must repair it on the page.
The numbers (audit of 2026-04-25)
Across 296 authored lessons at audit time: 5.1% identity, 84.1% structural, 10.5% intuition, and exactly one lesson missing a bridge (since fixed). Tier mismatches between what a lesson claims and what the canonical registry says: zero.
The structural tier dominating is by design. Identity bridges are rare because ML mostly is not frontend engineering in disguise. Intuition bridges are rationed because they decay fastest. When a lesson uses one, the correction is part of the lesson, not a footnote.
What changed since
The audit surfaced 293 lesson-invented analogies that were not in the canonical registry. The reusable ones were promoted; the rest were kept lesson-local on purpose, and every Deep Orbit lesson now carries a labeled bridge with its tier shown in the lesson header. Where an analogy breaks, the break is named in the text. That habit came directly out of this audit.
If you find a bridge that oversells, email support@tensorcraft.app. We would rather publish the correction than defend the analogy.