Predicting from Tables

A two-part series for applied ML engineers and technical leaders: why a model comparison needs more than a score, and how Quibble makes its evidence inspectable.

Same Accuracy, Different Mistakes: the decision problem, changed predictions, and the limits of a small smoke check.
Making Prediction Comparisons Worth Trusting: row identity, source fingerprints, failure/device accounting, and local-data workflows.

The articles explain motivation and engineering decisions. The documentation index remains the reference for current behavior and validation status. Start with Getting started to try the workflow.

Figures and curated evidence

The asset directory contains the original SVG diagrams, PNG exports, a portable Wine smoke table, and its provenance and verification notes. The historical example mixes CPU and CUDA, uses only 32 distinct test rows, and is not a full benchmark, speed ranking, or replacement approval. New GPU-only runs exclude the incompatible CPU methods shown in that historical example.

The evidence files are a curated summary, not a complete reproduction bundle. Original prediction arrays, model weights, environment lock, and local result snapshot are not included. Hashes identify supplied files; they do not reconstruct missing inputs or prove execution independently.

These are public source files in the docs tree, also exported by the static-site build. Private references, editorial notes, and social-platform drafts remain outside this public collection. Source: ushakrishnan/quibble. The site becomes live only after the Pages workflow successfully deploys.