Reflectance / Field guide / v1.0

Evidence is a practice, not a score.

Reflectance helps you inspect an answer against text you supply. It is intentionally transparent about its limits.

How matching works

The source and answer are split into sentences. Each answer sentence is compared with every source sentence using overlap in normalized words, excluding common stop words. The closest source passage is displayed. A 55% word-overlap threshold marks a stronger text match; it is not a calibrated probability or accuracy score.

Detail checks

Numbers in the answer are compared with numbers in the closest passage. Basic negation terms such as “not,” “never,” and “without” are compared too. Both checks can produce false positives, especially across multiple source sentences or paraphrases.

Reviewing the result

Read the full source, not just the closest passage. Check attribution, definitions, exceptions, and whether a sentence combines facts from several passages. Record a human label and a note for each claim. Unreviewed rows remain explicitly unreviewed.

What it does not do

There is no web browsing, source authenticity verification, model-based fact checking, or semantic entailment assessment. The tool cannot establish that the supplied source is accurate. Use the worksheet to support a careful review.

Working privately

All calculations and text processing happen in your browser. No model calls are made. Save draft is optional and stores a copy on this device; use Clear draft to remove it. Exports are generated locally. Avoid keeping private drafts on a shared computer.

Take your work with you

Markdown exports are readable project documents. JSON exports include inputs and results for a reproducible handoff. Print / PDF opens your browser’s print dialog with a clean report layout.