oracle

Oracle — the deterministic plane

The instinct is to start with the screenshot judge, because that is the part that looks like verification. Start here instead.

The worst thing a data-dense product can do is not a misaligned button; it is a well-rendered page stating a figure no source supports. A vision judge cannot catch that, because nothing on the screen looks wrong, and the best published labelled UI-display-defect detectors reach about 85% precision and 84% recall anyway (C17) — a ceiling that says nothing about unsupported numbers, which were never a visual defect. That class is closable by arithmetic (I7), and this skill is the arithmetic.

Running as a Gemini model? Read gemini.md in this directory first, then follow this file with the overrides it names. It routes step 1's template markup pass to another model, gives the four-attribute sweep a per-surface denominator, and turns every coverage claim into a pasted exit code. Other models skip it.

Procedure

1. Mark up the surface once. Four attributes, and they are a change to the product rather than to the pipeline:

   Attribute	On every displayed figure
   data-figure-id	names the figure, and is what a mismatch report cites
   data-source-ref	names the record it came from. A figure without one is the defect this plane exists to find
   data-source-field	names the field within that record
   data-source-expr	for a derived figure, the expression instead of a field: sum(segments.amount) or count(segments)

   The expression form is what catches a total that no longer equals the sum of its parts, which is the most common wrong-number-on-a-right-looking-page defect there is.

2. Extract and gate the lineage.

   python3 scripts/lineage_extract.py --root <repo> --input <rendered.html> --json
   python3 scripts/lineage_gate.py --root <repo> --input <rendered.html>
   python3 scripts/lineage_gate.py --root <repo> --glob 'dist/**/*.html'   # a whole surface set