Live demo

The review viewer with ~30 preloaded synthetic errors — the same tool we use for human verification of generated data. Every example carries its handler type, §-level schema tag, and applicability rating; the annotation buttons (or keys ++1++–++4++) record a verdict per example.

Open the demo full-screen ↗ — or try it right here:

Notes:

• The examples are unfiltered generator output over news text — judging a few yourself is the fastest way to understand what synterr produces.
• Annotations save to your browser's localStorage only; nothing leaves the page. Export/Import round-trips them as JSON.
• The sample data is also downloadable: sample.jsonl. The viewer itself ships in the repo as tools/diff_viewer.html — drop any of your own generated JSONL onto it.

Quality

Every handler has gone through three independent review passes: an internal audit against the underlying grammar reference (June, 73 findings, all fixed); a native-speaker annotator's pass using this same viewer (2,724 items — 98.4% corruption validity, 91.6% intended-type precision); and an external model-based adversarial audit (July, 47/51 findings confirmed, all fixed). Together they're the basis for treating synterr's output as trustworthy training and evaluation signal, not just plausible-looking noise.