A QUIETER WAY TO CATCH UP
Spot the repeat.
Keep the change.
Two versions of the same update? Let a small local model suggest what sounds familiar. Keep every original line, especially when the details change.
Local open-weight modelNo account or API keyEvery line stays yours
01 / COMPARE
A little less rereading.
0 notes
0 notes
English prototype · up to 50 nonblank notes per side · 500 characters per line · 16,000 per box. Compare important details yourself. Names, relationships and subtle meaning can still fool the model.
Ready. The model loads only when you compare with AI.
02 / READ WITH CONTEXT
Every current note, accounted for.
A similarity score is a cosine, not a probability. “Possible repeat” means check the two quotes. No notes are removed or rewritten.
SMALL MODEL. VISIBLE REASONS.
A suggestion,
with the receipts.
- Match the meaning. A pinned, open-weight MiniLM model embeds both sets locally in your browser. Model files are about 23 MB, plus the runtime.
- Check the details. Lexical guards look for changes to quantities, dates, negation, selected status words and uppercase identifiers. Close or reused matches go to review.
- Keep the originals. Every current note and its earlier reference remain visible. The audit includes text, scores, reasons, settings and measured timings.