NoticeProofLocal deadline triageEvidence and limits

Keep the date. Keep the proof.

A small local model ranks dated sentences. You review the original wording before exporting an all-day reminder.

Date reminders only. Always check the original cutoff time and timezone. No automatic calendar changes.

Loading the local model. No notice text leaves this browser.

1. Paste the notices

2. Review the dated text

No review yet. Add a notice or load the fictional example.

Model evidence and limits

The 11.6 KB multinomial naive Bayes model was trained on 80 fictional sentences. A fixed 32-sentence synthetic evaluation gave 15 true positives, zero false positives and one missed deadline, compared with five false positives and five misses for a keyword baseline. A broader post-hoc rule diagnostic scored 27/32 with no misses and five false positives. The model missed one true deadline, so all dated sentences remain reviewable. This is a tiny related-template test, not real-inbox validation or a guarantee. Raw scores are not calibrated probabilities.

Every dated sentence remains visible, including model misses. Only literal YYYY-MM-DD and DD Month YYYY dates are recognized. Missing years, relative dates, times and ambiguous slash dates need manual handling. Multiple dates in one sentence and exact normalized-sentence conflicts are blocked; paraphrased conflicts can be missed.

All processing is local. No login, API key, external model, inbox access or calendar account connection. No event is selected by default. Editing any source clears the prior review and approval. Up to 10 notices and 40 dated sentences per batch; this interface accepts two notices, each up to 12,000 characters.

Source: code, model, labels, fixed evaluation and baseline results. Earlier today this unsubmitted project used a proposed cloud extraction route; this version replaces it with actual local inference. No cloud inference result is claimed.