How Vortyx reduces false positives in follow-up and commitment detection

A customer may ask for an update. A founder may say they will send a deck. A teammate may promise to review something by Friday. Someone may say 'let's revisit this later' without creating a real commitment.

These signals can look similar, but they should not all become tracked work. Vortyx treats detection as a review-first workflow. It surfaces likely follow-ups, commitments, waiting items, next steps, reminders, draft replies, and recurring responsibilities for user review before they become tracked items or external actions.

Vortyx detection quality is evaluated across the main work signals the product is designed to surface.

• Follow-up detection: whether a message or capture implies that someone should follow up.
• Commitment extraction: whether someone clearly committed to doing something.
• Waiting-item classification: whether the user is waiting on another person, reply, decision, approval, document, or update.
• Ownership detection: whether the user owes the action, someone else owes it, or ownership is unclear.
• Timing extraction: whether a due date, expected response time, or reminder cue is present.
• False-positive reduction: whether casual language is avoided when it does not imply trackable work.
• Review outcomes: whether users accept, edit, dismiss, snooze, or mark suggested items handled.

Vortyx is currently evaluating these signals with a closed set of users. Quantitative metrics will be published only after there is sufficient repeatable evidence, including a labeled evaluation set, annotation rules, sample size, precision, recall, false-positive rate, and model/version context.

Vortyx separates strong work signals from softer conversation. A casual phrase is not automatically a commitment:

Let's catch up soon.

This may be useful context, but by itself it does not clearly say who owes what or when. A stronger commitment contains an actor, action, and often timing:

I'll send the deck by Friday.

This is more suitable for a proposed commitment because it includes an owner, an expected action, and a timing cue.

A soft intent is also different from tracked work. 'We should revisit this' may not deserve tracking unless there is a concrete next step. 'Please follow up next week' is clearer because it contains a requested action and timing cue.

Waiting states need care. A thread may say 'Waiting for approval from Finance.' That can become a waiting item if the approval is still unresolved. But if a later message says the approval was received, Vortyx should avoid resurfacing it as still waiting.

Recurring responsibilities are also different from one-time reminders. A note like 'renew the report today' may be one task. A statement like 'send the weekly client update every Friday' can become a recurring responsibility, but only after user confirmation.

Draft suggestions are another separate category. Vortyx may suggest a reply or follow-up draft, but it does not treat a draft as a user-approved action.

A useful Vortyx suggestion should be traceable to evidence. For each suggested item, Vortyx should preserve enough context for the user to understand why it appeared.

• Source message, thread, email, capture, or note.
• Detected owner.
• Expected action.
• Timing cue, if present.
• Whether the suggestion needs review.
• Reason the item resurfaced, such as overdue, waiting, needs setup, blocked, or recurring.

This traceability matters because detection is not perfect. Users need to see what Vortyx inferred and decide whether it is correct.

Vortyx uses review-first design to contain uncertainty. Suggested items are not automatically converted into final tracked work without user review. Draft replies remain editable. Email actions are not sent automatically.

Users can confirm, edit, dismiss, snooze, mark handled, or leave an item pending. This reduces the risk of false positives because uncertain detections remain reviewable. A weak or ambiguous suggestion may still be useful, but it should not execute externally or change the user's workflow without approval.

Vortyx is not designed to silently act on behalf of the user.

• It does not automatically send emails.
• It does not automatically approve suggested drafts.
• It does not treat every detected phrase as a task.
• It does not assume every future-looking sentence is a commitment.
• It does not remove the user's judgment from ambiguous communication.

The goal is to keep important work visible while preserving user control.

False-positive reduction in Vortyx is not only a model-quality problem. It is also a product-design problem.

The system should identify possible unfinished work, show the evidence, preserve the context, and let the user decide. That review-first approach allows Vortyx to be useful even when communication is ambiguous, without silently creating tasks or taking external action.