Task Ontology v1 (Core Labels)¶

Version: 1.0 Status: Stable Last Updated: 2026-02-23

This document defines the universal, user-agnostic core task labels used by the global classifier.

These labels are:

• Stable across users
• Observable from logged signals
• Designed for time tracking
• Separable using current feature set
• Independent from user-specific taxonomies

1. Core Label Set¶

The classifier predicts exactly one of the following core labels per window.

Label IDs are stable and MUST NOT be reordered without version bump.

2. Observable Labeling Rules¶

Labels must be assigned using observable interaction signals. Intent or mental state must NOT be used.

Below are operational rules for human labeling consistency.

0 — Build¶

Primary characteristics: - Editor or terminal foreground - Sustained typing - High shortcut usage - Moderate-to-low app switching

Examples: - Implementing features - Refactoring - Writing structured code

Excludes: - Investigating logs (Debug) - Reading documentation (ReadResearch)

1 — Debug¶

Primary characteristics: - Terminal-heavy or log-heavy interaction - Frequent switching between editor/terminal/browser - Moderate typing - Higher context switching than Build

Examples: - Investigating failing test - Checking logs - Running commands repeatedly

Excludes: - Writing new feature code (Build)

2 — Review¶

Primary characteristics: - Reading diffs or technical material - Light edits - Moderate scrolling - Lower sustained typing than Build

Examples: - Code review - Reviewing PR - Reviewing technical material with occasional edits

3 — Write¶

Primary characteristics: - Sustained typing - Low shortcut rate - Lower terminal usage - Structured text writing

Examples: - Writing documentation - Writing design notes - Writing blog posts

Excludes: - Coding (Build) - Messaging/chat (Communicate)

4 — ReadResearch¶

Primary characteristics: - High scrolling - Low typing - Browser-heavy - Low shortcut usage

Examples: - Reading documentation - Researching topics - Consuming articles

5 — Communicate¶

Primary characteristics: - Short bursts of typing - Frequent app switching - Messaging/email tools - Moderate click activity

Examples: - Slack/Teams chat - Email responses - Coordination messages

Excludes: - Long-form writing (Write)

6 — Meet¶

Primary characteristics: - Very low interaction - Meeting app foreground - Long uninterrupted window - Minimal keyboard/mouse activity

Examples: - Video calls - Voice calls - Screen-share sessions

7 — BreakIdle¶

Primary characteristics: - Near-zero interaction - Idle gap exceeded threshold - No meaningful task signals

Examples: - Lunch break - Walking away - Computer idle

3. Ambiguity Policy¶

The model outputs probabilities for all core labels.

A window may be marked as rejected if:

max(core_probabilities) < REJECT_THRESHOLD

Rejected windows: - Are labeled as Mixed/Unknown at inference layer - Are NOT part of training label space - Are logged for active learning

Training data MUST NOT include Mixed/Unknown as a class.

4. Label Transition Guidelines¶

For time tracking stability:

• Adjacent windows with identical predicted labels should be merged.
• Label changes shorter than MIN_BLOCK_DURATION should be smoothed.
• BreakIdle should override other labels if idle threshold exceeded.

5. Non-Goals¶

This ontology does NOT: - Encode user-specific goals - Encode productivity judgment - Encode domain-specific categories (e.g., PianoTeaching, Investing) - Encode emotional states

Personalization occurs via mapping layer, not by changing core labels.

6. Versioning Rules¶

Any of the following require version bump: - Adding/removing a core label - Reordering label IDs - Changing semantic definition - Changing reject policy behavior

Minor clarifications that do not change semantics do not require version bump.