Annotation quality
Precision lives in the hard cases.
Useful annotation is consistent about boundaries, uncertainty, scope, and evidence. The examples below are general training principles, not client-specific instructions or a substitute for a future task's schema.
Draw the visible object—not its shadow
For a box annotation, include the visible object extent according to the task schema. Keep truncation (the image frame cuts an object) distinct from occlusion (another object blocks it). A shadow, reflection, or neighboring object is not part of the target unless the brief says otherwise.
Preserve exact text boundaries
Transcribe what is shown, including case, punctuation, and confusable characters when the schema asks for literal text. Do not silently normalize a code or serial number. If the source is genuinely unreadable, use the task's uncertainty convention rather than guessing.
Keep scope, subject, and time separate
A negated, suspected, historical, or family-history mention is not automatically a current fact about the speaker or patient. Record the stated subject and temporality; do not convert an order into proof of administration.
Label what the evidence supports
In response evaluation, distinguish a supported statement from a plausible inference. ‘Up to’ is not a universal guarantee, a small pilot is not proof about everyone, and a reported symptom is not a confirmed diagnosis unless the source establishes it.
Keep overlapping speakers distinct
When voices overlap, preserve both attributable utterances and their timing. If identity is genuinely unresolved, use the task's unknown-speaker label rather than assigning a voice based on topic or frequency.
Learn how the assessment is reviewed or see the current task-status disclosure.