AttentionKind in src/visionset/kernel/domain/summary.py has three members: a batch holding
frames whose review has not happened, a queued unit of machine work that stopped and said why, and
one still going. There is no member for a batch holding frames a model labelled and nobody has
looked at.
Home's attention list is the answer to what is waiting on me, anywhere — it is the one surface
that can say so across projects, which is what earns it a rail entry. A batch of forty-eight frames
that a text-prompt model wrote minutes ago, where every label is a guess no person has judged, is
exactly something waiting on someone. The list says nothing about it.
Before batch pre-labeling this could not arise. Frames reached a state needing attention one at a
time, by a person pressing submit, and review_pending covered it. A single pre-labeling run now
puts a whole batch into pre_labeled at once, which is the case the vocabulary has no word for.
The fix is a member and its row, and the wording is the interesting part rather than the code. The
existing review row says a person is waiting on a reviewer; this one says a person is waiting on
an annotator, and the two should not read the same. What the row offers as its next step matters
too: for a pre-labelled batch the useful move is to open it and start correcting, not to review it.
Related: the decision that introduced the state is recorded on the batch pre-labeling issue
(#425), and the surface that produces it shipped in #673.
AttentionKindinsrc/visionset/kernel/domain/summary.pyhas three members: a batch holdingframes whose review has not happened, a queued unit of machine work that stopped and said why, and
one still going. There is no member for a batch holding frames a model labelled and nobody has
looked at.
Home's attention list is the answer to what is waiting on me, anywhere — it is the one surface
that can say so across projects, which is what earns it a rail entry. A batch of forty-eight frames
that a text-prompt model wrote minutes ago, where every label is a guess no person has judged, is
exactly something waiting on someone. The list says nothing about it.
Before batch pre-labeling this could not arise. Frames reached a state needing attention one at a
time, by a person pressing submit, and
review_pendingcovered it. A single pre-labeling run nowputs a whole batch into
pre_labeledat once, which is the case the vocabulary has no word for.The fix is a member and its row, and the wording is the interesting part rather than the code. The
existing review row says a person is waiting on a reviewer; this one says a person is waiting on
an annotator, and the two should not read the same. What the row offers as its next step matters
too: for a pre-labelled batch the useful move is to open it and start correcting, not to review it.
Related: the decision that introduced the state is recorded on the batch pre-labeling issue
(#425), and the surface that produces it shipped in #673.