mirror of
https://github.com/superdesigndev/treg.git
synced 2026-10-02 03:24:35 +08:00
SToneX
a8800fcad9
decide's last rule read "nothing kept" as "not a task". Most such queries are
tasks on a platform the catalog lists, for a job it does not: posting to
Threads where only reading it is listed, a Telegram bot, an options quote,
a sound effect from text. The judge names the platform with confidence and
keeps nothing, and the answer was "try describing the data you want" (a
person) or the lexical page (an agent): for "publish post to Threads" the
Threads search endpoints, which the agent then re-queried past, and the miss
list never saw the gap.
A confident platform Choice with nothing kept is now a gap (rule 8; "not a
task" is rule 9), so the empty page says so, the agent's hint names the
platform ("the catalog has Threads but no tool for ..."), and the search is
recorded as a gap to add. On the agent eval set: 29 of 33 "not a task"
answers become gaps, job-hit unchanged, false-none 6% -> 7% (one of them a
lexical page that did carry the endpoint called).
Fragments: find.md (the rules), search-experiment.md.
