For a day and a half this page sampled the Bluesky firehose and asked a calibrated model eight typed questions about every post, one call each, a few hundred milliseconds apiece: intent, topic, tone, engagement bait, hostility, sarcasm, bot-ness, and whether it was safe to show. Anything the model would not commit to went to a human lane.
82,727
posts judged
73M
input tokens
$3.09
total model spend
32%
sent to a human
It is paused while the budget refills. The code, the API, and the nightly dataset dumps are open.
