A small Safari extension, Jev, three probabilities and a bouncer.

This is starting off as, possibly, one of the lamest jokes ever but I’ll try to keep it together.

jev

If you have been around X a bit or similar (sorry about that), you are probably familiar with jev and the hype around all of that other train. My feed won’t shut up about a model that won’t talk, funny.

I shipped already Shortless and wanted to slap jev into it, as cool kids usually do ๐Ÿ˜‚ –> Slop Patrol was born ๐Ÿ‘ฎ

Slop patrol looks for evidence AI-sloppiness on your feed, given the latency requirement it doesn’t do anything too fancy (i.e. doesn’t analyze the video itself) but it’s a step up from the naive approach of trusting Big G with their “AI content” labels they sometimes slaps on videos.

The clanker has a fairly limited job description.

sys design

The bouncerโ€™s commute

  1. โ–คSafari cardVisible metadata
  2. โ—ˆOn deviceSettings ยท cache ยท queue
  3. โŒ˜Edge serviceCache ยท shared limits
  4. ๐Ÿค–JevThree questions
Fresh check โ†’ client policy

Eligible metadata travels through the native client and edge service to Jev. Validated signals come back; the client decides whether to hide.

Network: yesModel call: yes

The usual Shorts filter still runs locally, Slop patrol is asynchronous.

you can have three questions

for jev I used three noul (bernoulli) questions and each returns a probability for a yes/no outcome.

questions are:

  • Low value: does the video metadata offer real proof of filler stuff or low-quality/info output?
  • Substantive: does it suggest useful stuff or something along those lines?
  • Insufficient evidence: is the available text too little or ambiguous to make the call? just bail out broski

Nothing too complicated here.

follow the slop patrol around

shortless.app shows actual near-real-time stats if you’d like to see how it is chugging along.

Snapshot of Shortless's live production metrics: 80 checks received, 705 ms average Slop Patrolling time and 6 cache hits, with the bouncer's routine illustrated below.

The robot finally got a job. It does not get the final word. ๐Ÿ‘‹