Version 3.3.1

3.3.1 is a focused follow-up on the trust signals behind your match score and a quiet rework of the ingestion pipeline. Most of this is invisible plumbing — but the visible piece (no more "100% Great" on a one-skill extraction) was a UAT finding worth its own release.

Improved

• Match scores now show how much they're based on — A job whose AI extractor pulled only one or two required skills used to display "Match: 100% (Great)" if you happened to have that one skill. Mathematically true, but it overstates confidence on sparse data — and it dragged down trust in every other score you saw. The badge now switches to "Match: 1 of 1 (limited data)" (or "2 of 2", etc.) below an internal confidence threshold, with a neutral color so the green "Great" cue stops firing for sparse extractions. Above the threshold, the percentage and qualitative label work exactly as before.

• Email-forwarded batches now drive the same "Scoring N of M…" pill as auto-search — Forwarded jobs were getting parsed and added to your feed, but the per-batch progress indicator that lights up during auto-search wasn't firing for email ingest. The "Matching…" pill stayed put while scoring quietly ran in the background. Email batches now route through the same SQS-backed scoring path as the rest of the platform, so progress events resolve correctly and the dashboard tells you when it's done.

• Auto-search no longer rescores already-scored jobs — When a new search surfaced a job already in your feed (different aggregator, same role), the system would queue it for scoring again. Real fix: jobs surfaced via dedup/merge no longer re-trigger AI scoring. New jobs get scored once. This was eating roughly half of our scoring budget on busy accounts; you'll see it as faster scoring overall.

• Streaming auto-search now honors "exclude remote" the way the regular path does — If you'd unchecked the include-remote box, the live auto-search you saw on screen would still poll RemoteOK and Jobicy under the hood (they only return remote jobs, so they couldn't help you). The polished, non-streaming path got this right; the live one didn't. Both paths now apply the same source-eligibility logic and skip remote-only adapters when your intent excludes remote.

Behind the scenes

• Inactive Preview accounts no longer pile up auto-search load — On Preview specifically, accounts that haven't been active for over a week now skip the hourly poll cycle. This doesn't affect Production accounts (where any paid user keeps polling regardless of login frequency); it's a Preview-only cleanup.

• Local development defaults match production posture — Some scoring background sweeps that we paused in production are now off by default in local-dev environments too, so dev parity stays clean.

---

Most of 3.3.1 is plumbing. If you notice the match badge changing on a job you've been watching, it just means the data we pulled wasn't enough to call a confident percentage — open the job and you'll see the underlying skills.