Changelog
v3.1.1.1 - Recommendation Engine V6 and reliable personal Discovery
- Recommendation Engine V6 requires fresh Learned/Neural fitting over Activate runs; media, watch history, settings and spent acceptance IDs remain intact, and loading does not rewrite older parameter files
- One shared 41-feature contract, rebalanced bounded heuristic scoring, and Neural 41 -> 82 -> 96 -> 48 -> 24 -> 1 with 17,269 parameters
- Coherent story keywords and confirmed next-part collections support recommendations without language/genre-only confidence or skipping known collection gaps
- Cold/warm personal mixtures and admin/demo diagnostics distinguish global priors, provisional participation and independent verification; saved ranking diagnostics never trigger generation
- Personal provider-title Discovery routes reuse existing retrieval slots; 20 optional details share a 15-second per-user budget across fair lanes without weakening certification or request checks
- Complete Adam checkpoints preserve early-stopping and restart continuation; normalization/clamp coordinates and fitting-only imputation priors remain consistent
- Inactive-model retirement retains trust receipts and exact acceptance ledgers, allowing healthy global fallback only when withdrawal/recovery restrictions permit it
- Partial or ambiguous active Arr roots withhold orphan verdicts; remote Windows/UNC and POSIX paths keep independent comparison semantics
- Local verification: 8,109 backend, 1,771 UI, 112 browser/documentation fixtures, 64 multilingual demo checks and 4 targeted per-user filesystem E2E tests passed
- Published Codecov coverage: 94% backend and 97% UI; full Docker E2E suite inventory: 411 tests in 58 files
- Requires Jellyfin 12.4
Checksums
- MD5:
504bf1a25ea80acd15384232224e0301 - SHA256:
a6affcdeff8b4c13b3880b482d8eab42222707c26fe1b888c2967df472075392
Installation
Download JellyfinHelper_3.1.1.1.zip and extract it into your Jellyfin plugins directory.