How recommendations work

Curator uses a deterministic, inspectable pipeline. You do not need to understand the scoring to use it, but every score and explanation can be inspected. Missing metadata is treated as unknown rather than negative, and an unplayed scene is never assumed to be disliked.

What teaches Curator

Curator uses Stash ratings and viewing history, playback captured in the browser, scene feedback, and direct tag preferences. Performer ratings and favorites, and studio favorites, provide additional hints. Tags, performers, and studios connect those signals to other scenes. Better metadata and a few deliberate answers give it more useful evidence; you do not need to rate the whole library.

Tag ratings express a preference for that feature. Strong dislike lowers preference; Never is a hard exclusion for matching content. Clear answer removes the direct preference or block. These are separate from Never show, which suppresses one scene.

Scene descriptions are display metadata, not recommendation signals. Description term extraction and ratings have been retired; saved term answers are preserved but no longer affect scores or exclusions. Scenes with sparse tags can still use performer, studio, and viewing evidence. After updating, rebuild the model and refresh Expand to replace previously computed scores.

Understand the scores

Appeal is the long-term estimate of how satisfying an item is likely to be. It combines bounded evidence from content, performers, studios, similar scenes, and direct item history. Strong direct outcomes can override weaker inferences.

Tag, performer, and studio evidence has two sources. Outcomes say how the scenes you watched went compared with your average watched scene. Watch lift says what you choose: a tag, performer, or studio on a larger share of the scenes you watched than of your library counts in its favour, and a common one you keep passing over counts against it. Lift is shrunk for those with few watched or expected scenes, and it adds a small amount of evidence next to outcomes. Your Tag sentiment answers, Why these? reasons, and Stash performer and studio favourites and ratings still apply on top.

Current Fit is how suitable the item is right now. It starts from Appeal and adjusts for exact-scene cooldown, recent performer or content repetition, and Not now feedback. Time changes today’s fit; it does not erase learned taste.

Confidence is the strength and variety of the evidence behind the estimate. It is not another preference score. A high estimate with thin evidence belongs in a different lane than a high estimate backed by varied outcomes.

The Curator affinity gauge on performer cards and pages shows the performer signal used in recommendations: confidence-weighted scene evidence, your performer rating and favorite bonus, and a weaker estimate from similar performers. Similarity confidence increases with strong, agreeing matches and is capped at 65%. Missing performer attributes are unknown: similarity compares the usable attributes on both performers, without requiring every field to be populated. Conflicting matches reduce confidence. Strong direct evidence reduces the inferred contribution. Estimates never become training evidence for other performers. The similarity estimate appears after rebuilding the model. Its tooltip breaks those contributions down while confidence is shown separately below the value. A performer without evidence shows 0.00 · No evidence: they contribute zero, but this is not a learned neutral preference. A dash means the value is loading or unavailable. Scene evidence measures outcomes relative to your library baseline; a negative learned contribution can coexist with a positive overall preference. Favorites, rated performers, and performers with similarity evidence can have a value without scene history.

This is a signed recommendation contribution: zero is neutral, positive values favour the performer, and negative values reduce their contribution. Its size is separate from confidence. A favorite adds +0.18; a performer rating adds between -0.10 and +0.10, with 50/100 neutral. A favorite rated 100/100 therefore starts at +0.28 before scene evidence. There is no +0.35 cap. The raw combined signal includes a confidence-weighted similarity estimate; the ring fills at magnitude 1.0. For scene scoring, direct performer evidence and similarity remain separate contributions bounded to 0.30 and 0.16 respectively, alongside content, studio, and other scoring signals.

Lane policy

• Best Bets requires strong fit supported by corroborating signals or reliable direct positive evidence, and excludes anything with recorded viewing history.
• Revisit requires a prior strong positive and enough cooldown recovery.
• Stretch keeps a familiar anchor while naming one confirmed tag or studio the model challenges — either a dimension it has learned to dislike, or one it has too little evidence about — and requires that named challenge to exist.
• Blind Spots surfaces studios or confirmed tags you have barely played, gated on at least two independently corroborating facets so a single noisy field can’t qualify a scene alone.
• Dormant offers unseen scenes with a performer, studio, or confirmed tag you had a real positive history with but haven’t touched in a while. The model build checks how long that preference has been dormant.
• New Arrivals ranks unplayed scenes by Current Fit among those most recently added to Stash: everything added in the last 14 days, or the 50 most recent when fewer arrived. Scenes not yet marked Organized wait until they are, unless Include unorganized scenes in New Arrivals is enabled. A new scene has no feedback of its own, so its fit comes from performers, studio, tags, and similar scenes.
• For You mixes those policies with conservative items early, one New Arrivals slot in every 20, and only a small Blind Spots and Dormant share.

Hard exclusions, unavailable files, explicit suppression, and Prune state are checked before lane scoring.

Variety is presentation, not taste

After candidates qualify, Curator builds the ranked slate one card at a time. It avoids adjacent performer repeats and softly varies studios and very similar content. The page then draws from the strongest candidates: higher rank improves a scene’s odds, while recent qualified views that did not lead to a play or thumbs-up lower them temporarily. These presentation choices do not change Appeal or the evidence shown in Why this?.

Balanced uses the varied order; Score-first removes the variety adjustments. Filters narrow the qualifying candidates without teaching a preference.

How explanations work

Recommendation cards show a short reason for their selection, including positive preference matches for Best Bets, prior outcomes and cooldown recovery for Revisit, the named past preference and time away for Dormant, and when the scene was added for New Arrivals. For You uses each card’s source lane. The visible reason is capped at two lines; Why this? retains the full explanation.

Stretch names a familiar anchor and its selected challenge, distinguishing too little evidence from a negative preference estimate. Blind Spots names the underexplored studio and tag, with play counts in the expanded explanation. These facts belong to the model that selected the card; missing evidence is reported as unavailable.

Negative scenes in Sentiment review name contributing content features, performers, studios, and examples of similar scenes. Similar-scene outcomes are compared with the library average; below-average outcomes need not be negative. Cooldown and Not now affect Current Fit, so they are not reasons for negative sentiment. These explanations describe estimates, not necessarily explicit dislikes.

Every explanation is planned from reason codes derived from published model evidence. Curator derives it when you expand Why this?, keeping the recommendation page fast. The plain-language summary names the strongest facts; the score tree exposes the contributions, confidence, timing changes, exploration reason, and final lane. The summary describes evidence stored in the model. It explains an estimate of your preferences; use feedback to correct that estimate when a suggestion misses.