Your taste,
decoded.

6-star and 1-star ratings carry 3× base weight when Cue builds your taste vectors, so your strongest opinions move your genre, network and era preferences hardest. Auteur and anti-taste clusters are built from anchor ratings exclusively, so there the multiplier is uniform and does not change their relative ordering.

A title's community score is a Bayesian posterior rather than a mean: (C·m + Σratings) / (C + n), where C = 10 virtual ratings held at m — the community's running average, sampled from the 1,000 most recently updated titles, and 3.5 until there are enough titles to measure one. Σ and n are taken over a title's 500 most recent ratings. Against that 3.5 starting prior, two 6-star ratings land at 3.92 rather than 6, while at a few hundred ratings the prior moves the score by less than a tenth of a star. Recomputed nightly at 05:00, and it carries a 15% weight in the ranker's five-layer score.

An exponential decay model with an 18-month half-life. Each of your ratings' influence on your taste profile is calculated as weight × 0.5^(months_since / 18) — fractional months, no floor — so the profile evolves as you do. Other people's ratings and TMDb vote data are not age-weighted.

Masterpiece ratings are clustered by attributed person (creator, director, top-4 billed cast), each cluster weighted by the sum of its temporally-decayed ratings. Cue seeds every recommendation pool with work from your eight strongest loyalties, so those titles are in the running regardless of genre, and your strongest single loyalty on a title — not an average across its cast — drives the boost.

When you rate something 1 star you can tag what went wrong, and those tags cluster by temporally-decayed weight into an anti-taste profile. On the other side, a title carries the criticisms at least two distinct raters independently agreed on — the same tag vocabulary on both sides, so nothing has to be inferred or translated. Where the two overlap, the candidate's score is multiplied down in proportion to how strongly you hold that aversion, floored so a title is deprioritized and never excluded. Tags naming a whole genre are dropped at both ends.

Content boundaries are applied as a 0.3 score multiplier rather than a hard filter: a matching title is ranked down, not removed from the pool, and its card carries the reason "Contains content you prefer to avoid." Because the shelf shows only candidates above a 0.12 confidence threshold, a heavily demoted title can fall below that line and not appear — deprioritized by the scorer, never blocklisted.

A pure numeric scorer does the ranking: five weighted layers (0.35 taste affinity, 0.25 auteur loyalty, 0.15 social, 0.15 community score, 0.10 external audience data) summing to 1.0, scaled by era, content-constraint and anti-taste multipliers. The order is total — descending score, then TMDb id, then media type — with no randomness and no rotation, so a fixed profile and a fixed candidate set always produce the same order. The Recommended for You shelf renders that ranking directly with no model involved; where a language model is used it receives the finished list and narrates it in order without introducing titles.

Text you type — queries, moods, saved memory, support-chat history — is wrapped in XML tags, and the prompts that answer you are prefixed with a standing instruction that tagged content is untrusted data to be analysed, never obeyed as instructions. Most AI actions are also constrained to a fixed JSON response schema; the long-form ones — recaps, summaries, spoiler-safe explanations — return prose.

Marking an episode, changing a rating or untracking a show is appended to an operation log in the browser BEFORE anything else happens — then applied to local state, then queued for the server. There is no offline mode and no connectivity branching anywhere in the command layer: the identical code runs on wifi and in a tunnel. Opposing actions collapse, so tapping watched and unwatched five times sends one operation rather than ten, and each account gets its own physical database on the device.

Every operation carries a client-generated id and a per-device sequence number, and the server keeps a ledger keyed on that id, so replaying a batch after a dropped connection is acknowledged as a duplicate rather than applied again. The backend offers no unique index, so the ledger uses a reserve-then-claim protocol that resolves every replay in which either request can see the other's reservation; in the narrow case where neither can, every operation but a rewatch is idempotent on its own logical key and converges anyway. Pushes retry with exponential backoff — 10 seconds, doubling to a 15-minute ceiling — and a network failure retries indefinitely while reconnecting cancels the wait immediately. Episode totals are always recomputed on the server from your actual history, never trusted from a number a device sent.

Show and movie metadata flows through an in-memory LRU (per instance, 500 entries, sub-millisecond) → a shared database cache → the TMDb API, and a stale entry is served instantly while it refreshes in the background. TTLs adapt to the content: 5 minutes for a search, 6 hours for a season, and up to 90 days for a series that has already ended. A nightly janitor expires stale rows and collapses duplicates. Separately, the app keeps season data and reference shelves in the browser's IndexedDB so Watch Next and the Discover shelves work offline. AI answers are cached too, in their own database table.

A probabilistic data structure (10K capacity, 1% false-positive rate) stored per-user as a hex string on their taste profile. Candidate titles are checked against it in O(k) time before scoring, so nothing you've already watched reaches the ranking — backed by an exact check, so a false positive never silently drops a title.

Each user's rated title IDs are hashed into a 128-component MinHash signature, split into 32 bands of 4 rows, for approximate nearest-neighbor lookup in taste space. A shared bucket is only a candidate — the actual overlap is then measured, and weak matches are dropped rather than trusted.

A token-to-title mapping built from cached TMDb metadata (titles, genres, overviews), stored as database records queried by exact token match. It turns a natural-language search into candidates without a round-trip to TMDb.

A partitioned bitmap splitting the 32-bit ID space into 2¹⁶ chunks. Search unions the ID sets of your query's tokens through it — O(n) with excellent cache locality on sparse integer sets — before the Bloom filter removes what you've seen.

Find Similar narrows to the same primary genre, then ranks by distance across dimensions where distance actually means something — era, runtime, and critical rating. Genres are compared by name, never by a numeric code, because two genres with adjacent IDs have nothing to do with each other.