What this is. A fictional elimination of bracket entries. Nobody is being removed from anything; a DID is only a string in a list. The system knows nothing about anyone's posts, avatar, followers or popularity.
Roster. Entrants are full DIDs, deduplicated and sorted by string order, then frozen. The snapshot id is the SHA-256 of that list. The synthetic roster uses the reserved did:example method and is always labelled as demo data. The real roster is whatever this browser has indexed from the relay's repo list: indexed public DIDs, not "all Bluesky users". A DID in that list may be a deactivated, migrated or non-Bluesky AT Protocol repo; only active repos are kept.
Bracket. Entrants are placed by a seeded Fisher–Yates shuffle (not DID order). Each round pairs neighbours; an odd last survivor gets a bye, which is not a match. N entrants always play exactly N−1 matches and leave one champion. N = 0 has no champion; N = 1 wins unopposed.
Matches. The winner of match (round, index) is picked from a hash of the tournament seed and that coordinate (LP1, a murmur3-finalizer chain, not cryptographic). Because a result depends only on its coordinate, any batch size or worker split gives bit-identical brackets. Math.random is never used.
Chaining. Tournament 1's seed is SHA-256 of a published phrase. Tournament k+1's seed is SHA-256 of the algorithm version, previous seed, previous champion DID and previous tournament id. This is deterministic entertainment, not a lottery or a contest for prizes.
Simulation vs. playback. All matches are computed first, in a worker, with no network calls. The picture then plays back the finished bracket as fast as frames can convey it; the status line says so. Max speed skips intermediate visual states, never matches.
Rendering. One canvas, no per-entrant DOM. Zoomed out, a round's matches are aggregated into pixel-row bands (sampled, labelled AGGREGATED). Zoomed in you get real elbows, match nodes and truncated DIDs.
Verify. VERIFY REPLAY recomputes the tournament from its seed with a different batch size and compares a checksum over every round's survivors.
Not included. Sound; server-side or shared persistence; a complete real roster (millions of DIDs would take thousands of relay requests per visitor, so you index what you choose to). Benchmark numbers come from your own machine, in the SOURCE panel.