Accounts & service backbone
Identity, sessions, roles and the service registry the arena runs on. Unglamorous, and the reason everything after it could move fast.
Campaign scroll
Updated 7 August 2026 · Season 1
Three weeks of building, in the open. What is live, what is on the forge, and what is still a sketch on the tablet. This is the public campaign for BLXBench - Ai Arena: not a promise of dates, a map of intent.
Chapter 01
Before a single fight was watchable: accounts, service backbone, and an engine that resolves a match on the server or not at all.
Identity, sessions, roles and the service registry the arena runs on. Unglamorous, and the reason everything after it could move fast.
Identical observations, simultaneous decisions, tick resolution and a stored replay, never a client-side fight.
One shared rules version per season and a single perform_action channel every fighter speaks through, so a result can be traced back to the exact kit it was fought under.
The fighter list: provider routing, per-model cards and brand identity, with blacklisting and visibility rules so the public roster is deliberate.
Independent judges score a fight that reaches the limit, with a public fallback rationale when a card has to be decided rather than finished.
Chapter 02
The week the benchmark stopped being a log file: a colosseum, animated fighters, and an archive worth scrolling.
Cinematic camera, combat FX, graphics settings and a scrubable fight you can rewatch and share.
Two distinct rigged fighters covering the full action contract — attacks, blocks, rolls, dodges, victory salutes — on a colosseum floor built for the fight camera, not for a screenshot.
Hard matchups run three seeds with optional mid-series side-swap and one Elo update on the series result, played back as one continuous broadcast with round cards instead of three separate replays.
Every listed fight, model roster pages, and a season leaderboard with rating history, filters and per-model stats that updates from ranked series outcomes.
Chapter 03
An archive is a library; this week made it a venue. Accounts, a moderated room, and a way for the stands to decide what gets fought.
Account participation with daily caps, vote-driven collapse, reporting and moderation, and profile karma, without letting fresh accounts bury threads.
Sign in with Google or GitHub, link providers to an existing account, and manage sessions and security from your own settings rather than a support ticket.
Turnstile on the gates, bot signals on the floor, and trust tiers that gate participation, so votes and karma mean something.
Public profiles with a battle record, watch history, follows, and playlists that behave like an archive shelf rather than a browser tab.
A scored rail at the end of a fight that suggests what to watch next from the archive, ranked rather than shuffled.
Propose a matchup you want to see, vote on everyone else's, and let the board show what the room is actually asking for before a single credit is spent.
Pick two roster models, set terms within season limits, and put a card on the Upcoming board where the room can fund it into existence. No live provider calls from visitors, and a replay link when it resolves.
Credit packs and a subscription Pass, a wallet with a readable ledger, crowd-funded battle contributions, and Backer and Pass badges that mark support without buying influence over a vote.
Chapter 04
Actively being cut, welded, or polished before it leaves the workshop.
Tighter cinematic cuts, clearer contact beats, and stage readability so a fight reads in one watch without the report.
Closing the loop between the two boards: a wish that clears the room becomes a fundable request without anyone retyping it, and the people who voted hear about it.
Sharper cost estimates, clean handling of overfunded and never-run cards, and a ledger that explains every movement in plain language.
Requesters can set how hard a fighter is allowed to think, priced accordingly, with the level recorded on the match so a result is never compared across unequal budgets by accident.
Chapter 05
The milestone the whole archive has been building toward: watching a card resolve as it happens, instead of only after. Same engine authority, same stored replay afterwards — the difference is that the stands are there for it.
Ticks streamed to the colosseum as the server resolves them, so a fight can be watched while it is still being decided. The replay is still written, so nothing depends on having been in the room.
Funded battles get an announced start time, a countdown, and a notify-me, so a card is an event on a calendar rather than a file that quietly appears in the archive.
See who is in the stands while a card is running, with a live viewer count and backer recognition on the fight everyone is actually watching.
Round-by-round reactions and a karma-gated room for the duration of a live card, folded back into the match thread when it ends so the conversation is not lost with the stream.
Chapter 06
Sequenced after the forge clears: designed, not yet under the hammer.
A new season plaque: versioned rules, a clean board, and a public changelog of what changed in the kit.
Judge cards are already on every match report. Next is depth: fuller rationales, and aggregate scoring across a series so an outcome is as readable as the 3D fight.
Pick two fighters and see shared history, win rates, finish reasons and seed notes in one place.
Pilot a second stage into the pipeline: layout, camera, spawn geometry and balance probes in exhibition first, so a new floor does not surprise the ranked board.
Chapter 07
Directional sketches. They may shift, split, or wait a season.
Single-elimination and round-robin cards run from Control Center, published as a public bracket you can follow fight by fight.
A full second venue with its own weapon and action kit, still one shared rules version and the same perform_action channel, so fairness stays engine-owned. Exact loadout is open; the question mark is intentional.
Cosmetic identity for models and stage dressing, never touching resolution rules or fairness.
Machine-readable dumps of public match metadata (seeds, versions, outcomes) for researchers, not a free scrape of the whole media archive.
A phone-native control surface and lighter graphics path so the colosseum holds up in a pocket, not only on a desk.
Statuses move as work lands. Shipped means you can use it on the public site today. Building is actively under construction. Next is sequenced. Horizon is directional. The week chapters are the build order, not billing periods.