ReplayLab turns playtest sessions into evidence your team can act on: rewind any moment, ask questions across sessions, and check design intent against how people actually play.
The problem
A playtest today:forty minutes of video,a page of notes,and one sentence:“it felt unfair.”
Somewhere in there is the moment that matters. Finding it means scrubbing. Proving it means guessing.
ReplayLab · the loop
ReplayLab turns playtest sessions into evidence your team can act on: rewind any moment, ask questions across sessions, and check design intent against how people actually play.
Seed, build, inputs and meaningful events: kilobytes per minute instead of gigabytes of footage. Player voice and comments land on the exact tick they talk about.
Session export and race diagnostics already run inside AccoBolide.
Open a session at the exact tick. Filter events, follow one driver, read the evidence behind each marker. Next: watch it again from any camera, because you are re-running the game.
The session viewer is live on this site.
Who took the most hits?
Le Cantonnier took 5 of the race’s 22 melee hits.
t=00:03.85t=00:05.05t=00:10.07t=00:44.05t=01:08.88
“Where do new players lose the lead in the last ten seconds?” The answer is a list of replayable moments, and every sentence cites its ticks. No evidence, no claim.
Cross-session comparison comes first.
200 re-runs of the same build, AI drivers only: an illustration of the method, not a balance measurement.
Write design intent as rules (“no single hit costs more than three places”) and check every session against them. Replay last week’s human sessions on today’s build and see exactly where they diverge.
We already work this way: decision rules written before the numbers, 200 seeds minimum.
Why it’s different
A recording shows what the screen looked like. A session knows what the game did, and can do it again.
Every chart, cluster and answer opens on the replayable moments behind it.
Design intent becomes a rule. Rules get checked against every session, every build.
The ideal tool
From the first prototype to the complete lab: the tool we want for our own games, and for yours.
Timeline, speed, driver focus, event filters and raw evidence for every marker.
Seed, build, inputs and semantic events: kilobytes per minute. Keyframes for engines that are not deterministic.
Every bug report becomes a replay link that opens on the right frame, in the browser.
Re-run the incident from above, from the rival’s seat, in slow motion. It’s the game, not a video.
Think-aloud, surveys and chat messages anchored to the moment they’re about.
Line up sessions, builds and players; spot what changed and when.
Ask across every session. Answers come back as moments, and every sentence cites its ticks.
“37 sessions · contact at the bridge, turn 4”, with the heatmap and the clip that sums it up.
Write the intent, get the violations, with the moments that broke it.
Replay last week’s human sessions on today’s build and see where, and why, they diverge.
Fork any moment and let bots finish the race 200 times with one setting changed.
Replay link, clip from the right angle, build and steps to reproduce, sent to your tracker.
Works today
This is the first ReplayLab prototype, running on an archived AccoBolide test race. Scrub the race, follow a driver, filter the events and inspect the evidence. You can also open a compatible report from your computer: it is read locally and never uploaded.
Loading the session…
Select an event to read its raw data.
Source: archived automated test race, build 9d21f05, simulation convoi-reprise-2026-09-27. The map uses 30 Hz trajectories re-extracted from an exact replay of the same race (same final hash); the report itself samples positions every second. The viewer describes the report: it does not re-run or certify an imported file.
Who it’s for
“Does my mechanic do what I think it does?”
“Show me exactly how to reproduce this.”
“What do we fix this week, and why?”
“What did players say, at the moment it happened?”
Why now
Racing, fighting, strategy, puzzle and browser games can replay sessions almost for free.
Language models can now answer questions over structured logs. Replay lets you check every answer.
Modern engines and the browser make recording a session cheaper than recording a screen.
Design partners wanted
We are building ReplayLab with a small group of studios who playtest often. Tell us how you test today, and what you wish you could see.
VECTVR · France · est. 2026