Questions

Against Nexto, yes. It means KiariBot scored about seven of every ten goals in the series, and the interval that comes with it is roughly nine tenths of a point either side, so the reading sits a long way clear of half.

What it does not tell you on its own is anything about a different opponent. It is one measurement against one bot, repeated often.

No. Goal share counts goals. The two policies play until ten thousand goals have gone in between them, and the number is the fraction that were KiariBot's.

Nothing on this machine keeps score by match, so there is no such figure to report. If you see one attached to this bot somewhere else, it did not come from here.

Because without it the number claims more precision than it has. A ten thousand goal series carries an interval about nine tenths of a point either side, a little under two points end to end, and some of the changes made to this run are worth less than that. Print the number alone and you cannot tell an improvement from a re-run.

The chart on the front page holds one measurement repeated: about ten thousand goals against Nexto from varied starting situations, every thirty five minutes or so. The harness runs other evaluations against the same opponent to the same goal count, and the most frequent of those plays nothing but kickoffs. Its middle reading is a couple of points above the headline and it is a far rougher number than that sounds, having come back as low as 38% and as high as 99%. Four more are one offs: a longer series, two against a different bot, and one against this run's own earlier checkpoint. All real, none of them on that line, because a trend built from mixed protocols averages different questions together.

Inside the one protocol the harness still moves around more than a single interval suggests. Across the last thirty series the middle reading is 69.2%, the lowest 66.7% and the highest 70.4%, and nothing on that line has ever come back above 71%. A reading near 100% attached to this bot came from a kickoff only run having a good day. The headline is the last series confirmed under the charted protocol, not the best number the harness has ever produced.

They are read out of the running trainer every twenty seconds. Step count, iterations, steps per second, entropy, KL and learning rate come from the process itself. The goal share and its interval come from the evaluation harness's own state file for the checkpoint it last confirmed. Nothing here is typed in by hand.

When the reader cannot reach the trainer the page prints "not reporting" where the number goes. Not a zero, and not a spinner that never resolves.

One decision by one car in one simulated match. The machine gets through about a hundred thousand of them every second, and the counter on the front page has never been reset.

Itself, as it is right now. It plays a situation out against a car running the same current policy, gets scored, and the score shapes the next version. The opponent improves at exactly the rate it does, which is the point.

It has never trained against human games and it has never trained against Nexto. Nexto is used only at evaluation time, because a bot trained against the thing that grades it produces a grade worth much less.

Around Grand Champion, probably the upper half of it. That is an estimate rather than a placement, and the chain behind it is short: the community generally puts Nexto around Grand Champion, KiariBot takes roughly seven goals in ten against Nexto with an interval nowhere near half, so it is clearly ahead of a Grand Champion bot without being a tier clear of one.

The link that estimate cannot check is the last one. Turning a goal share against a bot into a place on a human ladder assumes the two map onto each other, and nobody has measured that. One rank of headroom either way is as much as this supports, which is why the scale on the front page draws a band instead of a pin.

Unless you are around Grand Champion, most likely. The rank estimate and the reasoning behind it are on the front page.

The shape of it, if you want to plan: it is heavily practised at kickoffs, air dribbling, and coming off a wall with the ball still on the car. About one part in sixteen of its practice is defending, which is the thinnest part of the mix, and patient play that makes it commit first works better than trying to out-mechanic it.

The arena is free and takes about a minute, which is a shorter route to the answer than this paragraph.

Yes, in a browser, at arena.kiaribot.com. The match runs on the training machine and is streamed out, so there is nothing to install.

Eight people fit at once and the ninth is told it is full, because one encoder feeding eight streams out of a house is what the connection carries. Driving is one person at a time: if someone has the wheel you get a queue position that updates as the line moves, and you watch while you wait.

The next live viewer takes the wheel and is told so, and you go to the back of the queue. You stay connected and you keep watching. Nobody is disconnected from the stream for having finished a turn.

If you close the tab you lose your place, because your place in the line is your open connection and nothing else.

That is what the Local tier is for. The policy runs on your hardware, in your own copy of the game: freeplay, private matches, custom playlists, the same situation as many times as you want.

You get the checkpoint that has been evaluated plus each new one as it is measured, and you can pick an older and weaker checkpoint if the current one is not fun yet. The training code and the reward function are not part of it.

Yes, and the record is the reason to believe it rather than my saying so. Fitted across every recorded series the goal share gains about a point per ten billion steps, and the most recent thirty billion came in a little above that. The run has not stopped and the curve has not flattened.

Any window shorter than that says nothing either way. A single series carries a band about nine tenths of a point either side and ten billion steps is worth about a point of progress, so two readings that close together can come out in either order. Where the ceiling is, the record cannot answer yet.

Because that is the computer there is. It means one run at a time, so there is no control run sitting beside this one and a change that looks good has to be judged by continuing and measuring again later.

The confidence intervals printed all over this site exist partly to make fooling myself about that harder.

It stops. The counter freezes, the front page says the trainer is not reporting, and the last evaluated checkpoint stays exactly as strong as it was on the day it was measured. A trained policy does not decay, it just stops improving.

If you have the bot locally it keeps working with no connection to any of this.

No. That is the one thing that stays on the machine.

Everything else is public and on this site: the step count, the learning signals, the curriculum with its real names and real emphasis, the whole evaluation history, and the score of every checkpoint that has one. The reward terms and their weights took the longest to get right and would be the quickest to copy. They are not in any tier, at any price.

You should not have to. This is one person's machine reporting on itself, and being suspicious of that is the right instinct.

The part you can check without trusting me is the arena. It is free, it is the live policy rather than a recording of a good day, and it will either beat you or it will not. That is why it is free and why it stays free.

No. KiariBot is an independent project with no connection to Psyonix or to the game as a product. Training happens inside a physics simulator that reproduces the game's mechanics, not inside the game itself.