Your mistakes, measured.

A mistake is a number, not an adjective

Language models are notoriously bad at calibrating severity: they call everything important, or nothing. So severity isn't left to the model. Game state is converted into gold-equivalent advantage, then into win probability, and the mistake is the drop it caused.

The evaluation is never a lone number, because a coach has to say why. The same position at 20 minutes, decomposed:

Gold	−5,550
Experience	−486
Towers	−2,000
Dragons	+1,300
Herald	−450
Grubs	−360
Advantage	−7,546

It also separates blame, because dying yourself and your team losing Baron while you pushed top are different problems:

direct

You died, or you handed over the kill or the objective. An individual decision.

positional

It happened away from you, and you were on the wrong side of the map. Map reading and rotations.

team

You were there and it still went wrong. Fight execution.

The report doesn't stay in a tab

Reading that you misplayed at 14:54 is one thing. Watching the card appear on screen, at the minute it happened, with the wave still pushing and the dragon still on the timer, is another. Open the replay in the League client and RiftCoach draws on top of it.

This only works with a replay running, never during a live game, and the refusal is enforced in code. It's the same distinction a football coach makes when pausing Sunday's tape: the match is over, there is nothing left to win.

Against whoever stood in front of you

Gold, experience and farm against your direct opponent, minute by minute. Above the line you were ahead; below it, behind.

It is not a comparison against challenger players, and that is deliberate: matching your curve against theirs would need their match timelines, minute by minute. Your direct opponent is the against someone your own match already supports, and the one who actually decided your lane.

Every claim carries its evidence

Riot's API has no wave-state field. So a tool that confidently tells you to freeze is guessing, and you have no way to know which parts to trust. Each statement is tagged with how it was obtained.

1. T1
   Measured

   Read straight from Riot's telemetry. "CS at 10 minutes was 61." Nothing between the number and the source.

2. T2
   Derived

   Calculated from telemetry plus a stated assumption. "Lost priority at 14:10, derived from CS pace and position." The assumption is mandatory.

3. T3
   Inferred

   Estimated from video or heuristics. "The wave looked like a slow push." Useful, and never presented as fact.

Item and champion numbers are injected from a patch-pinned database, and a validator rejects anything that doesn't exist in your patch. It will never tell you to build something removed six patches ago.

It tells habit from accident

One report can't answer the question that matters: was that a pattern, or was it one bad game? From your fifth analyzed match onward, RiftCoach counts.

The counting happens in Python, never in the model. The model receives the finished number and writes on top of it. That separation is what stops a confident "you always do this" built on two games. Your history is a file on your machine, and one command erases it for good.

What it doesn't do

• It never runs during a live game. Not as a policy, but architecturally. The only component that talks to the League client refuses to work outside replay mode, and it rechecks before every single request.
• It can't judge trades or combos from telemetry. Riot's timeline emits no ability-use event at all, and health is sampled once per minute, which is useless for a three-second trade. Those only get real evidence from the video path, and even there they stay inferred.
• Position resolution is one point per minute. That's what the timeline gives, so every measured mistake is accurate to the minute, not the second, and the minimap draws an area, not a dot.