How scoring works
A vision model only reads your screenshot. The pick itself comes from a fixed scoring model: eight weighted factors — nine if you asked for a carry pick — plain arithmetic, no randomness. The same board and the same mode always produce the same champion, and every sentence you are shown is generated by the factor that earned it.
Where the AI stops
Reads the screen
Vision model
Champion names, your lane, the bans. It never gets asked what to play.
Decides the pick
Scoring engine
12 traits8–9 factors × weightstanh% fit
What it is given
- Your lane. The one you chose, if you chose one — a role you state always beats the role we read, because a wrong role does not just shuffle the ranking, it removes every champion you could actually pick.
- Both teams. The champions read off the screenshot: up to four allies — your own slot is empty by definition — and up to five enemies.
- The bans. Only when we are at least 70% sure of what we read, a higher bar than anything else on the screen: recommending a banned champion is embarrassing but obvious, while a ban we imagined would quietly delete a legitimate pick and you would never know.
- The champions to choose from. Everyone played in your lane — or your own pool, if you have told us which champions you play. Anything already picked or banned is removed first.
The factors
Each one rates a champion from 0 to 100, where 50 means “this factor has no opinion” — which is also what it returns when the board is too unread to say anything. That is why a half-read screenshot degrades instead of guessing. The weights shown are team mode’s, live right now and tuned without a deploy; the carry-mode figure is beside it wherever the two differ.
Role fit
×1.30Is this champion actually played in your lane?
A gate rather than a nudge — the heaviest factor in the model, so a champion nobody plays in your role cannot win on the strength of everything else. Above the gate it still prefers a champion's main lane to the edge of its range, which is why you are not offered a support for top. If we do not know your lane, it has no opinion.
Enemy matchup
×1.15 · carry ×1.38How does it fare against the enemy team?
Five paired terms — their lockdown against your mobility, their frontline against your anti-tank damage, their divers against your crowd control, their poke against your engage, their engage against your durability — plus a head-to-head with whoever is in your lane.
Team synergy
×1.00 · carry ×0.45Does it fill a hole in your own team?
Looks for the missing piece: nobody to start a fight, not enough crowd control to make a fight stick, no peel or utility for the carries. If your team already has engage, it rewards the damage that follows up instead.
Damage balance
×0.95 · carry ×0.47Does it move your team toward mixed AD and AP?
Compares your team’s magic/physical split before and after the pick, and also judges the split it leaves behind — a fifth AD pick cannot make the ratio worse, so improvement alone would call it neutral instead of bad.
Lane win rate
×0.90 · carry ×2.70Does it actually win in your lane?
The champion’s published win rate in your lane, pulled toward 50% the less it is played so a niche pick’s lucky week cannot dominate. The only factor that comes from measurement rather than from authored traits — moderate in team mode, the heaviest factor in carry mode. Silent when the champion was not measured in that lane.
First pick
×0.90Is it a safe opener when nothing is picked yet?
Only active on an empty board, where every other factor honestly has no opinion. It rewards being self-sufficient at both ends of the game, hard to punish once the enemy has seen it, and playable in more than one lane.
Frontline
×0.85 · carry ×0.38Is there a body to stand in front?
About one and a half real frontliners is treated as healthy. Two-sided: it also marks down a fifth squishy on a team that has nobody.
Scaling
×0.70 · carry ×0.59Does it answer the game length the draft is heading for?
Compares how the two teams age. If the enemy out-scales you, you want either scaling of your own or a faster clock; if they are the early-game team, staying alive to your own late game is the win condition.
Objectives
×0.40 · carry ×0.20Does it add the dragon and baron pressure your team lacks?
The lightest factor in the model — a tiebreaker between close picks, not an argument on its own.
Lane counter
team: none · carry ×1.45Does it beat the champion standing in your lane?
Carry mode only, and the heaviest factor there. Two halves added: how strong the champion is in a lane at all, and a five-term duel against every enemy who plausibly plays your role — weighted by how likely each is to be the one opposite you.
Where the % fit comes from
The ratings become one number by weighted average, then get stretched onto a scale a human can read:
raw = Σ (rating × weight) / Σ weight fit = 50 + 49 × tanh((raw − 0.5) × 9)
The stretch is there because most factors are neutral in most games, so the raw scores of every candidate crowd together just above 0.5 — a percentage taken straight from them would tell you nothing.
The pick and the alternatives
Every candidate is scored, sorted by fit — ties broken alphabetically, so the answer never flickers — and the top of the list is your pick. The next 9 become the alternatives, labelled by how much fit you give up by taking one instead: within 2 points is strong, 4 is medium, 7 is neutral, further behind is weak.
The gap is used rather than the raw number on purpose, because a fit is not comparable between two screenshots — see below. “What am I losing if I take this instead?” always is.
The reasons under the pick are the factors that had something to say, ranked by how far each one actually moved the number. “Why this pick?” shows the same factors as a breakdown, so the explanation is a view of the maths and never separate copy.
The two modes
The toggle above the drop zone picks which question the model is answering. It is not a filter or a display setting: it changes the weights, and carry mode adds a ninth factor that team mode does not compute at all. The same board can honestly produce two different champions.
Team · default
“What does this draft need?”
The model as it has always been. It fills the hole in your composition — the engage nobody brought, the damage type you are missing, the body that has to stand in front — and weighs your lane as one term among many.
Carry
“What beats the player opposite me?”
The lane becomes the argument. Composition still counts — a champion that contributes nothing after twenty minutes is still a bad pick — but at roughly half the weight, and the lane duel outweighs everything except whether you can be played in the role at all.
What carry mode does to the weights
| Factor | Team | Carry |
|---|---|---|
| Role fit | ×1.30 | ×1.30 |
| Enemy matchup | ×1.15 | ×1.38 |
| Team synergy | ×1.00 | ×0.45 |
| Damage balance | ×0.95 | ×0.47 |
| Lane win rate | ×0.90 | ×2.70 |
| First pick | ×0.90 | ×0.90 |
| Frontline | ×0.85 | ×0.38 |
| Scaling | ×0.70 | ×0.59 |
| Objectives | ×0.40 | ×0.20 |
| Lane counter | — | ×1.45 |
Role fit is untouched in both. Carry mode is about winning your lane, and putting you on a champion nobody plays there is not a way to win it.
How the lane counter is scored
Only in carry mode, and the heaviest factor there at ×1.45. Two halves, added.
1. Is this a lane-winning champion at all?
Half early game, because that is what winning a lane is. Then whichever of poke or engage is higher — a champion needs one way to take a trade, not both — plus mobility to take those trades and live, and a little scaling, because a lane you win is only worth winning if you can close the game out.
This is not the same question as role fit, which asks whether anybody picks the champion in your lane. This asks whether it wins there once picked.
2. Does it beat the champion opposite you?
Champion select never tells us who is standing in which lane, so every enemy whose role list contains yours is treated as a possible opponent and weighted by how much of a mid laner they really are. If two of them qualify, the duel is averaged rather than summed — two suspects are still one lane, and an ambiguous read must not score twice as hard as a clear one.
- TempoYour early game against theirs. The heaviest term, because it is the one that decides a lane rather than a teamfight.
- PokeYour chip damage against theirs — scaled down by how mobile they are, since poke only lands on someone who cannot step out of it.
- All-inYour engage against their escape: whether committing onto them sticks.
- Anti-tankDamage that scales on their health, counted only to the extent they are actually durable. Near zero against a squishy opponent, which is correct.
- SafetyYour durability or mobility against their all-in — and only when they have one. Being slippery opposite a passive laner is wasted, not a liability.
Beating two of the champions who might be in your lane scores higher than beating one, because both duels are counted. That is the case the mode exists for: a pick that is good into everything the enemy could still put in front of you is a better carry pick than one that is good into a coin flip.
Where the win rates come in
Both halves also read measured win rates, aggregated from 0 ranked games through Riot’s own match API: how often a champion wins in your lane, and how often it wins there against that specific opponent. Both are pulled toward an even 50% by how little we saw of them, so a matchup observed a handful of times barely moves the number. They are added to the trait terms above rather than replacing them — the traits say why a champion should win a lane, the win rate says whether it does, and when the two disagree the model correctly stops making a strong claim either way.
So carry mode is, today, an argument from judgement — the same authored traits the other eight factors read, asked a sharper question. That is worth knowing before you trust it over your own read of the matchup.
What it does not know
- Whether you can play it. Mastery and ownership are invisible to us, so the pool is every champion in the lane unless you tell us yours.
- The current patch. No tier list, and the only statistics involved are the lane and matchup win rates carry mode reads — a snapshot of Riot’s match data taken on one patch, not a live feed. The traits are judgements either way, and they move when we revise them, not when Riot ships a hotfix.
- How to compare two fits. A fit says how strongly the model prefers this champion given what it knows, so an emptier board scores higher — there is simply less that can be argued against.
- Draft order. The model sees a set of champions, not a sequence of picks, so counter-pick timing and comp archetypes are not modelled.
A 👍 or 👎 on a recommendation is how this gets better — it is measured against real drafts, not against a benchmark we invented.