TFT CODEX

Every number here came from a game someone played.

What is behind the figures on every other page: where the boards come from, what a comp is, what the floors are, and what Riot’s match data does not contain.

Rebuilt every two hours. Updated 19 Nov 2026, 15:18 UTC.

Where the numbers come from

We sample ranked Teamfight Tactics matches through Riot’s match-v1 API and store every final board whole: champions, items, star levels, traits, the level the player finished on, and where they placed. The boards are kept, so a question nobody has asked yet can still be answered later.

Every board carries a rank bracket. Where the sampler already held the tier of the player whose board it is, that is the tier on the board; otherwise it is the tier of the player the match was discovered through. Diamond+ — Diamond, Master, Grandmaster and Challenger — is the slice every page opens on, and Emerald+, every game we sample, is one click away: both numbers ship inside the page, so the switch swaps text rather than loading anything. Most of those games are Diamond+ as well, and we hold too few games below Emerald to say anything about those ranks.

Every number is from the patch being played now. The window opens 2 days after Riot publishes a patch’s notes, once every region is on it, and never covers fewer than 3 days: in a patch’s first days it reaches back into the patch before, so no number stands on a day or two of games. Comp numbers cover at most the last 7 days of it. Champion, item and trait numbers pool across every board, so they carry far larger samples than any single comp does, and cover at most 14 days. Today both cover 3 days. Everything is rebuilt every two hours.

Ability, trait and augment numbers for Set 18 are read straight out of Riot’s own TFT game files after each patch, because CommunityDragon’s export carries the sentences without them; the text itself comes from CommunityDragon.

What a comp is here

A comp is a family of boards, not a name. Boards are clustered into cores, and a core sharing all but 2 of its units with a larger one is a version of that one: its boards count for the family, and the comp page lists it as a variation with what it adds and what it leaves out.

Within a family, a champion on 60% or more of its boards is core — that is the board the page draws. Between 25% and 60% it is flex: a slot people fill differently from one game to the next.

Under a champion on the final board are the champions its boards play in that slot instead. Which slot a champion takes is measured, not guessed: on the boards that ended at the final board’s level, we compare how often each drawn unit is there with how often it is there beside that champion. A board has as many slots as its level, so a champion that comes in pushes about one unit out, and it goes under the unit it pushes out most — when that unit is at least 20 percentage points rarer beside it. A champion is only asked about when it is on 10% of those boards, or is one of the comp’s own flex units, and on 50 of them either way; one that pushes out no single unit is not put under any of them.

Each family has an id that does not change, and its name comes from its own boards — the carry it plays, and the trait it reaches furthest. A name is kept for as long as it still holds: its trait among the family’s top three and the champions it names on the board, one of them carrying. Only then is it renamed, and the old address points at the page of its name for good. Identity is rebuilt from 14 days of boards, because a comp that drifts a unit a day should not become a different comp overnight. A family whose carry leaves is a new comp with a new id; the old address points at the page of its name, or of the comp that now counts its boards.

The floors

A comp is published at 50 games and graded at 400. A grade also needs the comp on 1 board in 100 of its bracket’s boards over the days comp numbers cover; a comp that already has a grade keeps it down to 1 in 125, so one sitting at the edge does not gain and lose its letter between refreshes. That is why a comp with thousands of games can show no grade: it is measured and shown, but played too rarely to rank. A comp whose core holds fewer than 5 units is not graded either. A champion, item or trait row needs 400 boards, and the Diamond+ split on those pages needs 400 of its own. Rows under a floor are kept and counted; they are shown without a number.

A grade is set on the conservative average — the top of the 95% interval around the mean, mean + 1.96σ/√n — so a thin sample sinks instead of spiking: S to 4.25, A to 4.50, B to 4.75, C to 5.00, and D above that.

Placement has a standard deviation of about 2.29, so a 163-board average carries a 95% interval of roughly ±0.35 — wider than the gap between two grades. That is what the floors are for.

Placement against level

Riot’s match data keeps the board a game ended on, and the level it ended at nearly decides where it placed: a game that reached level 9 lasted long enough to get there. So a comp’s average partly says when it completes. Beside it, each comp page and both lists show vs level: the comp’s average minus what every board at its final levels placed over the same days — negative, it placed higher than its levels alone predict. Level is partly how a game went and partly the plan (a reroll comp stops at 7 or 8 on purpose), so this is context beside the grade and never part of it. The boards every comp is measured against:

Final levelAvg placeBoardsDiamond+ avgBoards
101.8830,3321.8728,777
93.96170,0703.96161,900
84.95150,8734.95142,040
75.6458,1525.6354,035
67.1810,0497.189,405
57.929207.93885
48.0018.001

Over the last 3 days, every board filed under a comp.

A board at every level

A comp page shows the comp’s board at every level from 4, and only some of those boards are measured. Riot’s match data keeps the board a game ended on, so the levels a comp’s games end at — usually 8 to 10 — are measured: the units on at least half of the boards that ended there, beside the share of games that did. Nothing records the boards on the way, so the levels below are planned by one rule, and marked plan wherever they are shown:

Read upwards, that is an opener on the comp’s own traits, a unit a level, and one roll into the board the comp is played for. A planned board has no average placement: nothing records how it did, so none is shown.

Some lower levels do not need the rule. Where at least 100 games went out at level 5, 6 or 7 still holding 4 or more of the comp’s final units, that level shows the board those games ended on instead — the units on 30% or more of them — marked real. They are real boards on the way to the comp, not the comp’s own games.

Under a measured level are its options: what the comp’s games put in the slots past that level’s core, on the games that went out at that level holding the whole core, this patch. An option is the whole set of units in those slots — with two or three slots to fill, the combination people actually played together, never each slot’s top unit put side by side. Each shows its games, its average place and how often it won, best placed first, beside the level’s own average; the level’s board draws the first in a gold ring, and the most played is named among them. Every game compared went out at the same level, so an option is not flattered by the head start of having got there. A level needs 100 such games before anything is ranked, and an option 30 of them and 3%; before ranking, each average is pulled toward the level’s own by 30 games’ worth, so an option on a few dozen games cannot top the list on luck. The final board’s flexible slots follow the same rule.

The level above a comp’s top row gets a row of its own, marked best: that row’s whole board and the champion the games that reached the next level added to it, best placed first. Those games are counted by the comp’s core rather than by which comp they were filed under, because a level-10 board has usually grown past the core it was matched on. It is still a comparison between games that got to the same level: what they placed with, not what caused it.

How a comp is played

The chip beside a comp’s name is measured, not chosen. A comp rolls at the lowest level where one of its 1-, 2- or 3-cost champions is 3-starred on at least 50% of the boards that ended there (with 100 such boards at least), and the chip names the cost and the level: 1-cost reroll at level 6. Where every champion it 3-stars is a Riftbeast it says Riftbeast shops instead: that trait turns a shop into Riftbeasts every three combats, so those copies come from the trait’s own shops rather than from rolls. Every other comp is labelled by how high its players are when they go out in stage 6: Fast 10 at 9.6, Fast 9 at 9.0, Fast 8 at 8.4, and Level 8 below that. A comp page’s Pace line prints that level for each stage with 30 boards or more.

Riot’s last_round is the round a player went out on. Stage 1 has four rounds and every later stage seven, so round 5 is 2-1 — and almost nobody goes out on a carousel round, which is how we know. Player pages print it as Out at 5-3.

Roll plans and the calculator

Match data records no rolls, no gold per round and no shops; the sites that print roll counts read them from desktop trackers, which nothing on a phone may do. What Riot does publish, in the game’s own files, is how a shop is drawn: the odds of each cost at each level, and how many copies of each champion the pool holds. A comp page’s Roll line is a model built from exactly that — 1,500 simulated roll-downs at the level the comp rolls at (or 8 or 9 for a comp that levels), from a two-star to a three-star for a reroll and from one copy to three for a carry, with a Wisp in every other shop’s last slot — and it is marked model wherever it appears. The roll calculator uses the same rules and works the chance out exactly instead of simulating it.

The rest of the lobby

We keep all eight boards of every match we sample, so a comp page can say how the comp placed when nobody else in the lobby held its carry, when one other player did, and when two or more did; each of those needs 50 games to print. Part of the gap is the lobby rather than the copies, so the page says the comp placed that way, never that it did because. The builds under a carry are its full three-item builds, each on at least 3% of the carry’s full builds and 20 games, with their own average placement; the components are what its most played builds are made of. Play it when names those components, and adds “nobody else is on its carry” only where the comp placed at least 0.25 better that way than against two or more.

How we rate difficulty

The Difficulty chip under a comp’s name counts four things we measure, a point each:

No point is Easy, one is Medium, two or more is Hard. The line beside the chip prints the number behind each point. A signal we have no numbers for scores nothing, and a comp with none of them gets no label. The last one comes from the lobby table above, so it says the comp placed worse: part of that gap is the lobby rather than the copies.

What is not measured

Riot’s match data reports the final board only. It does not say which augments a player took, where units stood, or what a board looked like before the last round. So there are no augment win rates here and no measured positioning: where units stand on the drawn boards, and the boards below the levels games end at that are not marked real, are planned by stated rules, and say so. The augment reference lists what each augment grants, from Riot’s own data files, and carries no numbers.

A comp page’s Augments that fit puts every augment through the rule the overlay uses when it reads the three on offer, and says which it is: measured for an augment that gives a trait, beside the comp’s own placement at that trait’s levels; by rule for an attack-damage or ability-power augment against the damage its carry’s items are built for, a reroll augment against whether it 3-stars a champion on 50% of its boards, and an XP or gold augment against an average level of 8.9 or more. No augment has a number of its own.

One queue is sampled — ranked — and the numbers are not split by region.

Who is named

Riot IDs are public, shown in every lobby and on Riot’s own leaderboards, and we resolve them only for Challenger, Grandmaster and Master. Each name on the leaderboard opens that player’s recent games as our sampling caught them — placement, board, level — and their LP across our daily snapshots; those pages exist only while the player is on the board and are kept out of search engines. Everyone else in our data is a one-way pseudonym we could not name if we wanted to. The privacy policy says exactly what is stored, and how to be removed.

Who makes this

TFT Codex is an independent fan project by a solo developer, not made or endorsed by Riot Games. Questions, corrections and removal requests go to support@tftcodex.com.

What changed

Recent changes to the site and the app, newest first.

Rule changes

When the way we count or grade last changed, newest first.