How to read it. The bar runs from the 50% line: right and blue means the choice won more than it lost, left and red means it lost. The pale band behind each bar is the 95% interval, so a wide band is a thin sample. n is how many times the choice was observed; Ο is how often it gets drafted when it is available β the price the room puts on it.
The number is a value, not a win rate. Each ability is scored against
the average pick in its own slot, holding the other thirty-nine abilities in the
game and both hero shells fixed β so it answers "what does taking this do for
me", which the raw win rate does not: that one carries whoever tends to take the
ability and whatever else they took with it. Ultimates and basics are scored
inside their own group and cannot be compared across the two. Ranking on this
instead of the raw rate changes the top basic in 44% of real pools and the top
ultimate in 39%, worth about 1pp of win rate a pick. The raw rate is still in the
hover. Rows near the middle are still thin:
at the median n of 72,486 the interval is roughly Β±0.4pp. Abilities are drafted in priority order, which flatters popular picks by about +7.2pp per 100pp of draft rate β so a low-Ο ability with a high win rate is the interesting one, and a high-Ο ability below 50% is the expensive mistake.
The two sides are not equally safe. Ο divides by the matches in which the ability's hero put anything into the pool, and heroes contribute partial kits β measured 2026-09-17, the pool draws roughly 48 abilities from 10β16 heroes, not twelve whole kits. That denominator is too large, so every Ο here reads low. The cheap-and-good side is therefore optimistic: some of it is normally-drafted abilities wearing a deflated Ο. The expensive-mistake side clears 90% despite the deflation, so it is the conservative half. Read the high-Ο negatives as findings and the low-Ο positives as a screen.
AGH and SHD. 162 of the 514 abilities here have an Aghanim's
upgrade written against them β AGH for Scepter, SHD
for Shard β and the hover says what the upgrade does. Two things are worth
separating. Measured: an ability carrying a Shard upgrade is worth about a
point more than one without, and the room drafts it at exactly the same
rate (81% against 81%), while it does price the Scepter (87% against 80%). Not
measured: why . The obvious reading is that the item upgrades what you
drafted, but this corpus keeps no item data at all, so nobody here has checked
whether a Scepter or Shard was ever bought. A further 64 upgrades grant a whole
new skill β Crystal Clone, Ice Spire, Swiftslash and 61 others β and are not
marked, because each is a separate ability, none is in any draft pool, and none
has appeared once in 833,165 matches.
Shell effects. 361 of the 514 abilities on this board shift by
1pp or more with the body they land on, and those rows say so in a line under the
name: what the ability is worth on a ranged body and on a melee one β or on an
int body and a non-int one β blue above 50% and red below. Smaller shifts get no
line but still move the number. Until
you set a body the row's number weights the two by how likely you are to be on
each β the share of your pool that is ranged, or int, or of every hero before a
pool is picked β not 50/50, because only about a quarter of heroes are int, and
not by where the room drafts the ability, which is their choice and not your body. Pick a row in
Hero shells to set the body you are playing and the row takes your side's
number, underlined in the line. The hero column reads the same numbers from the
other side. A hero is one pick, like the ultimate, and its number is its own
strength plus how much better the kit you could build on it is than on an average
body from your pool: what you already hold, and the best of what is still on the
table for your open slots, each valued on that body β ranged or melee, int or not,
its own spells, body-specific combos. Its hover lists both. Heroes
other players took get the same Γ as abilities. An ability that shifts on both axes shows the
larger split, starred while the other axis is held at your pool's mix, and its
hover lists all four bodies. Every number is the adjusted value from a paired
match-level regression that removes the ability's own value, the shell's own
strength and every other ability in play. Body terms are best estimates rather
than significance calls: each is pulled toward zero by its own noise, so a thinly
measured split reads small instead of being hidden. The hover sets the raw win
rate each body actually posted, and its n, beside it.
Combos. Click any ability to put it in your kit β one ultimate,
three basics β and everything still on the table is re-priced for what you
hold; the dashed kit badge is the change, and the hover says which
pairing moved it. The Γ at the end of a row marks an ability another player took:
it leaves its column and the combo suggestions and is listed under Taken by other
players below the columns, where a click brings it back; right after the click,
an Undo note at the bottom of the screen does the same. Two layers, both gated before they ship. 34 mechanism
pairs β attack modifier with cleave +1.4pp, two crits β1.3pp; each instantiated by 28β36,010 distinct ability pairs β clear a
placebo that shuffles which ability carries which property, at a 5% false-discovery
rate against it, and rest only on tags checked by hand; the other 317 do not say
anything. 2305 named pairs ship, overriding the mechanism layer
where they exist; the ones on your table are listed above the columns. They
arrive by two routes and both are judged on wins. One screens on per-player
impact in the first half of the corpus and judges on the second, which the
screen never saw. The other nominates by how much more often players actually
take the two together β a screen that never looks at the outcome, so its judge
gets the whole corpus instead of half. The second route exists because this
page leads with a per-ability ranking, and a pair whose value is all
interaction is invisible there: Acorn Shot and God's Strength are each below
50% on their own, and about +5pp together. Between those two layers
sit 222 enablers β one ability against a whole class of partners or of bodies β which is
where Sticky Napalm lives: it adds damage to every damage instance, so it pairs
with anything that ticks fast, and above all with Ember Spirit's and Phoenix's
bodies, whose innates tick every 0.2s. Every ability also carries an own-hero term β what it is worth on its own
hero's body against bodies of the same attack type and attribute, pulled toward
zero by its own noise β and it counts once you set that body; 313 of them
move it by 1pp or more, because the innate stays with the hero here: Lina's Slow
Burn feeds her spells and Shadow Fiend's souls feed Shadowraze. The hover gives
each ability's term and the raw win rate it posted on its own hero. 1144 combos clear 5pp on top of their
parts and are marked broken . 39 of those were only ever seen
on one body and apply only when you set it: Ball Lightning with Arcane Orb is a
combo on Outworld Devourer, whose innate stays with the body in this mode and
refunds mana on every cast, and shows nothing on the 21 Storm Spirits it landed
on. Every number is an empirical-Bayes estimate from the half of the corpus
the screen never saw: the combo's own held-out result, pulled toward what the
screen's picks deliver on average by however noisy it is. Fitted on one quarter
of the data, these estimates predict the next quarter at slope 0.99; the raw
held-out gaps manage 0.35, which is why none of them is printed. The hover says
how sure the data is that each combo clears the broken bar.
Hero shells are picked too. Players choose their body from the pool, and the data shows it: the two heroes left over each game run 1.4pp weaker as shells than the ten taken, and shells that get picked more often win more (r = 0.53) β so shell win rates carry the roomβs preferences, like everything else here. The skill side of that selection is tightly bounded: mean owner strength varies 0.12pp across shells, within noise of the 0.03pp expected from chance alone. The popularity side is not bounded, and it flatters the shells the room wants to play.