There's no fixed pick count where a champ-select read "starts meaning something" — instead Should I Dodge widens its standard deviation continuously as picks drop below ten, by the square root of ten over picks known: ×1.291 at six picks, ×1.581 at four. So a four-pick reading needs to sit nearly 60% further from 50% than a full lobby's to earn the same "High" or "Low" label, because the four still-empty slots carry real variance, not an assumed 50% average.
The intuition, and why it's backwards
Score a lobby with only six of ten champions picked and it feels like it should be a more confident read, not less — fewer unknowns, a cleaner signal, less noise to average over. Should I Dodge does the opposite: the fewer picks it has, the wider it makes its bands. That's not a bug or a conservative fudge factor. It's the correct answer to a question that's easy to ask wrong.
The question that gets asked, and the one that matters
The tempting question is: "if I sum six champions' log-odds instead of ten, how spread out is that six-champion sum?" Since the app's win-probability output is a sum of one log-odds term per champion, a smaller sum built from fewer independent terms genuinely does have less variance than the full ten-term sum — that part of the intuition is correct in isolation. Answered honestly, that question implies narrowing the standard deviation by a factor of roughly the square root of six-tenths, not widening it.
But that's the wrong question for what these bands are actually judging. A six-pick reading during champion select isn't a final answer about a six-champion game — it's an early, incomplete look at a lobby that is still going to end up with ten champions in it. The four empty slots aren't going away; they're going to be filled by real picks with their own real win rates, and those picks haven't happened yet. Scoring the unfilled slots as if they contribute nothing — effectively treating them as a flat, average 50% — throws away the fact that they still carry the same variance any of the other six champions did. A partial reading is a noisy early estimate of where the eventual full-lobby number will land, not a tighter, smaller-scope measurement in its own right. Narrowing the bands as picks come in would have the app delivering its loudest, most confident verdicts exactly when the fewest picks are locked in — precisely backwards. So sigma widens as the champion count drops below ten, by the square root of ten over the pick count, floored so it never goes below the full-lobby value once all ten are in.
"Picks known" here means recognised champions across both teams combined, not per side — a lobby with three allies and three enemies locked in is a six-pick reading, the same as one with five allies and one enemy. Should I Dodge will score a lobby with as few as a single recognised champion; there's no minimum beyond that, only a wider and wider band as the count drops, right up to a lopsided one-sided read that carries the least evidence of all.
How much wider, exactly
The scaling factor applied to a tier's full-lobby standard deviation, at various pick counts, is the square root of ten divided by the number of champions actually known:
- 10 picks known: ×1.000 (the calibrated, full-lobby value — never adjusted further)
- 8 picks known: ×1.118
- 6 picks known: ×1.291
- 4 picks known: ×1.581
That scaling is the same regardless of tier — going from a full lobby down to four known picks always widens the standard deviation by about 58%, whether the tier's baseline is a tight 4.40 or a loose 12.23. Applying it to a few representative tiers' full-lobby sigma gives the widened standard deviation at each pick count, and half of that widened figure is exactly how far from 50% a reading has to sit to clear the Low/High boundary — the same 0.5 sigma cut used at a full lobby, just measured against a wider sigma. Really Low/Really High sits at three times that distance (1.5 sigma), so it scales identically.
| Tier (baseline sigma) | 10 picks known | 8 picks known | 6 picks known | 4 picks known |
|---|---|---|---|---|
| All tiers combined (4.78) | 2.39 | 2.67 | 3.09 | 3.78 |
| Diamond (4.40) | 2.20 | 2.46 | 2.84 | 3.48 |
| Gold (5.45) | 2.73 | 3.05 | 3.52 | 4.31 |
| Challenger (12.23) | 6.12 | 6.84 | 7.89 | 9.67 |
Concretely: at Diamond, a full ten-champion lobby only needs to reach 52.20% to register as High. A four-pick early read needs to reach roughly 53.48% to earn that same label — not a huge difference in absolute terms, because Diamond's baseline spread is already tight. Do the same comparison at Challenger and the gap is much larger in absolute points: a full lobby clears High at 56.12%, but a four-pick read needs to clear roughly 59.67% — almost four full points further out, on top of Challenger's baseline spread already being the widest on the board. See the full boundary table by tier for every tier's full-lobby cuts, and why Challenger's baseline is so wide in the first place for why that combination compounds the way it does.
A label that changes as the lobby fills in
The Really Low/Really High boundary sits at three times the Low/High distance (1.5 sigma against 0.5 sigma), which makes it possible for the same raw reading to carry two different labels depending purely on how many picks are locked in. Take a 58% reading at Diamond. Off a full ten-champion lobby, Diamond's really-high cut sits at 56.60% — 58% clears it, so the label is Really High. Off just four known picks, Diamond's widened standard deviation pushes the high cut out to roughly 53.48% and the really-high cut out to roughly 60.44% — 58% clears the first but not the second, so the same number reads as plain High instead. Nothing about the lobby changed between those two readings except how much of it was known; the label moved because the app is correctly less willing to call an early number "really" anything.
What this means while you're actually drafting
In practice this means treating an early read the way you'd treat an early poll result: a genuine data point, not a settled outcome. A 55% reading with four picks locked in is weaker evidence than the same 55% once the lobby fills out to ten, even though the app is looking at the same underlying quantity both times — see how to read champion select as picks come in for the non-numeric signals worth tracking alongside the number while the draft is still moving. It's the same underlying principle as the tier-level noise floor: less evidence means more caution, not more confidence, and the honest response to a thinner sample is a wider margin of doubt, not a narrower one.
The practical rule of thumb: don't trust a "High" or "Low" label on a partial lobby as strongly as the same label on a full one, and expect the gap between them to matter more at wide tiers than tight ones. See how the calculation and its bands are built for where the partial-lobby scaling sits inside the full pipeline.