⚡ All features are free during our launch period — no account required to explore.
⚡ All features are free during our launch period — no account required to explore.
Research
403 professional Dota 2 matches analyzed across 5 roles. These are the correlations that power DOTApulse's AI coaching engine. Every threshold shown has been validated against real match data — not conventional wisdom.
The Strongest Signal We Found
An offlaner with fewer than 8 total assists at end of game wins only 11.2% of games. The cliff starts even earlier — fewer than 6 assists is essentially unwinnable at 4.5% (198 samples). With 8+ assists, win rate climbs to 71.8% — a 67-point swing from the critical zone.
This holds across all three tournaments independently and is the largest win rate gap of any single metric across all five roles.
Total Assists at End of Game : Win Rate
X-axis shows total assists — not KDA. A "6-7" bar means the player finished the game with 6 or 7 total assists.
Red = below 50% win rate. Gold = above 50%. Dashed line at 50% (even odds). Hover bars for sample size.
* 6-7 bucket win rate derived from cumulative thresholds rather than direct measurement.
WHAT THIS MEANS FOR YOU
If you play Position 3, your job is not to farm — it's to be present. Join every smoke, every teamfight, every objective. Show up for Tormentor. TP to defend. Follow up on your team's initiation.
Assists are a proxy for one thing: were you there? The data says if you weren't there enough times, your team loses nearly 90% of the time. There is no stat in Dota more punishing to ignore.
The Pos 3 Absolute Floors — Never Won Once in 403 Matches
Teamfight participation below 0.3 (30%): 0.0% win rate
Across every match in the full dataset, an offlaner who participated in fewer than 30% of their team's teamfights has never won. Not in ESL Birmingham, not in Wallachia, not in DreamLeague. This is the most absolute floor in the dataset — not “very unlikely to win” but statistically zero.
Tower damage above 8,000: 100% win rate (n=18)
The ceiling mirrors the floor. Offlaners who dealt over 8,000 tower damage won every game in which they did. While small-sample (18 games), it illustrates the same principle from the other direction: an offlaner who is winning fights and converting them into objectives is essentially unbeatable.
Carry Lane Efficiency: The Hard Floor
The threshold moved from 75% to 70% in the expanded dataset — the 75% cutoff showed a weaker signal (37.5% win rate) while 70% is the sharper break point.
Carry Lane Efficiency : Win Rate
70-79% win rate derived from known aggregate 70%+ (54.3%, n=258) and known sub-buckets.
WHAT IS LANE EFFICIENCY?
Lane efficiency measures how much gold you earned in the laning phase compared to the maximum possible gold available in your lane. A score of 80% means you captured 80% of the creep gold available to you.
It accounts for last hits, denies, and whether you were present in lane or rotating/dying elsewhere. A carry who leaves lane early, dies repeatedly, or misses last hits will have low lane efficiency even if the game feels normal.
WHAT THIS MEANS FOR YOU
Below 70% is a critical failure — not just “a rough lane.” You are handing the enemy team a structural advantage that is extremely hard to recover from.
Focus on: staying in lane longer, avoiding unnecessary deaths, last hitting under tower, and not rotating unless it's a guaranteed kill. Your team needs you to reach that 80%+ threshold before the game can be won.
Pos 5 Deaths: The Cliff
Staying alive is a positive performance signal, not just the absence of a mistake. Every support death removes vision, save potential, and map presence simultaneously.
Total Deaths at End of Game : Win Rate
X-axis shows total deaths for the hard support across the full match.
Win rate collapses after 9 deaths. Dashed vertical line marks the cliff point. 7-8 death range (~48%) falls close to the league average.
WHAT THIS MEANS FOR YOU
Every time a hard support dies, three things disappear simultaneously: vision, save potential, and map presence.
You don't need to play scared — you need to play smart. Position behind your cores. Don't contest runes you can't win. Use your save items before you die, not after. If you're dying more than 9 times, you're not supporting — you're feeding.
Mid CS at Minute 10: New Finding
Mids who reach 60 CS at the 10-minute mark win 55.9% of games. Those who fall below win only 41.5% — a clean, high-sample gap across 608 player records.
Mid CS at Minute 10 : Win Rate
WHAT THIS MEANS FOR YOU
CS at 10 is the mid lane equivalent of carry lane efficiency — it measures whether you controlled the lane or were disrupted by the enemy. Missed rune trips, poor equilibrium, or taking unnecessary fights all show up here.
60 CS at 10 minutes is achievable with one rune trip and solid last-hitting. If you're consistently falling short of this, the root cause is almost always rune trips (leaving lane at the wrong time) or equilibrium management (wave pushed too far under tower to last hit safely).
Pos 4 First Blood: Early Game Signal
First blood involvement is a positive indicator — a real but moderate edge, not a dominant signal. The gap is meaningful (7.3 points) but the effect is less dramatic than early analysis suggested when the sample was smaller.
Pos 4 First Blood Involvement : Win Rate
WHAT THIS MEANS FOR YOU
The soft support sets the tempo of the early game. First blood is a positive signal, but the bigger picture is fight involvement throughout the game — winning Pos 4 players average 18.6 assists vs 10.1 for losers.
This doesn't mean force a fight you can't win. It means: have a plan for the first 2 minutes. Know your kill threat. Know your lane matchup. Coordinate with your offlaner or mid for an early kill opportunity — if it's there, take it.
Map Geometry
Across 403 matches: Radiant 52.5%, Dire 47.5%. The finding was consistent across all three tournaments independently (52.0% in ESL One Birmingham, 52.9% in PGL Wallachia Season 8, 51.9% in DreamLeague Season 29), suggesting this is a stable structural signal rather than noise.
The 52–53% Radiant advantage mirrors what is observed in public matchmaking data, which is surprising at the pro level — teams at the highest skill bracket still show the same map-side skew. This points to structural map geometry advantages (Roshan position, high-ground angles, jungle pathing) rather than player-level decision-making differences.
Coaching implication: When drafting or planning rotations, Dire teams should be more conservative about early aggressive plays that depend on map control, and should prioritize securing Roshan vision aggressively given the geographic disadvantage.
A Note on Side Selection in Pro Play
Unlike public matchmaking where sides are assigned randomly, professional tournaments use a structured side selection process. The loser of a coin flip chooses between first pick or side selection in Game 1. After each subsequent game, the losing team chooses their advantage for the next game — pick order or side.
This means pro teams are actively and strategically choosing Radiant when they believe it advantages their draft. Despite this deliberate optimization by elite teams, Radiant still wins 52.5% of games — consistent across all three tournaments independently (ESL Birmingham: 52.0%, PGL Wallachia S8: 52.9%, DreamLeague S29: 51.9%).
The implication is significant: the Radiant geometric advantage — ancient positioning, jungle camp layout, Roshan pit angle — persists even when the world's best teams are actively trying to account for it. This is not a pub matchmaking artifact. It is a structural feature of the map.
New Findings — Expanded Dataset
These signals emerged from analysis of the complete dataset and have been incorporated into the coaching engine.
Mid Lane XPM: Strongest New Mid Signal
XPM captures everything that matters for mid: farm, rune control, level advantages, and fight presence. A mid who finishes below 550 XPM has failed the core tempo function of the role. The threshold tightened from 600 to 550 with the expanded 403-match dataset — the signal is cleaner and the sample size nearly doubles.
Mid XPM : Win Rate
Carry GPM Floor: The Hard Minimum
Carries finishing below 550 GPM have never won a match in 50 samples across ESL One Birmingham, PGL Wallachia S8, and DreamLeague S29. Below 500 GPM: still 0% across 17 samples. This is the hardest floor in the entire dataset — no other metric produces a complete zero across 50+ samples.
Win rate by carry GPM threshold across 403 matches. Zero wins recorded below 550 GPM.
WHAT THIS MEANS FOR YOU
550 GPM is not a “high” bar — it's the survivability floor. Falling below it means your farm rate was so low that the game was already structurally lost. If you hit this threshold, it's almost never fixable mid-game.
The implication: carries need to treat GPM not as a vanity stat but as a binary pass/fail. Below 550 and your team has essentially been playing 4v5 from a resource standpoint. Focus on early-game GPM: jungle timing, safe lane control, and minimizing deaths before the 15-minute mark.
Carry Tower Damage: Strongest Carry Signal
This is the single most lopsided threshold in the entire dataset for any role. A carry who does not convert their power spike into structure damage is farming for themselves, not for the game.
Carry Tower Damage : Win Rate
WHAT THIS MEANS FOR YOU
Tower damage below 2k doesn't mean you played badly. It almost always means you never had the game state to push — your team was losing fights, losing lanes, or your carry items were delayed. It confirms a loss, not creates one. But if your power spike arrived and you still didn't have 2k tower damage, you were farming when you should have been sieging.
Mid Rune Control: Strongest Threshold in the Dataset
Of all the threshold analysis run across 403 matches and all five roles, rune pickup count produces the cleanest separation between winning and losing mids. Winners averaged 8.8 rune pickups; losers averaged 6.1. The 31-point gap between the bottom bucket (under 7 runes: 37.4%) and the top (10+ runes: 68.9%) is larger than the gap for XPM, CS at 10, or hero damage. Rune control isn't just good mid play — it's the single best predictor of whether a mid laner wins.
Mid Rune Pickups : Win Rate
Rune pickups counted across the full match duration. Bounty, power, and shield runes included. “7-9” bar reflects the 7+ threshold average — the 10+ bucket separates further.
WHY RUNE CONTROL MATTERS
Every rune you pick up is gold, XP, or tempo — and every rune the enemy picks up is the same denied to you. At the pro level, the gap between the most and least active rune-contesting mids is 2.7 runes per game. That delta compounds into item timings, fight windows, and ultimately game outcomes.
The coaching implication: mid players should treat contested rune fights as mandatory, not optional. Going out of position for a rune is almost always correct — missing 3 runes across a game costs more than a single death.
Mid Rune Pickups
10+ runes: 68.9% win rate
The strongest single threshold in the full dataset. Winners averaged 8.8 rune pickups vs 6.1 for losers. Below 7 runes: 37.4% win rate — a 31-point gap. See the dedicated rune control section above.
Hero Damage Threshold
Carry — under 20,000 damage:
34.3% win rate
Mid — under 20,000 damage:
32.9% win rate
Both roles show the same cliff. Carry above 20k: 61.2% win rate. Mid above 20k: 59.5% win rate. Hero damage reflects fight presence — if you're under 20k, you were absent from fights or never reached your power spike.
Carry APM
300+ APM: 61.6% win rate
Below 300 APM: 37.2% win rate. Low APM indicates idle time between actions — farming patterns and attack-move mechanics directly drive this metric.
Offlane Teamfight Participation
Below 40% fight presence: 10.9% win rate
Nearly unwinnable. An offlaner missing teamfights is not generating assists, not creating space, and not converting fights into towers — the three core functions of the role.
Offlane Tower Kills
2+ tower kills: 74.5% win rate
3+ tower kills: 82.2% win rate. Tower kills separate offlaners who convert fight wins into map control from those who win fights and stall the game.
Pos 4 Healing
3,000+ healing: 61.7% win rate
Below 3k healing: 50.3% win rate. High healing means active fight presence with save items deployed — Force Staff, Lotus Orb, and Urn all contribute.
Pos 5 Healing
5,000+ healing: 77.3% win rate
3,000+ healing: 70.5% win rate. Healing output confirms the hard support was alive, in fights, and deploying save tools — all three simultaneously.
Buyback as Losing Indicator
Used buyback: 26.5% win rate
No buyback: 56.5% win rate. Buyback usage is a symptom of being in a losing game state, not a strategic tool. Never frame it positively.
Net Worth Gap by Role (Winners vs Losers, End of Game)
Net worth gap between winning and losing players at the same position, averaged across all 403 matches.
Advanced Signals
These signals required the full 403-match dataset to surface. Some confirm instincts. Others challenge them directly.
Vision Quality: Dewarding, Not Warding
The point-biserial correlation between total observer wards placed and win rate is effectively zero. Placing more wards does not predict winning. What does predict winning is deward efficiency— killing the enemy's wards. Hard supports with more than 4 observer kills win 65.5% of games. The story the data tells: winning teams deny vision, not just provide it. A ward placed and left alive is as useful to the enemy as to you.
Wards Placed
≈ 0.00
Point-biserial correlation with win
Placing more wards does not predict winning or losing. The metric is essentially noise.
Observer Kills > 4
65.5%
Win rate when observer kills exceed 4
Killing enemy wards — not placing your own — is the vision metric that predicts winning.
THE VISION INSIGHT
Buy sentry wards. Contest high ground vision before objectives. When the enemy places an observer ward in a key location, removing it is worth more than placing your own in a generic spot. Vision control is active, not passive.
Courier Kills: High-Leverage Microplay
Courier kills are among the highest-leverage early actions in Dota — they deny gold, items, and tempo simultaneously. Mids who secure even one courier kill show a 59.6% win rate, well above the 50% baseline. Hard supports with more than 1 courier kill win 71.1% of games — the second-strongest vision-related signal after deward efficiency. The asymmetry is real: the cost of chasing a courier kill is small; the reward when it connects is enormous.
Mid — Any Courier Kill
59.6%
Win rate when mid secures 1+ courier kills
At the pro level, the mid lane courier kill is often set up by the hard support with a level 1 ward on the enemy courier path. One coordinated kill shifts the entire lane phase.
Pos 5 — 2+ Courier Kills
71.1%
Win rate when Pos 5 secures 2+ courier kills
A hard support consistently threatening courier kills forces the enemy to delay item deliveries, scout courier routes, and buy courier items defensively — all significant tempo losses.
The Pos 4 Stacking Paradox
This is one of the most counterintuitive signals in the dataset. Conventional Dota coaching tells soft supports to stack camps for their carry. But Pos 4 players who stack more than 5 camps win only 40.2% of games — below the 50% baseline.
The interpretation: over-stacking is a reactive behavior of players who are losing. When a game is going badly — lost fights, missed rotations, enemy ahead — a soft support who “can't fight” defaults to stacking camps. The camps themselves aren't bad. But a Pos 4 stacking 6+ camps is a player who was unable to be where the game actually needed them. The metric is a symptom, not a cause.
1-3 camps stacked
52.4%
win rate (baseline range)
5+ camps stacked
40.2%
win rate (losing indicator)
Note: stacking 2–3 camps as part of normal rotation is good play. The signal emerges above 5 stacks — where volume of stacking indicates the player was absent from fights.
Combination Metrics: When Two Signals Compound
Individual metrics are informative, but when you combine a carry's two strongest independent signals — XPM (farm pace) and tower damage (objective conversion) — the spread becomes the largest in the entire dataset. An 84-percentage-point gap separates carries who hit both thresholds from those who miss both.
84pp spread between best and worst carry outcome combination. Farm pace gets you to your spike; tower damage proves you used it.
THE CARRY MANDATE
GPM gets you to your item timing. Tower damage proves you used it. A carry who farms to full items but never pushes is playing a private farming game — not winning Dota. The data is unambiguous: 3.6% win rate when both signals are absent is not recoverable regardless of other factors.
Comeback & Throw Analysis
From 403 pro matches — ESL One Birmingham 2026 + PGL Wallachia S8 + DreamLeague S29 · Patch 7.41. The Opus analysis mapped every comeback (deficit overcome) and every throw (5,000+ gold lead lost) across the full dataset.
When Comebacks Actually Happen
In 128 games that ended before 30 minutes, there were zero comebacks. Not one team that was behind at minute 20 won before the 30-minute mark. As games extend, the deficit-to-win window opens: in games reaching 50+ minutes, the comeback rate rises to 15.9% — roughly 1 in 6 games.
When comebacks did happen, the key variable was teamfighting, not farming. Comeback winners averaged a teamfight participation rate of 0.673 — substantially higher than baseline. The data says you can't farm your way back into a deficit at the pro level. You can only fight your way back.
Comebacks Under 30 Min
0
out of 128 short games
Comeback Rate 50+ Min
15.9%
~1 in 6 long games
Avg TF Participation
0.673
for comeback winners
Radiant's Advantage Vanishes in Long Games
Radiant wins 83.3% of games that end before 25 minutes. By 50+ minutes, that advantage has inverted to below 50% at 46.3%. The structural map geometry that favors Radiant in early fights and objective control is gradually neutralized as both teams reach full itemization. Dire teams that fall behind early have a rational strategy: survive, extend, and equalize.
Radiant Win Rate by Game Duration
Radiant's 52.5% overall win rate is the average of a 83.3% early-game advantage and a 46.3% late-game deficit. The cross-over point is approximately 45 minutes.
Throws Found
16 / 302
5.3% of matches
Games where a team held a 5,000+ gold net worth lead and still lost. Throws are rare — but when they happen, the pattern is consistent.
Average Flip Speed
7.9 minutes
from peak lead to momentum flip
Leads don't erode slowly — they collapse. The average game went from peak advantage to losing momentum in under 8 minutes.
Tower Kill Signal
94%
15 of 16 throws
Tower kills happened near the peak advantage in nearly every throw. The most dangerous moment is right after winning a fight and pushing high ground.
Buyback Signal
71%
10 of 14 tracked throws
A high-value core died without buyback in 71% of throws. When you're ahead after minute 25, dying without buyback is the single highest-risk play you can make.
⚡ Fast Collapse
Flips in 2–4 minutes
One bad teamfight, no buyback, sudden swing. The team wins a fight and immediately overextends — pushing high ground or chasing kills deep. The enemy buybacks, picks off two cores in the chaos, and the game flips before anyone can react.
🐌 Slow Bleed
Lead erodes over 10–15 minutes
The team gets ahead but stops converting. No high ground pushes, no Roshan follow-up, no throne pressure. The enemy farms back quietly. Classic “can't close” pattern — the lead is real but never cashed in, and late-game itemization slowly equalises the game.
Key Insights
94% of throws happen while towers are falling — the most dangerous moment is right after winning a fight
71% of throws involve a high-value core dying without buyback — when you're ahead, always have buyback after minute 25
50% of throws happen around a Roshan fight — Roshan is highest-risk when you're already winning
The average throw takes only 7.9 minutes from peak to momentum flip — leads evaporate faster than you think
Every throw in our dataset was by Dire — consistent with Radiant's 52.5% map advantage
DIRE THROW FINDING
100% of throws in our dataset were by Dire. Combined with the 52.5% Radiant win rate finding, this suggests Radiant's geometric map advantage doesn't just help them win — it helps them close out games even when behind. Radiant teams facing a deficit may be more structurally capable of mounting comebacks, while Dire teams with leads may face additional friction converting those leads into wins.
Draft Synergies
Synergy analysis requires a minimum of 10 co-appearances to filter noise. Pairs that meet this threshold show genuine portfolio-level signals that pro teams' drafting already reflects — but that are now quantified.
Top Hero Pairs by Win Rate (minimum 10 co-appearances)
Kez + Phoenix
Phoenix ultimate creates safe setup windows for Kez engage
Doom + Tiny
Doom silences key targets; Tiny burst closes immediately
Batrider + Tiny
Classic lasso into toss — still one of the cleanest kill combos in pro play
Notable Anti-Synergy
Pangolier + Tiny
36.4%Despite both being aggressive melee cores, Pangolier and Tiny appear to create mechanical conflicts when drafted together — competing for the same team fight windows and requiring similar positioning. At 11 co-appearances, the 36.4% win rate is a genuine statistical signal, not sample noise.
ON DRAFT RESEARCH LIMITS
403 matches is sufficient for individual player metrics but borders on small for hero-pair analysis. The pairs shown here meet the n≥10 threshold and are directionally reliable — but hero-specific thresholds should be treated as preliminary until the dataset expands to 1,000+ matches per patch.
Hero Performance Anomalies
These findings highlight cases where conventional “good game” metrics diverged from outcomes. Each represents a coaching insight that raw win rates alone wouldn't surface.
Sniper (Mid)
10% win rate
Sniper in the mid lane won only 10% of games in the dataset despite being picked into seemingly favorable matchups. The data suggests pro teams have solved positioning around Sniper — or that the hero's slow item timing in mid creates a predictable weak window that elite opponents consistently exploit.
Terrorblade
28.6% win rate — highest tower damage in dataset
Terrorblade produces the highest tower damage numbers of any carry in the dataset, yet wins only 28.6% of games. The anomaly suggests a selection effect: Terrorblade is drafted into matchups where winning requires reaching the late game — but those games are already structurally difficult. High tower damage is a symptom of a game that went long, not a cause of winning.
Enchantress — Position Matters Enormously
Pos 4: 69.2% vs Pos 5: 22.2%
The same hero played one position apart produces a 47-point win rate swing. Enchantress as soft support (Pos 4) wins 69.2% of games. As hard support (Pos 5), 22.2%. This is the largest position-specific split in the dataset and suggests Enchantress's strength comes from having farm to build tempo items — not from a pure support role.
Kez (Mid)
72.7% win rate (Pos 2)
Kez in the mid lane is the highest win rate mid hero in the dataset with sufficient sample size. The hero's ability to create and exploit fight windows aligns with what the data shows separates winning mids: rune control (for tempo) and active fight presence (for hero damage). Kez in the mid appears to naturally incentivize both behaviors.
What the Data Corrected
Here's what the data actually showed.
CS at 20 minutes (carry)
Winners and losers nearly identical: 202 vs 199 CS at 20 minutes. The difference is statistically meaningless. CS@20 is removed from carry coaching metrics.
Offlane damage taken
Winners and losers averaged within 1% of each other across all damage taken buckets. No threshold showed a reliable win rate difference. Damage taken is removed from offlane metrics.
Courier kills (mid)
Exactly equal: winners averaged 0.32 courier kills, losers averaged 0.32. The metric is symmetric — courier kills reflect early aggression patterns, not mid lane performance. Removed.
Camps stacked (soft support)
Reverse correlation — losers stacked slightly more than winners. The interpretation: players who can't fight stack camps instead. Stacking is a symptom of not being able to rotate, not a performance signal. Removed.
Offlane stun duration
Win rate sat at 50–52% regardless of stun seconds dealt. Flat across every threshold tested. Stun duration has no meaningful correlation with offlane win rate — the metric is removed from coaching feedback.
Offlane Blink timing
We initially thought a minute-18 Blink was the key threshold. At 403 matches, there's essentially no win rate difference at any threshold before minute 22. Having Blink matters. When exactly you get it (before minute 22) doesn't.
Mid lane Blink timing
Same story. The 13–18 minute "sweet spot" we saw at 20 matches disappeared entirely at scale. Sample noise.
Pos 4 save items as a winning indicator
At 20 matches, winners had 2× more save items. At scale, losing Pos 4 players actually completed slightly more — they were building Force Staff and Glimmer reactively while behind. The correlation was backward.
Pipe being a winning offlane item
At 20 matches, zero losing offlaners built Pipe. At scale, losers built it just as often. Pure sample noise.
Last hits at 10 minutes (mid and carry)
Pos 2 LH@10 showed no signal at any threshold — 46.9% win rate below 40 CS at 10, essentially identical to baseline. Pos 1 LH@10 below 60 shows 48.0% WR (152 samples), also indistinguishable from noise. Early last hit counts are a much weaker predictor than sustained GPM or XPM over the full game. LH@10 has been removed from coaching feedback for both roles.
Context
Public tools like DatDota, STRATZ, and Dotabuff provide raw statistics — hero win rates, pick rates, team records. What they don't publish is role-specific behavioral threshold analysis: the exact point at which a metric crosses from “losing” to “winning” territory, validated across a large sample of elite matches.
Top Tier 1 organizations employ dedicated analysts who run proprietary versions of exactly this analysis internally. That institutional knowledge has never been publicly available.
DOTApulse publishes it.
Our target audience is the serious player who doesn't have an org behind them — the top Immortal grinder, the semi-pro in open qualifiers, the coach running a Tier 2 team on a budget. The gap between what pro analysts know and what individual players can access is real. We're closing it.
Future Research Plans
Bracket-specific benchmarks (Divine, Ancient, Legend) for GPM/XPM/CS calibration — making coaching feedback relevant at every skill tier
Expanded tournament dataset as new Tier 1 events are indexed
Hero-specific threshold analysis — does the offlane assists signal hold equally for initiators vs. durable heroes?
Patch-over-patch threshold drift analysis — how do signals shift when Valve makes major changes?
Methodology
DOTApulse analyzes professional tournament matches exclusively for threshold validation. Elite coordinated play produces clean, role-faithful data — players honor their positions, drafts are intentional, and execution reflects genuine strategic decisions.
Public matchmaking — even at Immortal rank — introduces significant noise: off-role picks, smurfing, uncoordinated drafts, and players optimizing for individual stats over team outcomes. This chaos dilutes signal and produces thresholds that don't reflect how Dota is actually supposed to be played.
Future plans: We intend to run separate bracket-specific analyses (Divine, Ancient, Legend) to calibrate GPM/XPM/CS benchmarks for each skill tier — making coaching feedback more relevant to players at every level.
Sample size caveat: 403 professional matches is a meaningful dataset for role-level patterns but insufficient for hero-specific thresholds. These findings apply to role behavior, not individual hero performance.
Consistency check: Key findings (offlane assists, Radiant advantage, Pos 5 deaths cliff) were validated independently across all three tournament subsets before being reported as findings.