Guide

Game ranked placement and calibration systems explained

Harbor Arena’s Season 4 reset looked generous on paper: every player who finished Diamond or above received a Gold III placement seed regardless of last season’s peak. Veterans who had grinded Immortal spent their first ten ranked games deleting opponents rated 800 MMR lower. New players who won three of five placements still landed in Silver because the algorithm averaged visible rank, not hidden skill. Churn telemetry showed 44% of players who lost their first two post-placement matches never queued ranked again that season — not because they hated the game, but because the ladder lied about where they belonged. Meanwhile smurfs on fresh accounts hit Platinum in six games while matchmaking still believed they were Silver.

Ranked placement (also called calibration, provisionals, or placement matches) is the short onboarding window that estimates a player’s hidden MMR before steady-state matchmaking takes over. It is distinct from the visible rank badge players show off: placement is where uncertainty collapses fast. Done well, it puts you in fair games within a dozen matches. Done poorly, it poisons retention for a whole season and floods support with “I got placed too low” tickets. This guide covers the placement FSM, K-factor inflation and streak rules, carry-over from prior seasons, integration with decay and season resets, smurf acceleration, the Harbor Arena refactor, a technique decision table, pitfalls, and a production checklist.

Hidden MMR vs visible rank during placement

Competitive ladders almost always split two numbers:

Hidden MMR (matchmaking rating) — a continuous skill estimate the server uses to find opponents. High confidence = small rating swings per game; low confidence = large swings.
Visible rank (tier, division, LP) — the badge on the profile. Often quantized into Bronze through Challenger with promotion series between tiers.

During placement, MMR should move quickly while visible rank catches up or stays hidden entirely. A common failure mode is tying placement outcomes only to visible tier averages — e.g. “win 3 of 5, get Gold” — without updating MMR proportionally to opponent strength. A win against Immortal-rated players should jump MMR more than a win against Bronze.

Uncertainty is the third variable. Glicko and TrueSkill expose a rating deviation or sigma: wide during provisionals, narrow after calibration completes. Matchmaking during placement should prefer opponents near your current MMR estimate but with compatible uncertainty bands so games remain informative.

The placement FSM: states, gates, and completion

Treat placement as a finite-state machine separate from the main ranked loop:

UNRANKED — account eligible but no games played; optional skill survey or last-season prior applied.
PROVISIONAL — N games (typically 5–15) with inflated K-factor; visible rank may show “?” or a wide band.
CALIBRATING — optional extended window if uncertainty remains above threshold after N games (streaky players, parties).
PLACED — MMR sigma below cutoff; visible rank revealed or LP accrual begins at tier mapped from MMR.
STEADY — normal K-factor; full integration with decay rules and promotion series.

Gates between states:

Complete minimum games (e.g. 10) and sigma < threshold
Or hit maximum games (e.g. 20) — force placement at best estimate
Abandon or AFK during provisional may count as loss without revealing rank early

Party placement is harder: queue as a group should calibrate on party MMR (often inflated mean + variance penalty) and complete only when all members exit provisional or when the party leader’s uncertainty dominates.

K-factor inflation, streaks, and informative opponents

In Elo-style systems, the K-factor controls how many points change per game. Steady-state ranked might use K=16–32; provisionals often use K=80–160 or multipliers (e.g. 4× until 10 games).

Design rules that improve calibration speed without overshooting:

Opponent-strength weighting — beat higher MMR, gain more; lose to lower, lose more. Never flat +25 per win.
Streak dampening — five wins in a row against similar-rated foes suggests underrating, but cap daily MMR delta to prevent smurf rockets from one lucky afternoon.
Loss streak floor — after three losses, widen matchmaking search slightly so the next game is winnable-in-principle; avoids death spirals that abort calibration.
Performance subscore (optional) — in team modes, adjust K slightly by contribution metrics (damage share, objective score) only when correlated with win probability; avoid rewarding stat-padding.

TrueSkill’s update naturally shrinks sigma each game; you may not need manual K inflation if sigma-driven match quality is tuned. Hybrid stacks (TrueSkill mu/sigma internally, Elo-like LP externally) are common in AAA shooters.

Season carry-over, soft reset, and re-entry placement

New seasons interact with placement in three patterns:

Hard reset — everyone reprovisions; high churn risk for veterans unless MMR priors carry hidden skill forward.
Soft reset — compress MMR toward population mean (e.g. pull everyone 30% toward 1500) but keep ordering; short reprovision or skip if sigma still low from last season.
Visible-only reset — LP and tier drop, hidden MMR barely moves; placement is cosmetic. Fast but can mismatch badge and match quality if MMR drifted during offseason.

Harbor Arena’s mistake was a visible seed without MMR prior: returning Immortal players got Gold badges but their hidden rating was still ~3200 while matchmaking during provisionals treated them as ~2100. The fix applied a season carry-over prior:

MMR_start = α × MMR_end_last_season + (1 − α) × population_mean

with α = 0.75 for players with 50+ ranked games last season, plus a 3-game reprovision only if peak tier exceeded calibrated tier by two divisions. New accounts still ran full 10-game calibration from mean + optional survey offset.

Coordinate with decay systems: players who decayed from inactivity should re-enter provisionals with higher sigma, not the same narrow band as active grinders.

Smurf detection and acceleration during calibration

Fresh accounts with veteran skill are a placement problem and a fairness problem. Signals during provisionals:

Win rate and performance z-score vs cohort after 3–5 games
Mechanical inputs (headshot %, reaction latency) where privacy policy allows
Hardware fingerprint, payment instrument, or phone verify (controversial)
Queue dodge and leaver history on linked accounts

Acceleration means ending provisionals early when confidence exceeds threshold — place a smurf at true MMR in 4 games instead of 10. Clamping caps how high a new account can calibrate without identity verification (e.g. max Platinum until SMS link). Balance acceleration against false positives: a player returning on a new platform is not always a smurf.

Harbor Arena refactor: from fixed seeds to prior-driven calibration

After Season 4 churn spiked, Harbor shipped:

Prior map — last-season peak tier → MMR prior with α compression; removed flat Gold seed.
Dual-track display — provisionals show “Calibrating (7/10)” instead of misleading tier icon.
Opponent-weighted updates — replaced win-count tier table with per-game TrueSkill updates.
Party sigma lock — parties calibrate on leader sigma until all members placed or 15 games max.
Telemetry — track “expected win% vs actual” over first 20 ranked games post-placement as calibration quality score.

Outcomes two patches later: misplacement churn 44% → 11% (players who quit after two losses in first week); average games-to-fair-match (|expected win% − 50%| < 8) fell from 18 to 9; support tickets tagged “wrong rank” down 62%. Smurf time-to-true-tier dropped from 12 to 5 games without raising low-skill player stomp rate in provisionals.

Technique decision table

Scenario	Prefer	Avoid
Brand-new ranked mode	Full provisionals from population mean	Importing casual MMR blindly
Annual season reset	Soft MMR compress + short reprovision	Hard reset with fixed visible seed
Returning veteran	Prior from last season peak + α blend	Same calibration as new account
Party queue	Leader/party MMR with shared sigma rules	Calibrating solo members independently
Suspected smurf	Accelerate + optional verify cap	Ban without review pipeline
Low population region	Widen search, fewer required games	Forcing 10 games at 8 min queue each
Cross-play ranked	Separate pools or input-adjusted priors	Single pool with no device prior

Common pitfalls

Visible rank without MMR movement — badge lies; matchmaking breaks.
Flat win-count placement — ignores opponent strength.
Same seed for veterans and newbies — stomps or sandbags.
Ignoring party uncertainty — duos calibrate at solo pace.
No abandon policy in provisionals — dodging rerolls easy tier.
Over-long calibration — 20+ games before ranked “counts” feels grindy.
Under-long calibration — three games leave sigma too wide for steady state.
Decoupled from decay — inactive players skip reprovision they need.

Production checklist

Define placement FSM states and transitions (min/max games, sigma cutoff).
Separate hidden MMR updates from visible tier reveal timing.
Apply inflated K or TrueSkill sigma reduction schedule for provisionals.
Weight game outcomes by opponent MMR and expected win probability.
Implement season carry-over prior with documented α compression.
Handle party placement with shared or leader-weighted uncertainty.
Wire smurf acceleration with human-review appeals path.
Count abandons and AFK as losses during provisionals.
Telemetry: calibration quality score over first 20 post-placement games.
Align with decay and promotion series so players never see conflicting ranks.

Key takeaways

Placement estimates hidden MMR — visible rank is secondary during provisionals.
FSM + sigma cutoff — know when calibration ends.
Inflated K and opponent weighting — converge fast without flat win tables.
Season priors — soft reset beats fixed tier seeds for veterans.
Harbor Arena misplacement churn 44% → 11% after prior-driven calibration.