The redesign: one product, two numbers, two jobs

Hidden Skill Rating for fairness. Visible Season Rank for motivation. Don't cross the wires.

Everything so far points at one conclusion, and Convocados' own ADR already states it: the failure wasn't a bad formula, it was asking one number to do twojobs. ADR 0031separates them — Skill Rating is hidden and drives balancing;Season Rank is visible and drives motivation. This page makes both concrete.

	Skill Rating	Season Rank
Job	Predict performance; balance teams	Make progress felt; give a goal
Audience	Organizer, matchmaking	Every player
Visibility	Hidden (admin only)	Visible to members
Horizon	Lifetime; never resets	Per Season; soft-resets
Engine	Glicko-2 / OpenSkill	Transparent point payouts
Failure to avoid	Unfair teams	"+0 after a win"

1 · Choosing the Skill Rating engine

Convocados rates individuals whose teams are reshuffled weekly — that is the team-inference problem from the uncertainty page, where plain Elo is weakest (it applies a team result to each individual as if the team were one entity).

Engine	Uncertainty	Individual from team result	Cost
Elo (today)	No	Naive (winner takes all)	Lowest
Glicko-2	Yes (RD, σ)	Per-player, head-to-head	Moderate
OpenSkill / TrueSkill	Yes (Gaussian)	Yes — the point of it	Moderate–high

Recommendation. Move the hidden Skill Rating toOpenSkill (fallback: Glicko-2). It models 5-a-side team results and per-player uncertainty — the two things Elo fakes. Keep the change invisible: it only feeds balancing, so a bad model shows up as unfair teams, not as player confusion. Gate the migration behind a recompute so history stays consistent.

2 · The Season Rank blueprint

Unit: Rank Points (RP), whole ladder ×10 — seeds, payouts, and tier edges all in RP, so the maths is unchanged but the numbers are worth chasing.
Payout: K · (S − clamp(E, 0.1, 0.9)). A win always pays at least +30 RP; even games pay most; upsets pay more.
Provisional: an unlock counter (not a hard 3-game cliff), with the fast-K window driven by evidence.
Seeding & reset: seed from hidden Skill Rating; soft-reset 50% between Seasons so a Season is a fresh race without discarding skill.
Tiers: percentile edges frozen at calibration, expressed in RP.
Feedback: a per-game Rank reveal at score time (before → after, progress to next tier), plus the existing tier-transition modal.

3 · Play a season

Play a season

Your skill · 1720

Field average · 1000

Payout

Tier band (illustrative) · 1500

Starting RP = 7200 · Start tier = Diamond

Play the eight-round Season.

Switch between current and Rank Points and play a few seasons. Current quietly pays the favourite almost nothing per win; Rank Points pays every game — which is the entire motivation argument, made measurable.

4 · Does it answer the two failures?

"Static Elo is demoralizing." Skill Rating is hidden; the public ladder is Season Rank, which soft-resets — everyone starts the race near each other, and improvement is visible per game.
"Two numbers confuse players." Players see exactly one: Season Rank (plus the Crew ladder). Skill Rating is an organizer tool.
Residual risk to watch: a skill-anchored reset still favours stronger players over a Season. That is honest — the fix for feeling stuck is visible per-game progress and tiers, not equalising ability.

Check yourself

The reason to run two systems rather than one is that they serve different…

Purposes and audiences. Two jobs — fair teams versus felt progress — pull in opposite directions.

OpenSkill suits Convocados' Skill Rating because teams are…

Reshuffled every week. Weekly reshuffles make the team a noisy observation of each player.

The visible Season Rank should never show a player…

A hidden Skill Rating. Skill Rating is hidden; exposing it recreates the static-number problem.

Primary sources

Re-read the repo's ownADR 0031(the settled model this redesign delivers), then theOpenSkill paper for the engine andMenke's GDC talk for the rank-as-experience layer.