VTSR v4 VT Stats

VTSR v4: Both Sides of the Coin

A running commentary on and analysis of the VTSR-T performance rating and the VTSR-C commander Elo based on continued research, player feedback, and empirical observations. The "both sides of the coin" references this document's attempt to weigh both the input from the community and the industry standards.

Personal note from Sev: this page is obviously drafted with AI-assistance, but each section is based on, or serves as a precursor to, my own work and research - and I have personally reviewed every section of this document. Every number on this page comes from python scripts/validate_elo.py (§1–§21) run on 2026-10-22, and anyone with the repository can regenerate all of them. The site shows VTSR-T as a band (a tier and its Upper or Lower half), never as a number, so this page does too: where a player is named, their VTSR-T appears as a band, and point figures appear only as averages over many games. Every design change is governed by a rule written before the numbers that judge it existed, as statistics dictates we must. BZCC is an obscure and unique game, with a tiny and intriguing playerbase. Quite humorously, AI often falls short of understanding it since it's training is often based on games with only a subset of BZCC's game mechanics, with playerbases in the millions. The trick is figuring out how to leverage what we know about the subsets, in order to accurately assess the system as a whole. That being said, BZCC didn't invent anything new, it's simply an interesting combination of existing concepts (FPS + RTS), with enough variation to never quite become stale - perhaps the sole reason we still play the game 25+ years later.

Quick answers

Commonly asked questions. Every answer links to the numbered section that carries the numbers, and to the ELO page where the mechanics are explained step by step.

Why don’t 1v1s and 2v2s count?
VTSR-T scores you against the room you played in. With two players in the room every stat comes out at exactly ±0.5 no matter what happened, and with four the spread is noise — the math has nothing to measure. The games are not worthless: the side with the stronger thugs won 14 of the 19 we have a result for. So the plan is a separate small-format record first, and a small-format ladder once there are enough results — the way StarCraft II, chess and Counter-Strike’s Wingman each handle a different format.
Details: §10 Include all games? §13 Roadmap, item 6
Why isn’t win/loss part of VTSR-T?
It is computed: a win/loss ladder runs beside the performance rating and updates on every decided match (237 so far). Its weight in the published number is zero. A team result gives every player on a side the same credit, so in a 5v5 it says little about one thug. Results also barely repeat for the same thug: half-vs-half reliability is 0.30 for win rate against 0.95 for the performance score, on the same 1,473 games. Mixing it in has been tested three times (2026-09-04, 2026-10-13, 2026-10-25) against a rule written in advance. It never called more winners. On the latest 236 matches, a 10% or 25% blend changes no call, and a 50% blend changes one and gets it wrong. The commander rating, VTSR-C, is pure win/loss.
Details: §6 Why win/loss is not in VTSR-T §5.1 Performance vs wins §13 Roadmap, item 2 ELO · How it works
Why did my rating drop when we got stomped, or when my commander welaped?
Because you are measured against everyone in the lobby, both teams, and a stomped team scores below the room on almost every axis. Your team’s result explains about a third of your per-match score, and spending a third of the match on foot at base waiting for a ship is, on average, the steepest loss on the page. Tier 2-and-up thugs on a losing team score about even with the room (P −0.042) and lose rating 84% of the time, because the bar their rating set was higher. The fix on the roadmap is to extend the ship-denied-time credit to every tier.
Details: §5.2 Team result §5.3 Time on foot §13 Roadmap, item 5 ELO · Commanders & fairness
Why not StarCraft-style faction ratings? Is Hadean overpowered?
Hadean does win more than its commanders’ and thugs’ ratings explain — about 44 commander-rating points on the latest full-sample fit (z 1.70, p 0.089) — and its pick share has roughly tripled in a year. But StarCraft’s per-race MMR works because it is 1v1, you pick your own race and there are millions of players; here the commander picks the faction for the whole team and there are 47 rated players, so three ladders would leave most people provisional forever. What transfers is the chess idea that the matchup itself carries an edge: one faction term in the commander ladder’s prediction, frozen at +60 Hadean and judged only on games played after this document.
Details: §7 Faction-based ratings §5.6 Faction is not neutral §13 Roadmap, item 1
Why isn’t there a support / utility rating?
Every role rating that lasted has the role declared before the match (Overwatch); every one that guessed the role from stats died (Dota 2’s Core/Support split in seven months, Riot’s positional ranks after one preview). Here your role is whatever ship and weapon the commander handed you, and a role the system has to guess is a role players can pose for. The data also says the premise is half right: PvE work is already rewarded — the winning team has more of it 80% of the time. Accuracy and efficiency rewarded clean fighting and agreed with winning about 39% of the time, so they left the composite in v2.12 and v2.13. That was a weights change on this ladder.
Details: §8 Role-based ratings §5.4 Four axes remain ELO · The 4 axes
Does commanding punish my rating?
Not on average: commander rows and thug rows beat expectation about equally often (42.5% vs 43.4%), and commander rows gain slightly more per game on average, because the commander adjustment lowers the bar. The catch is that the bar is set to the average commander, so strong fighters harvest it (blue beats expectation in 79% of his commander games against 33% as a thug) while a few still bleed (Lithium 20% as commander against 52% as a thug). Measuring each commander against their own thug numbers instead is roadmap item 3. Your win/loss record as commander lives in a separate rating, VTSR-C, where your thug rating is not touched.
Details: §5.5 Who commands §8 Role-based ratings §13 Roadmap, item 3 ELO · Commanders & fairness
Why no per-ship rating?
The commander builds your ship, it changes several times a match, and per-player-per-ship samples are far thinner than even per-faction ones. The one study that looked found champion-specific skill mattered in League of Legends and added little in Dota 2. Accuracy is not in the rating at all: it left the composite in v2.12. The verdict: the Loadout tab shows your per-ship record, and it never becomes a ladder.
Details: §9 By-ship ratings
What is the difference between VTSR-T and VTSR-C?
VTSR-T (thug) is a performance rating, not a win/loss Elo: four in-match statistics scored against the lobby you played in, so it moves when you play better or worse than expected, and winning the game does not move it. The site shows it as a band, never as a number. VTSR-C (commander) is a classic win/loss rating between the two commanders, with a handicap for how strong each side’s thugs were. Two ladders is the genre’s own answer — Allegiance and Natural Selection 2 both split the commander from the field player — and it is why commanding does not touch your thug rating.
Details: §1 Where we are §6 Why the design is defensible ELO · How it works ELO · VTSR-C ladder
Can the site balance the teams for us?
On request, yes. Tools → Team Balonce has a Suggest button. It tries every ordered pair of commanders in the lobby and, for each pair, every split of the remaining players into two even teams (one extra player on a side only when the lobby is odd, with the empty seat scored as a floor-rated thug) — 90 pairs and 70 even splits for ten players — scores each layout with the same commander-ladder formula that calls 64.2% of 730 duels (commander VTSR-C gap plus the gap in mean thug VTSR-T), and keeps the one closest to a 50% win chance; ties go to the smaller rating-sum gap, then to fewer people moved. Read it as one fair answer, not the answer: in a typical ten-player lobby about half the commander pairings have a split within half a point of even, so a six-point rating change — and ratings move after every pushed match — can swap the pick for an unrelated layout that is just as fair. What the site still cannot do is matchmake: with 20–25 active players, volunteer commanders and hand-split teams the ratings can only describe the room, and the research says perfectly fair matches are not even the engagement optimum, and that even paying people to take the unpopular role fades within weeks.
Details: §11 Matchmaking Tools · Team Balonce
Is this real statistics, or just what an AI said?
Don’t take either on trust; check it. Every number on this page comes out of a committed script you can run: scripts/validate_elo.py, scripts/study_wins_signal.py, scripts/vtsr_t_variants.py and scripts/study_pve_denial.py, on the processed data in the repository. Each rule carries a date before its results: the wins-blend rule was written 2026-09-04, before the first sweep, and its one amendment on 2026-10-25, before that sweep was read. Both are in critique/decisions/phase-5-wins-blend.md. Each mechanism has a published source, row by row in §2. AI tools drafted some proposals. Several of them are on §3’s held and refuted lists, rejected by the same gate as everyone else’s. For one worked case with the algebra, the data and the commands, see why win/loss is not in VTSR-T.
Details: §6 Why win/loss is not in VTSR-T §2 Where every mechanism comes from §3 How a number becomes a rule
How do I check any of these numbers myself?
Clone the repository and run python scripts/validate_elo.py. It rewrites _validation/report.md: §15–§19 are the numbers in section 5, §20 is the small-format test, §10 the commander ladder, §11 the axis table. Every design decision is a memo under critique/decisions/, and every outside claim on this page has a numbered source.
Details: §3 How a number becomes a rule §15 Sources cited ELO · Does it work?
What is actually going to change?
Only what passes its pre-registered rule. Queued: the faction term in the commander ladder (31 of the 60 confirmation duels are in), a fresh test of blending win/loss into VTSR-T (225 determined matches now, against 39 last time), and two what-if versions of the thug rating — the commander adjustment against your own baseline, and ship-denied-time credit for all tiers — plus a small-format record. Snipes, T-key, accuracy, and efficiency have already left the thug composite. Nothing on the do-not-do list is coming back.
Details: §13 Roadmap

1. Where we are

287 recorded matches, 248 rated, 47 rated players (22 of them on the VTSR-T ladder), 225 rated matches with a determined winner; F9bomber's community duels bring the commander ladder to 730 scored duels. [76]

Spearman ρ (R → P)
0.288
v3: 0.462
Self-consistency
0.712
v3: 0.804
Winner from team mean R
56.3%
n=224 · v3: 43.3% (n=30)
VTSR-C duel prediction
64.2%
n=730 · log-loss 0.629

Every validator run since August is a point. Winner prediction from team mean rating climbed from 37.5% on 32 determined matches to 59.1% on 210, then the four-axis re-rate reset it to 55.1%; it is 56.3% on 224 now. The commander ladder went from 55% on 38 duels to 64.3% on 705 once the F9 ledger and the mechanically-applied outcome pairings landed in mid-September, and holds at 64.2% on 730. [77]

Calibration error is 0.028; on the eight-axis rating it had reached 0.009. Mean-rating aggregation is 56.3%, hard-MAX 57.6% and softmax 55.4%, all inside each other’s error bars, so the v3-era preview that hard-MAX would win clearly has not held. Eleven matches have a team gap big enough to shift the win chance by about 14 points or more, and 10 of those 11 were called correctly — encouraging, still a small bucket. On the same-corpus retirement reprocess, winner prediction stayed flat (57.8% to 58.3%) while Spearman ρ moved from 0.412 to 0.275 and self-consistency from 0.787 to 0.672 (critique/decisions/thug-efficiency-retirement.md). The cards above are the later corpus.

2. Where every mechanism comes from

Nothing in VTSR is invented here except the four axis definitions and their weights — and those are exactly what §5 puts under test. Each row: a VTSR mechanism, and the shipped system or peer-reviewed method that uses the same idea.

Logistic expected score, 1500 anchor
Chess Elo [1]; the Bradley–Terry model it approximates [2]
K falls with games, rises with inactivity
Glicko / Glicko-2 rating deviation [4][5]; used by Lichess and US Chess [7]
Rating from in-match statistics, not only W/L
TrueSkill 2 (Microsoft Research / The Coalition): kills, deaths, quits, squads; Halo 5 prediction 52% → 68% [9]; HLTV Rating 2.0 [23]; PandaSkill [27]
Everyone measured against the room they played in
Elo-MMR for massive multiplayer contests [13]; PandaSkill's free-for-all OpenSkill update [27][11]
Separate commander and thug ladders
AllegSkill on Microsoft Research's Allegiance: “separate skill ratings for commanders and pilots” [50]; Natural Selection 2 Hive Skill 3.0 [58]
Commander role adjustment
Baseball WAR's positional adjustment — a catcher and a first baseman on one scale [19][20]
Team-strength handicap in the expected score
Bradley–Terry order effect: home advantage, white's first move [3]; Aligulac's matchup-constrained SC2 ratings [48]
Spectators, partials and idlers left out, not penalized
TrueSkill 2's quit modelling [9]; Hive 2.0's team-switch handling [60]
A zero-sum W/L ladder beside the performance ladder
Classic Elo [1]; the two-ladder structure of AllegSkill [50] and Hive 3.0 [58]
Validation by out-of-sample prediction
The Cambridge CS:GO benchmark [31]; TrueSkill 2's own test [9]; WHR [12]
Rules written before the data that judges them
Registered Reports in the sciences [74][75]
Never reward underperformance
Elo-MMR's incentive-compatibility guarantee [13]; opponent indifference [14]

3. How a number becomes a rule

Hypothesis → a rule written in a memo before any number exists → a forensic what-if rating file beside the real one → the validator scores both → PROMOTE / HOLD / DISCARD → only PROMOTE changes a published rating. Every proposal since v2 has gone through it. [79]

Shipped
  • Inactivity K-boost (v2 §7.1)
  • Row-exclusion gates v2.5–v2.10
  • Wins-ladder machinery (inert at α=0)
  • F9bomber ledger import
  • VTSR-C economy machinery (inert at αc=1)
  • Ladder eligibility + 90-day inactivity (display)
  • Balonce Meter (the validated VTSR-C formula)
  • Axis retirement v2.11–v2.13 (snipes, T-key, accuracy, efficiency)
  • VTSR-T shown as sticky bands, never as a number (display)
  • Balonce main reason + stacked-team record (display)
Held
  • ALPHA wins blend — three times
  • Rank-based lobby scoring — three times (June review)
  • VTSR-C αc economy scoring
  • Commander VTSR-T in the handicap
  • Phase 6 faction term (31 of 60 confirmation duels)
  • Reconstructed accuracy — HOLD (53.8%, Wilson lower bound under 0.50)
Refuted / stopped
  • Hard-MAX opponent reference (v2 critique) — ρ 0.46 → 0.19, +522 points of drift
  • Softmax opponent reference
  • Unlocked commander priors (v2 critique) — below noise floor
  • Converged re-rating
  • Softmax / MAX thug handicap — −2.5 to −11.7 pp

Several held and refuted entries were drafted by automated reviewers — the v2 production-systems critique, the June 2026 review, and this document's own exploratory faction fit — and went through the same gate as a proposal from the developer. The data overrules the tooling. That is the point of the gate.

The industry has not converged either. Valve split Dota 2 into Core/Support MMR in August 2019 [38] and merged it into one rating with per-role handicaps in March 2020 [40]. Riot previewed positional ranks and ended the preview after satisfaction fell 20–30 points [42]. Blizzard kept per-role SR in Overwatch because the role is locked before the match [44] and added per-race MMR to StarCraft II six years after launch [36]. NS2's developers said separate team and commander skills were “scrapped because there is no good way to implement any of it” in 2016 [61] and shipped exactly that in 2020 [58].

4. What changed since v3

Shipped (memo-gated, additive)
  • Wins ladder RW — real, 225 determined matches, blended at α = 0
  • F9bomber ledger → VTSR-C v3 (505 duels, full K); 50 outcomes adjudicated mechanically from F9 pairings under six gates
  • VTSR-C economy composite + opening-window amendment, αc = 1
  • Ladder bars (25 matches for VTSR-T; 8-or-25 duels for VTSR-C), 90-day inactivity, 3-game comeback — display only
  • Balonce Meter: one win-probability formula for Tools and the Outcome card
  • Input corrections v11–v14 (each moved ratings ≤ ~1 point)
  • VTSR-T bands replace VTSR-T numbers on every surface: a sticky tier and half (20-point promotion buffer), a fuzzy position bar, Beat expectation / On par / Below expectation verdicts for a game, and team gaps as their effect on win chance. VTSR-C keeps its numbers. Display only: ratings are byte-identical
  • The wins-weighted ladder preview and the VTSR-T The Champion career card are retired
  • Balonce main reason: each prediction names why the favorite is favored (commander edge, thug edge, both, even, or stacked — a better commander against much stronger thugs). Over 293 classifiable duels the favorite won 102 of 153 on a commander edge (66.7%), 49 of 76 when that edge faced stronger thugs, and 62 of 88 with both edges; stacked favorites won only 4 of 9. Shown in the Outcome card, Tools and the ELO page
  • Validator through §21, including the reconstructed-accuracy candidate
  • Axis retirement v2.11–v2.13: snipes, T-key, accuracy, efficiency leave the composite (schema 17, four weight keys)
Tested, not shipped
  • ALPHA ∈ {0.10, 0.25, 0.50}: HOLD three times (2026-09-04, 2026-10-13, 2026-10-25). On the latest 236 matches, no blend calls more winners (56.4% at best, same as α = 0); ρ and log-loss improve slightly as α rises
  • Rank scoring: HOLD three times — self-consistency up, bootstrap σ worse
  • Converged re-rating: STOP — reverted
  • VTSR-C αc < 1: HOLD — no axis clears the 0.55 lower bound on 81 confirmation duels
  • Balonce T-term variants: HOLD / DISCARD — softmax degrades monotonically
  • Reconstructed accuracy axis: HOLD — sign agreement 53.8% on 225 determined matches, Wilson 47.3–60.2%

5. The complaint, quantified

The community's reading: VTSR-T is a raw performance meter that punishes strong thugs for what their commander or team did. Six findings, each a validator section. [77]

5.1 Performance ladder vs wins ladder (§15)

Across the 25 players with ≥ 20 matches, the published composite (thug_elo) and the inert win/loss ladder (wins_elo) correlate at Spearman 0.518 (Pearson 0.483). The chart puts both on the same band scale and plots, per player, how many bands higher the performance rating sits than the wins ladder: Cloaket five bands (a 29–58 record), then Snake, VTrider, Domakus, Muffin, blue and Nomad three each. Only Lithium sits a band lower on performance than on wins; HappyOtter, Monkey and Certified Bad Guy land in the same band on both.

The mean gap is about 1.7 bands (132 points on average), and 24 of these 25 players rate higher on performance than on wins (21 of them by at least a band), after efficiency and accuracy left the composite (5.4). Commander share correlates 0.27 with the gap. This is basketball's box-score-vs-plus-minus argument [18][16][17]. Most of the gap is a level offset: the two ladders sit at different levels by construction, and centred on their own means only about 10 of 25 players are higher on performance (why win/loss is not in VTSR-T).

5.2 Your team's result moves your rating (§16)

1,474 determined thug rows. Winners: mean P +0.218 and an average change of +12.0 points. Losers: −0.185 and −8.9. 23% of winning thugs lose rating; 21% of losing thugs gain. Win/loss alone explains η² = 0.33 of thug P variance.

Tier 2-and-up thugs on a losing team score about even with the lobby (P −0.042) and lose rating 84% of the time (−10.2 points on average), because their expectation was higher. The rating does not fine a strong thug for losing; it fines them for not beating the room by as much as their rating said they would.

5.3 Time on foot at base costs rating (§17)

Thug rows banded by the share of the match spent on foot inside their own base — waiting for a ship. Under 5%: an average change of +12.5 points, team win share 76%. Over 30%: −18.5, win share 14%. Spearman(share, change) = −0.52. The gradient survives inside losses (−1.4 at <5% vs −20.8 at ≥30%).

Confound, stated plainly: ships go unreplaced more often on losing teams. What this establishes is that ship denial costs rating, and that the low-tier lift (Tier 4 and below) does not reach most of these rows — the worst band’s average pre-match rating sits in Tier 3 · Lower.

5.4 Four axes remain (§11)

Across 225 determined matches the four axes still in the rating agree with the winner 88.9% (net damage), 88.4% (mobility), 84.7% (kill rate) and 79.6% (PvE share). When all four are present they are 13.3%, 33.3%, 33.3% and 20% of the score.

Snipes and T-key left in v2.11. Accuracy left in v2.12 (sign agreement 39.5%; BulletHit never recorded non-human hits). Efficiency left in v2.13 (38.8% on 188 matches, Wilson interval entirely below a coin flip; critique/decisions/thug-efficiency-retirement.md). Those rates are the frozen pre-retirement measurements: the axes are no longer in elo_history.json. Reconstructed accuracy, scored on these 225 matches, agrees 53.8% (Wilson 47.3–60.2%) and stays display-only under its promote rule. PvE share remains, and it still agrees with winning.

5.5 Who commands, and what commanding does to a rating (§18)

Commanders sit at the 47th percentile of their lobby by pre-match VTSR-T (median 0.44) and below their own thug mean 52% of the time — the middle, not the weakest. 33 distinct commanders; the top four (Sev 53, Lithium 45, Cloaket 44, F9bomber 42) hold 37% of commander rows.

Commander rows beat expectation 42.5% of the time against 43.4% for thug rows, and gain slightly more on average (+2.8 points a match vs +1.6). The v2.4 adjustment is set to the average commander's shortfall, so the chart splits: strong fighters harvest it (blue beats expectation in 79% of his commander games against 33% as a thug, Domakus 64% vs 38%, Sev 57% vs 33%) while Lithium (20% vs 52%), judgeguns and Darkvale (28% vs 46%) still bleed; 7 of these 21 players beat expectation more often as commander than as thug — the known failure mode of a population-baseline positional adjustment [19][21].

5.6 Faction is not neutral (§19)

Raw: Hadean beats ISDF 60.7% (85–55) and Scion 56.6% (64–49); Scion vs ISDF 50.0% (92–92). Controlling for both commanders' VTSR-C and both squads' mean VTSR-T (the canonical expected score as a fixed logistic offset), Hadean-vs-ISDF is +0.252 logit (z 1.70, p 0.089) ≈ +44 commander-rating points; Scion-vs-ISDF −0.021. LR χ² 3.75 (p 0.153); log-loss 0.6238 → 0.6195; accuracy 65.2% → 64.5%. F9 alone is +0.347 logit (≈ +60 points); telemetry alone is +0.084 (≈ +15). Hadean's pick share has gone from ~10% of sides in late 2025 to 37% in September and 33% in October 2026. The full-sample fit is discovery. The frozen +60 candidate is judged only on games after 2026-10-04; that confirmation sample is 31 duels, and so far the frozen term does not improve their prediction [79].

6. Why the current design is defensible

It measures something persistent

Split-half self-consistency 0.712. Rating from per-match statistics is the move that lifted Halo 5 prediction from 52% to 68% [9].

It is calibrated

Predicted vs observed lobby-relative performance differ by MAE 0.028 on the −1 to +1 score. (The Balonce Meter’s printed probability rests on something else: the VTSR-C duel formula and its measured 64.2% accuracy.)

It predicts, and improved with data

56.3% winner prediction, with 10 of 11 calls correct once the team gap is worth about 14 points of win chance; the Cambridge benchmark's best win/loss-only system reached 64.1% on ten thousand pro matches [31]. VTSR-C's 64.2% over 730 duels sits right at that benchmark.

Two ladders is the genre's own answer

AllegSkill: “separate skill ratings for commanders and pilots” [50]. Hive 3.0: a high-ranked player commanding “will no longer be expected to carry to avoid losing skill” [58].

The handicap is validated

λ = 1.0 is the shipped handicap: 64.2% on 730 duels (log-loss 0.629), against 63.4% (0.633) with the handicap off. λ = 1.5 scores 64.3% with a slightly worse log-loss.

The exclusion gates do real work

Spectators, partials, idlers omitted rather than penalized — TrueSkill 2's instinct too [9]. Top-20 Jaccard 0.71 under resampling.

The axes that disagreed with winning left

Efficiency agreed with the winner 38.8% of the time and accuracy 39.5%. Both left the composite (v2.12 and v2.13), with snipes and T-key (v2.11). The four axes that remain all agree with the winner.

The process says no to its authors

The refuted list is longer than the shipped list. A system that mostly says no behaves like a measurement system should.

Why win/loss is not in VTSR-T

Every number in this block comes from one pipeline run: 2026-10-25, 260 rated matches, 237 with a decided result, VTSR-T on its single enemy-damage axis. Other sections were measured on earlier snapshots, so a few of their figures differ.

The short answer. Win/loss is computed on every decided match: a second ladder, RW, runs beside VTSR-T and updates from team results. Its weight in the published number is zero for two reasons. With five players a side, a team result cannot be pinned on one thug. And three tests, each written down before its data was read, found that mixing it in does not call more games correctly. Win/loss does count elsewhere. VTSR-C, the commander rating, is 100% win/loss, and the Balonce Meter’s win probability is built on it.

The math: one game is one equation for ten unknowns

Give each player a hidden “contribution to winning” θi. The standard team model, the one behind TrueSkill [8], says a team wins on the sum of its members:

P(A wins) = σ(x1θ1 + … + x10θ10)

Here xi is +1 if player i is on team A, −1 on team B, and 0 if absent, and σ is the logistic curve. Four facts follow:

  • Every teammate gets the same lesson. The learning signal from one game for player i is (y − p)·xi, where y is the result (1 or 0) and p the predicted chance. It is identical for all five players on a side. The best and worst player on the winning team receive exactly the same credit.
  • One game pins down one number. Its information matrix is p(1 − p)·x xT, which has rank 1. A game tells you about the difference between the two team sums and nothing about how either sum splits between players.
  • Players separate only through rotation. Over many games the information is XTWX, and one player is distinguishable from another only through the games where their teammates differed. Two players who always play on the same side have identical columns and can never be told apart.
  • One thug is a small slice. With ten similar contributions, one player is about 1/10 of the variance of the result. In a 1v1 they are 1/2. A win says five times less about one thug than a 1v1 win says about its player.

VTSR-T avoids this by scoring a per-player statistic, each thug’s share of the lobby’s damage to enemies. One game then gives ten separate measurements instead of one shared one.

The wins ladder doing exactly that

The most recent decided match with at least three rated players a side: 2026-10-25T02-56-16, 5v5, Team 2 won. Each row is RW’s update, Δ = K·(s − e), where s is the result and e the expected score from the two teams’ mean RW.

RowseΔ RWImplied K
Team 1 commander00.4306−5.1712.0
Team 1 thugs A, B, C, D00.4306−5.17 each12.0
Team 2 commander10.5694+5.1712.0
Team 2 thugs A, B, C, D10.5694+5.17 each12.0

All ten players moved by the same amount. s and e are shared by a whole side. K is the only per-player term, and it differs only while someone is new (24 falling to 12 over their first 10 decided games). Nothing in the update can tell who played well. Source: python scripts/study_wins_signal.py, B3.

The data
A thug’s results barely repeat; their performance does

Same 1,473 thug rows, 23 players with at least 16 each, split into odd and even games. A player’s win rate in one half predicts the other half at 0.30 (Spearman–Brown; r 0.18). Their performance score: 0.95 (r 0.91). Most of what a thug’s win rate shows is not a property of the thug. Source: study_wins_signal.py B1.

Who commanded explains a fifth of a record

On those 1,473 games, the expected result from the two commanders’ VTSR-C alone (no thug rating used) correlates r 0.47 with each thug’s actual win rate (r² 0.22, 23 players). Actual and commander-expected differ by 6.6 wins per 100 games on average. This is descriptive, not causal: the commander ratings were partly built from these same games. Source: B2.

Thug ratings do feed a win/loss prediction, at team level

The VTSR-C duel model uses the gap in mean thug VTSR-T as a handicap. With it (λ = 1): 63.7% of 742 duels, log-loss 0.632. Without it (λ = 0): 63.0%, 0.635. λ = 0.5 gives 63.8% (0.630) and λ = 1.5 gives 63.0% (0.639). This is where the thug side enters win/loss: as team strength. Source: validate_elo.py §10.

No thug rating calls winners well, wins ladder included

From the thugs’ pre-match team means, every VTSR-T candidate tested on 2026-10-24 separated winners from losers at AUC 0.51–0.54 on 186 decided matches. The wins ladder alone scored 0.508; 0.5 is a coin flip. Source: vtsr-t-enemy-damage.md.

The performance score already moves with winning

Win/loss explains η² 0.23 of a thug’s per-match score on 1,561 thug rows (§5.2 shows 0.33 from the earlier four-axis rating). Winning thugs average a match score of +0.29 and losing thugs −0.17. The rating already credits winning play, player by player. Source: validate_elo.py §16.

What §5.1’s “higher on performance” is and isn’t

The two ladders sit at different levels by construction. VTSR-T drifts up: losses count at 0.85 of gains, and there is a 1000 floor. RW is zero-sum around 1500 and compressed by the shared update above. For the 25 players with at least 20 matches, the means are 1697.5 and 1500.9 and the SDs 182.4 and 54.5. Raw, 23 of 25 rate higher on performance (mean gap 197). Centre each ladder on its own mean and it is 10 of 25, mean distance 128 points (1.3 half-bands). The ladders disagree about order, not consistently in one direction. Source: B4.

The tests: mixing RW into VTSR-T, three times

Each sweep publishes vtsr = α·RW + (1 − α)·thug rating for α = 0.10, 0.25 and 0.50, and scores it on the same matches as the current rating (α = 0). The winner call is the team with the higher mean pre-match rating.

DateVTSR-T thenMatches scoredα = 0Best αLog-loss, α 0 → 0.50ρ, α 0 → 0.50Verdict
2026-09-04Multi-axis composite3839.5%39.5% (all)0.7040 → 0.70440.447 → 0.459HOLD
2026-10-13Four-axis composite19957.3%56.8% (all)0.6816 → 0.68130.277 → 0.280HOLD
2026-10-25One axis, enemy damage23656.4%56.4% (0.10, 0.25)0.67798 → 0.677320.416 → 0.427HOLD

What favours the blend: ρ (how well the pre-match rating ranks players’ next performance) rises with α in all three sweeps. Log-loss improved in the last two; it worsened slightly on 2026-09-04. Self-consistency rose in the first and last (0.850 → 0.865 on 2026-10-25) and was flat on 2026-10-13. The log-loss gain at α = 0.50 on 2026-10-25 is 0.0007 per match. What it doesn’t do: call more winners. On 2026-10-25, α = 0.10 and 0.25 change no call on any of the 236 matches. α = 0.50 changes one call (2026-06-01T02-09-58), and that one goes wrong. A 2026-09-23 run that was never written up showed at most +1.4 points (two matches of 142); it is now recorded in the memo.

What would change the answer

Written 2026-09-04, before the first sweep. α moves only if some α in {0.10, 0.25, 0.50} meets all five:

  1. Winner-call accuracy improves by at least 5 points over α = 0.
  2. Log-loss is strictly lower on the same matches.
  3. Spearman ρ drops by no more than 0.01.
  4. Calibration error (MAE) rises by no more than 0.005.
  5. The mean published rating does not rise above the α = 0 mean on the same snapshot.

Any α dropping ρ by more than 0.03 is discarded. Condition 5 originally read “less than 50 above 1500”. It was amended on 2026-10-25, before this sweep was read, because the α = 0 pool already sat at 1615.9 for reasons unrelated to the blend. Next test: after the next VTSR-T re-rate. On 236 matches, condition 1 means about 12 more correct calls.

Check it yourself

From a clone of the repository (the processed data is committed):

python scripts/validate_elo.py
python scripts/validate_elo.py --elo-mode alpha10 --output-dir _validation/a10
python scripts/validate_elo.py --elo-mode alpha25 --output-dir _validation/a25
python scripts/validate_elo.py --elo-mode alpha50 --output-dir _validation/a50
python scripts/study_wins_signal.py

Each report’s clean_win_accuracy.aggregations.mean is the winner call; _validation/wins_signal.md holds B1–B4. Rules and every result: phase-5-wins-blend.md.

Limits, stated plainly. The current rating’s 56.4% winner call has a 95% interval of 50.0–62.5% on 236 matches. That is barely clear of a coin flip, and neither rating predicts winners well from the thugs alone. The reliability and commander numbers rest on 23 players. The commander comparison cannot separate a commander’s skill from the thugs they were given. Blends above α = 0.50 were not tested; α = 1 would be the wins ladder alone, which scores 0.508 AUC. None of this says winning doesn’t matter. It says a team result is the wrong instrument for measuring one thug, and the commander ladder is where it is measured.

7. Faction-based ratings

The case for

The full-sample estimate in 5.6 is about +44 Hadean points (p 0.089). The pre-registered candidate is still the frozen +60, judged on games after 2026-10-04. Skill is partly faction-specific. Blizzard added per-race MMR to StarCraft II in 2016 because “many people want to experience the ladder as other races without affecting the rank and MMR of their ‘main’ race,” seeding off-races from the main race with high uncertainty [36][37]. Aligulac rates pro SC2 per matchup, constrained so the overall rating is the mean [48]. NS2 shipped per-team skill for truly asymmetric sides [58].

The case against

SC2 is 1v1, the player picks their own race, the population is millions. Here the commander picks the faction for the whole team and there are 47 rated players: splitting them three ways gives 116 player-faction cells with a median of 9 matches; 34 reach the 25-match ladder bar; 12 players have ≥ 10 on all three factions. Small split pools are exactly where Glickman warns ratings drift [6]. And VTSR-T is lobby-relative — both teams' faction effects sit in the same lobby and cancel — so a faction axis there measures nothing and charges thugs for a commander's choice [81].

Designs, ranked
  1. Three per-player ladders (SC2 style) — rejected on the arithmetic above; the precedent does not transfer.
  2. Per-player faction offset (NS2 “average skill + team bias” [58][62], Aligulac's mean-constrained matchup ratings [48]) — the right design if per-player faction skill is ever wanted; viable later, display first.
  3. A global faction-advantage term in the VTSR-C expected score — the recommendation and the only thing v4 pre-registers: the Bradley–Terry order effect [3], applied to the matchup, not to either player. Frozen at +60 Hadean / 0 Scion / 0 ISDF; judged on ≥ 60 confirmation duels after 2026-10-04 by out-of-sample log-loss, accuracy and sign stability [79]. Verdict today: 31 confirmation duels, still short of 60, and so far the frozen term does not improve them.
  4. A faction axis or shift in VTSR-T — rejected; lobby-relative scoring makes it meaningless.

Where both sides agree: the faction matchup table and the pick-share trend belong on the Meta tab regardless of any rating decision.

8. Role-based ratings

The case for

Thugs do different jobs; a kills-and-damage composite rates the hunter above the utility player even when the utility player wins more. Dota 2 [38], League [41] and Overwatch [44] all built role-aware ratings; PandaSkill models every League role independently [27]; skill decomposes into player, hero and player-hero components [28]. Our own commander adjustment is a role adjustment, and 5.5 shows it is miscalibrated.

The case against

Every role rating that lasted has the role declared and locked before the match. Overwatch's survived for that reason [44][45]. Dota 2's Core/Support split lasted seven months before Valve “replace[d] the Core/Support separation with a single rank” [40]. Riot's positional ranks lasted one preview: players “frustrated with the grindiness, low satisfaction of off-position games, and teammates not taking off-position games seriously”; satisfaction down 20–30% [42]. In BZCC the role is whatever ship and weapon the commander handed you — a burst-gun Xypos goes to pools, a slicer Xypos hunts [80]. A role the system has to guess is a role a player can choose to flatter their rating, and a rating you can game stops measuring [70][71][73][13]. The data also says the premise is half right: PvE work is already rewarded (80%). The axes that rewarded clean fighting have left the composite (5.4).

Designs, ranked
  1. Separate inferred-role ladders — rejected: gameable, fluid roles, the Riot precedent.
  2. Role as a declared pre-match input — not available in this game's telemetry or lobby.
  3. Recalibrate the commander adjustment to the player's own thug baseline — the WAR analogy made honest [19]; Phase 7(a) forensic alt mode.
  4. Remove the axes that disagreed with winning — done in v2.11–v2.13 (snipes, T-key, accuracy, efficiency). The four that remain all agree with the winner.
  5. Ship-denied-time adjustment for all tiers — generalize the v2.8 lift; Phase 7(c).
  6. Display-only playstyle descriptor (pools / hunter / mixed) — vocabulary, no rating consequence.

9. By-ship ratings

For

Ships are where skill shows; champion-specific skill was a significant predictor in League of Legends [28].

Against

The same study found it added little in Dota 2 [28]. The commander builds the ship; it changes several times a match; per-player-ship cells are far thinner than per-faction. Accuracy is a display column: it left the rating in v2.12, and the reconstructed-hit candidate is still on hold.

Verdict

Display-only ship profile (already the Loadout tab); ship-normalized axis baselines as a possible future input; never a ladder.

10. Include all games regardless of size?

A match rates today at ≥ 6 real players and ≥ 240 s — 3v3 and up. The proposal: drop the size gate. The author's position: 1v1 and 2v2 change the game too much. Validator §20 tests both. [77]

What is left out. 27 small-format games (1v1 ×13, 2v2 ×9, five uneven), 19 with a determined winner; 18 of the 27 are from September and October 2026. The dynamics do change: median duration 1310 s vs 939 s; kills per player-minute 0.41 vs 0.31; damage per player-minute 2203 vs 1388; PvE share 0.69 vs 0.60; deaths 0.14 vs 0.17 and peak pools 4 on both. The author's premise holds on duration, kill rate, damage rate and PvE share.

But the ratings transfer, partly. Reconstructing pre-match ratings, the side with the stronger thugs (higher team-mean VTSR-T) won 14 of 19 (Wilson 51–88%, beside its 56.3% on full games); the higher-rated commander won 10 of 19 (32–73%). A 1v1 is a fighting duel far more than a command duel.

The hard constraint is math, not taste. VTSR-T's Pi is a lobby z-score: with two rated rows every axis is ±0.5 by construction (why the v14 gate counts real rows — a 1v1 with a camera-pod gallery used to rate), with four the σ is noise. The composite cannot score these games. The wins ladder and VTSR-C are size-agnostic, so “include” is a real option for them.

The industry's answer to a different format is a separate ladder, never a mix: StarCraft II places per mode, 1v1 / 2v2 / 3v3 / 4v4 [83][84]; FIDE split rapid and blitz in 2012 because “players have different levels of play depending on the speed of the game” [85]; TrueSkill 2 treats other modes as correlated-but-distinct [9]; Overwatch 2 keeps Open and Role Queue apart [47]; CS2 keeps Wingman on its own rank [86]. F9bomber's ledger dropped its 1v1/2v2 rows at import too [79].

For including

+11% corpus, 19 more known results; real, long, combat-heavy games; many of the players in them are the mid/low-tier regulars who already have the fewest rated rows; TrueSkill-family systems rate any team size [8][10]; thug skill demonstrably carries over.

Against mixing

Composite undefined below ~6 rows; dynamics differ in the weighted axes; VTSR-C's expectation assumes a squad; every precedent separates formats; commander ratings predicted these games at coin-flip.

Designs, ranked
  1. Rate everything in the published VTSR-T — rejected on the mathematics.
  2. Small games into the wins ladder RW only — mechanically sound; invisible at α = 0; cheap forensic alt mode. Phase 7(d).
  3. 1v1s as VTSR-C duels — hold: 10/19 says the ladder would learn the wrong skill.
  4. Separate small-format record, then ladder — the SC2 / FIDE / Wingman pattern; display-only now, rated at ≥ 30 determined small games with a rule written first.
  5. TrueSkill-2-style cross-mode prior — the principled long-run version of (4).

Where the author's argument lands: “too different” is supported as different enough to rate separately, not as too different to count. Do not throw the games away; do not mix them into a room-relative composite; record them separately first.

11. Matchmaking

Twenty to twenty-five active players, commanders volunteer, teams split by hand [80]. No rating design fixes that; it can only describe it honestly.

  • Who commands: mid-lobby players by VTSR-T (5.5). The weakest-player story is not supported; the concentrated-load story is. The volunteer's dilemma is the textbook description [69].
  • Records bend under the load: two of the four most frequent commanders hold the two lowest VTSR-C (Lithium 1303 over 144 duels, Darkvale 1317 over 107); F9bomber's ledger records Lithium 26–77 as commander [76]. Skill or burden, a W/L ladder cannot tell — the argument for displaying games commanded beside the rating.
  • Fairness is not the only objective: equal-skill matching is a special case of engagement optimization that “rarely holds in reality” [65]; optimal matchmaking under losing-streak churn differs from skill-based by 4–6% engagement [66]; slightly weaker opponents reduce churn more than fair matches do [67].
  • Bribing people into the unpopular role fades: Overwatch's Priority Pass “had a positive effect for several weeks, but its influence quickly faded,” and loot-box / XP rewards for tanking could not “put a dent in the demand for playing DPS” [47]. Make commanding measured fairly and visibly credited; do not pay for it with thug rating.
  • Closed pools drift: Glickman's committee tracks a fixed cohort's mean rating [7]; the validator's mean_vtsr history (1539 → 1584 points, a pool average) is the same instrument here.
  • What the site does offer: Tools → Team Balonce’s Suggest button, an exhaustive search over commander pairs and even thug splits for the layout nearest 50% under the VTSR-C duel formula. It is a suggestion from a flat landscape: on a 3 October ten-player lobby, 50 of 90 commander pairings had a split within half a point of even, and two players who clicked it across a mid-evening ratings push got two different, equally fair layouts. The card says so after every suggestion and stamps when its ratings were computed.

12. The FPS+RTS lens

Battlezone (1998) combined “a tank simulation game, a first-person shooter and a real-time strategy game” [56]; Battlezone II (1999, remastered 2018) kept the formula [57]. The lineage runs through Allegiance (Microsoft Research, 2000) [52], Natural Selection (2002) [53], Savage (2003) [54], Nuclear Dawn (2011) [55], NS2, Eximius [64] and Silica [63]. Two shipped public skill systems — AllegSkill [50] and Hive [58] — and both separate the commander from the field player.

What makes the genre hard to rate is not that it has two jobs; it is that they are coupled. A thug's measurable output depends on the commander's supply (5.3), and the commander's win depends on the thugs' fighting. Every thug is on a team whose coach controls their equipment mid-game — which is why the composite-vs-impact debate maps so cleanly onto VTSR-T vs the wins ladder [16][17].

Thug skill is FPS skill and is measurable per match — the statistics TrueSkill 2 found predictive in Halo [9] and HLTV built a rating from [23]. Commander skill is RTS skill: economy, timing, build order. RTS expertise is detectable from telemetry [32], and adding a macro-economic Spending Quotient lifted StarCraft II league prediction from 47% to 62% [33] — the long-run case for VTSR-C's economy axes, recorded, not yet scored, judged by a confirmation sample [79].

13. Roadmap v4

Ranked by evidence available × leverage ÷ effort. Nothing below changes a published rating without a pre-registered rule passing.

1. Phase 6 — faction-advantage term in VTSR-C

Memo written; validator §19 scores it every run. Trigger: ≥ 60 confirmation non-mirror duels after 2026-10-04. Authorizes the Balonce Meter's faction inputs if promoted.

2. ALPHA wins-blend re-sweep — done, HOLD

Run 2026-10-13 (199 matches) and again on 2026-10-25 on the single-axis rating (236 matches). Every blend failed the +5-point accuracy condition. Next trigger: the next VTSR-T re-rate. Why win/loss is not in VTSR-T.

3. Phase 7(a) — commander adjustment vs own thug baseline

Forensic alt mode; judge by §18's cohort gap closing without §1–§4 degrading.

4. Done — snipes, T-key, accuracy, and efficiency left the composite

v2.11–v2.13. On the same-corpus reprocess, winner prediction stayed flat and Spearman ρ and self-consistency fell (critique/decisions/thug-efficiency-retirement.md). Reconstructed accuracy remains HOLD.

5. Phase 7(c) — ship-denied time adjustment for all tiers

Forensic alt mode; §17's per-outcome split as the diagnostic.

6. Phase 7(d) — small formats

Forensic alt mode admitting ≤2-per-side games to the wins ladder only; display-only small-format record on the player page; a rated small-format ladder, rule first, at ≥ 30 determined small games.

7. Display work, no rating consequence

Faction matchup table and pick-share trend on the Meta tab; faction and playstyle profiles; commander burden beside VTSR-C.

Do not do
  • Mix ≤2-per-side games into the lobby-relative composite (two-row z-scores are ±0.5 by construction)
  • Three per-player faction ladders (median cell 9 matches)
  • Inferred-role ladders (gameable; the Riot precedent)
  • By-ship ladders (dilution; commander-assigned)
  • A faction axis in VTSR-T (lobby-relative; penalizes thugs for a commander's choice)
  • MAX or softmax as the opponent reference or the thug aggregation — both refuted
  • Refit any frozen constant on the sample meant to judge it

14. History: what v2 proposed and v3 found

Condensed from the v3 review [78]. Two reforms refuted, one shipped.

The June 2026 independent review, also in [78], left three proposals this page does not otherwise track: soften the campod and presence cliff gates into proportional row weights (not shipped); show the bootstrap per-player noise band on the leaderboard (shipped: it is the soft edge of the VTSR-T position bar); sweep ALPHA_PVE and the effective-kill window and finisher floor (not shipped). Recorded outcomes, a validator on every run, rank scoring and the low-tier lift are already in the shipped list or the roadmap above.

Locked commander priors — refuted

Re-rating with the two hand-tuned priors unlocked moved the most-affected player 3.9 rating points against a 27-point bootstrap noise floor, well inside one band. Kept canonical.

MAX vs median opponent reference — refuted

Hard MAX inflated the pool’s mean rating by +522 points, halved Spearman ρ (0.46 → 0.19) and doubled bootstrap noise. Median kept. Inactivity K-boost (+min(20, 0.05·days)) shipped the same phase.

15. Sources cited

original ↗ = the link as found on 2026-10-04  ·  local copy = PDF kept in critique/publications/.

A. Rating-system foundations

  1. [1] Elo, A. E. (1978). The Rating of Chessplayers, Past and Present. Arco (2nd ed. 1986). original ↗
  2. [2] Bradley, R. A., & Terry, M. E. (1952). Rank analysis of incomplete block designs: I. The method of paired comparisons. Biometrika 39(3/4), 324–345. doi:10.2307/2334029 original ↗
  3. [3] Davidson, R. R., & Beaver, R. J. (1977). On extending the Bradley–Terry model to incorporate within-pair order effects. Biometrics 33(4), 693–702. doi:10.2307/2529467 original ↗
  4. [4] Glickman, M. E. (1999). Parameter estimation in large dynamic paired comparison experiments. Applied Statistics 48(3), 377–394. doi:10.1111/1467-9876.00159 original ↗ local copy
  5. [5] Glickman, M. E. (2001). Dynamic paired comparison models with stochastic variances. J. Applied Statistics 28(6), 673–689. glicko.net original ↗ local copy
  6. [6] Glickman, M. E., & Jones, A. C. (1999). Rating the chess rating system. Chance 12(2), 21–28. glicko.net original ↗ local copy
  7. [7] Glickman, M. E. US Chess Ratings Committee report 2019 (report19.txt); A Conversation with Mark Glickman, Chess Life (cl-article.pdf). original ↗ · report19.txt local copy · report19.txt original ↗ · cl-article.pdf local copy · cl-article.pdf
  8. [8] Herbrich, R., Minka, T., & Graepel, T. (2006). TrueSkill™: A Bayesian skill rating system. NIPS 19, 569–576. paper original ↗ local copy
  9. [9] Minka, T., Cleven, R., & Zaykov, Y. (2018). TrueSkill 2: An improved Bayesian skill rating system. MSR-TR-2018-8. microsoft.com original ↗ local copy
  10. [10] Weng, R. C., & Lin, C.-J. (2011). A Bayesian approximation method for online ranking. JMLR 12, 267–300. jmlr.org original ↗ local copy
  11. [11] Joshy, V. (2024). OpenSkill: A faster asymmetric multi-team, multiplayer rating system. JOSS 9(93), 5901. doi:10.21105/joss.05901 original ↗ local copy
  12. [12] Coulom, R. (2008). Whole-History Rating. Computers and Games (CG 2008), LNCS 5131, 113–124. pdf original ↗ local copy
  13. [13] Ebtekar, A., & Liu, P. (2021). Elo-MMR: A rating system for massive multiplayer competitions. WWW '21. doi:10.1145/3442381.3450091; arXiv:2101.00400 original ↗ local copy
  14. [14] Bodwin, G., & Zhang, F. (2023). Opponent indifference in rating systems: A theoretical case for Sonas. ITCS 2023, LIPIcs 251. doi:10.4230/LIPIcs.ITCS.2023.21 original ↗ local copy
  15. [15] Rosenbaum, D. T. (2004). Measuring how NBA players help their teams win. 82games.com original ↗ local copy
  16. [16] Sill, J. (2010). Improved NBA adjusted +/− using regularization and out-of-sample testing. MIT Sloan Sports Analytics Conference. original ↗ local copy
  17. [17] Ghimire, S., Ehrlich, J. A., & Sanders, S. D. (2020). Does regularized adjusted plus minus isolate individual NBA player contributions? PLOS ONE 15(8): e0237920. doi:10.1371/journal.pone.0237920 original ↗ local copy
  18. [18] Hollinger, J. (2002). Pro Basketball Prospectus: 2002 Edition. Potomac Books; PER (Wikipedia). original ↗ local copy
  19. [19] FanGraphs Sabermetrics Library. Positional Adjustment; WAR for Position Players; Beginner's Guide. original ↗ · Positional Adjustment local copy · Positional Adjustment original ↗ · WAR for Position Players local copy · WAR for Position Players original ↗ · Beginner's Guide local copy · Beginner's Guide
  20. [20] Cameron, D. Position Adjustments. FanGraphs. original ↗ local copy
  21. [21] FanGraphs. The Issue of Positional Inequality. original ↗ local copy
  22. [22] Tango, T., Lichtman, M., & Dolphin, A. (2007). The Book: Playing the Percentages in Baseball. Potomac Books. original ↗

B. Rating from in-match performance

  1. [23] HLTV.org (2017). Introducing Rating 2.0. original ↗ local copy
  2. [24] HLTV.org (2024). Introducing Rating 2.1. original ↗ local copy
  3. [25] dave. Reverse engineering the HLTV 2.0 rating. original ↗ local copy
  4. [26] Sardegna, C. Exploring problems with Counter-Strike rating systems. original ↗ local copy
  5. [27] De Bois, M., Parmentier, F., Puget, R., Tanti, M., & Peltier, J. (2025). PandaSkill – Player performance and skill rating in esports. arXiv:2501.10049 original ↗ local copy
  6. [28] Chen, Z., Sun, Y., Seif El-Nasr, M., & Nguyen, T.-H. D. (2016/2017). Player skill decomposition in MOBAs. Meaningful Play 2016; arXiv:1702.06253 original ↗ local copy
  7. [29] Dehpanah, A., Ghori, M. F., Gemmell, J., & Mobasher, B. (2021). Evaluating team skill aggregation in online competitive games. IEEE CoG 2021. arXiv:2106.11397 original ↗ local copy
  8. [30] Dehpanah, A., et al. (2021). The evaluation of rating systems in team-based battle royale games. arXiv:2105.14069 original ↗ local copy
  9. [31] Bober-Irizar, M., Dua, N., & McGuinness, M. (2024). Skill Issues: An analysis of CS:GO skill rating systems. arXiv:2410.02831 original ↗ local copy
  10. [32] Thompson, J. J., Blair, M. R., Chen, L., & Henrey, A. J. (2013). Video game telemetry as a critical tool in the study of complex skill learning. PLOS ONE 8(9): e75129. doi:10.1371/journal.pone.0075129 original ↗ local copy
  11. [33] Chen, Y., Aitchison, M., & Sweetser, P. (2020). Improving StarCraft II player league prediction with macro-level features. AI 2020, LNCS 12576, 256–268. doi:10.1007/978-3-030-64984-5_20 original ↗ local copy
  12. [34] Huang, J., Zimmermann, T., Nagappan, N., Harrison, C., & Phillips, B. (2013). Mastering the art of war: How patterns of gameplay influence skill in Halo. CHI 2013. pdf original ↗ local copy
  13. [35] Lewis, J. M., Trinh, P., & Kirsh, D. (2011). A corpus analysis of strategy video game play in StarCraft: Brood War. CogSci 2011. escholarship original ↗ local copy

C. Role- and faction-specific ratings in shipped games

  1. [36] Blizzard Entertainment (2016). Patch 3.7: Separate MMR Per Race. original ↗ local copy
  2. [37] Blizzard Entertainment (2016). StarCraft II: Legacy of the Void 3.7.0 Patch Notes. original ↗ local copy
  3. [38] Dota Team, Valve (2019-08-06). Matchmaking Update (Core and Support MMR). Text mirrored at r/DotA2. original ↗ local copy
  4. [39] Valve (2020-01-16). Dota 2 Update – January 16th, 2020. original ↗ local copy
  5. [40] Dota Team, Valve (2020-03-02). Ranked Roles Update (text mirrored at devtrackers.gg). original ↗ local copy
  6. [41] Altorfer, E. “SapMagic” (2018-08). /dev: Position Ranks in 2019. Riot Games. original ↗ local copy
  7. [42] Altorfer, E. “SapMagic” (2019-03). /dev: State of Ranked. Riot Games. original ↗ local copy
  8. [43] Riot Games Developer Relations (2019-04-17). Positional Ranking Deprecation. original ↗ local copy
  9. [44] Blizzard Entertainment (2019). Introducing Role Queue. original ↗ local copy
  10. [45] Blizzard Entertainment, Overwatch forums (2019). Role Queue Update. original ↗ local copy
  11. [46] Dexerto (2019). Jeff Kaplan confirms role queue release date, role-based SR and more. original ↗ local copy
  12. [47] Mercer, S., & Kaplan, J. (2020). The Whats, Whys, and Hows of the Overwatch Priority Pass; Keller, A. (2024). Director's Take: Opening up the conversation on 5v5 and 6v6. original ↗ · Priority Pass local copy · Priority Pass original ↗ · Director's Take 2024 local copy · Director's Take 2024
  13. [48] Aligulac. Frequently Asked Questions. original ↗ local copy
  14. [49] Aligulac. Database status. original ↗ local copy

D. FPS+RTS hybrids and their skill systems

  1. [50] FreeAllegiance Wiki. AllegSkill. original ↗ local copy
  2. [51] FreeAllegiance Wiki. Stack rating. original ↗ local copy
  3. [52] Wikipedia. Allegiance (video game). original ↗ local copy
  4. [53] Wikipedia. Natural Selection (video game). original ↗ local copy
  5. [54] Wikipedia. Savage: The Battle for Newerth. original ↗ local copy
  6. [55] Wikipedia. Nuclear Dawn. original ↗ local copy
  7. [56] Wikipedia. Battlezone (1998 video game). original ↗ local copy
  8. [57] Wikipedia. Battlezone II: Combat Commander. original ↗ local copy
  9. [58] Unknown Worlds Entertainment (2020-10-16). Introducing Hiveskill 3.0: Per-Team and Commander Hiveskill. original ↗ local copy
  10. [59] Unknown Worlds Entertainment (2020-10-28). Update 335 – Hive Skill 3.0 Released! original ↗ local copy
  11. [60] Unknown Worlds Entertainment (2016). Build 310 is now live on Steam! (Hive 2.0). original ↗ local copy
  12. [61] Unknown Worlds forums (2016). So what about hive 2.0? original ↗ local copy
  13. [62] Moultano (2014-08-04). A skill ranking system for Natural Selection 2. original ↗ local copy
  14. [63] Steam. Silica (Bohemia Interactive, Early Access 2023). original ↗ local copy
  15. [64] Steam. Eximius: Seize the Frontline (Ammobox Studios). original ↗ local copy

E. Matchmaking, engagement, small pools, incentives

  1. [65] Chen, Z., Xue, S., Kolen, J., Aghdaie, N., Zaman, K. A., Sun, Y., & Seif El-Nasr, M. (2017). EOMM: An engagement optimized matchmaking framework. WWW 2017. arXiv:1702.06820 original ↗ local copy
  2. [66] Chen, M., Elmachtoub, A. N., & Lei, X. (2024). Matchmaking strategies for maximizing player engagement in video games. Management Science. doi:10.1287/mnsc.2023.02957 original ↗ local copy
  3. [67] Kang, H., Suh, C., & Kim, H. K. (2024). Match experiences affect interest: Impacts of matchmaking and performance on churn in a competitive game. Heliyon 10(3): e24891. original ↗ local copy
  4. [68] Jaffe, A., Miller, A., Andersen, E., Liu, Y.-E., Karlin, A., & Popović, Z. (2012). Evaluating competitive game balance with restricted play. AIIDE 2012. original ↗ local copy
  5. [69] Diekmann, A. (1985). Volunteer's dilemma. Journal of Conflict Resolution 29(4), 605–610. doi:10.1177/0022002785029004003 original ↗ local copy
  6. [70] Strathern, M. (1997). ‘Improving ratings’: Audit in the British University system. European Review 5(3), 305–321. original ↗ local copy
  7. [71] Campbell, D. T. (1979). Assessing the impact of planned social change. Evaluation and Program Planning 2(1), 67–90. original ↗ local copy
  8. [72] Hardt, M., Megiddo, N., Papadimitriou, C., & Wootters, M. (2016). Strategic classification. ITCS 2016, 111–122. original ↗ local copy
  9. [73] Kleinberg, J., & Raghavan, M. (2020). How do classifiers induce agents to invest effort strategically? ACM TEAC 8(4). doi:10.1145/3417742 original ↗ local copy
  10. [74] Nosek, B. A., & Lakens, D. (2014). Registered Reports: A method to increase the credibility of published results. Social Psychology 45(3), 137–141. doi:10.1027/1864-9335/a000192 original ↗ local copy
  11. [75] Chambers, C. D. (2013). Registered Reports: A new publishing initiative at Cortex. Cortex 49(3), 609–610. original ↗ local copy

F. Internal provenance

  1. [76] data/processed/elo_current.json (schema 17, computed 2026-10-22), elo_history.json, elo_commander_current.json / elo_commander_history.json (schema 10, 730 duels), matches.json (287), validation_summary.json; data/external/f9_ledger.json + f9_community.json (F9bomber). original ↗ local copy · elo_current.json local copy · elo_history.json local copy · elo_commander_current.json local copy · elo_commander_history.json local copy · matches.json local copy · validation_summary.json local copy · f9_ledger.json local copy · f9_community.json
  2. [77] scripts/elo.py, scripts/elo_commander.py, scripts/validate_elo.py (§1–§21; §15–§19 regenerate every number in §5 of this page, §20 the numbers in §10, §21 the reconstructed-accuracy candidate); _validation/report.{md,json} run of 2026-10-22. original ↗ local copy · elo.py local copy · elo_commander.py local copy · validate_elo.py
  3. [78] Prior reviews: v1–v3 of this analysis and their plain-language editions (markdown, retired 2026-10-05 — this page is the only edition; the files remain in the repository history); the June 2026 independent review (repository history); the external .docx critique. original ↗ original ↗ · fable local copy · Analysis of the VTSR-T Algorithmic Matchmaking and Rating System in Competitive Environments.docx
  4. [79] Decision memos in critique/decisions/: phase-2b (×2), 2c, 3, 4, 5, balonce-meter-t-term, f9-external-duels, vtsr-c-v2-composite, vtsr-c/t ladder eligibility, vtsr-inactivity-threshold, phase-6-faction-advantage-term, thug-efficiency-retirement, accuracy-recon-axis. original ↗ local copy · phase-6-faction-advantage-term.md local copy · phase-5-wins-blend.md local copy · balonce-meter-t-term.md local copy · f9-external-duels.md local copy · vtsr-c-v2-composite.md local copy · thug-efficiency-retirement.md local copy · accuracy-recon-axis.md
  5. [80] Operator answers recorded in the v4 planning thread (2026-10-04): faction chosen by the commander per team; volunteer commanders, hand-split teams; complaint mechanisms = lobby-relative scoring under a stomp + ship denial; roles are a blend of self-selection, commander direction and ship/weapon constraints.
  6. [81] ideas.txt §4 Faction Analytics (June 2026): faction balance is a meta property, not a player-skill property; the deferred faction-aware axis gate reassessed here. original ↗ local copy · ideas.txt
  7. [82] DEVELOPER_GUIDE.md §13; docs/DATA_DICTIONARY.md §11. original ↗ local copy · DEVELOPER_GUIDE.md local copy · DATA_DICTIONARY.md

G. Added in the format-gate addendum

  1. [83] Blizzard Entertainment. Leagues and Ladders FAQ (StarCraft II: placement per mode, random-team rank per mode, arranged-team ratings). original ↗ local copy
  2. [84] Liquipedia. Battle.net Leagues. original ↗ local copy
  3. [85] Chess.com (2012). First rapid and blitz rating lists to appear July 1st, 2012; FIDE (2012-06-30). July 2012 FIDE Rating List; Kosteniuk, A. (2012-07-01). FIDE inaugurates new Rapid and Blitz chess Rating Lists. original ↗ · Chess.com 2012 local copy · Chess.com 2012 original ↗ · FIDE July 2012 list local copy · FIDE July 2012 list original ↗ · Kosteniuk 2012 local copy · Kosteniuk 2012
  4. [86] CSDB.gg. How CS2 ranking works — Premier rating & skill group system (Wingman 2v2 on its own skill group). Third-party guide; Valve publishes no Wingman ranking documentation. original ↗ local copy