Every defensible choice · game 01

White's advantage, if you define it

White moves first. White stays above 50% in most broad definitions here. The famous percentage does not stay fixed. Change the population, rating floor, clock, draw rule, estimator or opening, and watch the instrument move.

54.1767%Jeff Sonas's 2002 fitted score for equal ratings. It is an intercept, not a raw average, and its 266,000-game source database is proprietary and unavailable.

Loading two local count cubes and crossing 5,250 definitions…

First, the anchor

His method. New populations.

Sonas fitted White's score against White's rating advantage, clipping the predictor below −460 and above +390. The intercept at equal ratings was 54.1767%; the slope was 0.001164 per rating point. His rows cannot be recomputed because they were never released. These are explicitly method reproductions on open Lichess data, not reproductions from the same rows.

Sonas prints · unavailable rows 54.1767% 0.001164 slope · about 35 rating points
Open broadcast archive · same estimator loading loading from Sonas · loading
Online rated samples · same estimator loading loading from Sonas · loading

The formula and clipping bounds are printed together in Sonas's original article. The open broadcast records carry FIDE ratings. The online records carry Lichess Glicko-2 ratings, so a rating number is not a shared axis between the two pools.

Make one definition

What does “White scores” mean?

Every menu is also crossed with every other menu below. Your choice is one point, not the answer.

The two rating systems are not comparable.

loadingdecoding the shipped counts
The hidden mover

The live draw rate is printed beside every estimate. A half-point draw rule barely moves a low-draw online pool, but can dominate a master pool in which draws are common.

The complete curve is computing.

Plot count pending. The solid horizontal line is Sonas's printed 54.1767%; thin dashed lines are the other closed-data references listed below. The gold mark is the nearest open-data analogue: all rated broadcast games, conventional draw scoring, clipped fit, openings pooled.

Sonas's printed value sits at the loading of the defined open-data results. The nearest open-data analogue is loading, at the loading. Those are two different locations because the printed database is absent.

A choice Sonas could not make

The clock changes the score, too.

Online PGNs say whether a game ended by normal play or time forfeit. Broadcast PGNs do not carry that header, so the clock comparison is available only for the online pool and is disabled as a cross-pool claim.

All endingsloading
Normal onlyloading
Time forfeitsloading

loading survive in the sampled panel. Their White score differs from normal endings by loading. This is descriptive of these Lichess samples, not a causal claim about clock pressure.

The scout's January 2013 result is reproduced: normal endings score loading, time forfeits score loading, a difference of loading. The direction reverses when all seven shipped samples are pooled.

Second layer · the trend fight

The direction is another choice.

Sonas made no claim that the first-move advantage was declining. Other authors did disagree: Streeter saw White's score rise across historical periods; Watson described a slip from 56% to 55%; a 2013 PLOS ONE paper found a rising advantage measured in centipawns, which is a different outcome again.

Computing the selected trend.

Each annual point must contain at least 250 effective observations, and a line needs three points. Online points inherit the time-ordered January-slice limitation.

All 5,250 trend questions

Computing trend signs.

This is a sign map, not evidence that one historical author “won.” The populations cover different years, ratings and forms of play. A slope here describes the shipped panel only.

Streeter, 1946: White's score rose across three historical periods.
Watson, 1998: the long-standing figure had slipped from 56% to 55%.
Ribeiro et al., 2013: White's centipawn advantage rose toward a plateau. Different unit, different claim.

Printed, not recomputed

Reference marks from closed datasets

These values provide historical bearings. Their underlying databases are not redistributed here, so they are visually and verbally separate from the live curve.

52.16%2009 World Blitz Championship, 462 games
53.4%Streeter overall, 5,598 games
54.1767%Sonas fitted intercept, 266,000 games
54.8%New In Chess Yearbook 55, pooled
55.7%Adorjan, players rated 2700+
56.1%New In Chess, after 1.d4

The check

The browser counts what it loaded.

separate filePGNs seenrated usedcellsSHA-256
decoding

Grid accounting pending.

Which choice moved the curve?

Variance decomposition pending.

Failures remain in the denominator of attempted specifications. They are not silently removed from the design, though summary values necessarily describe the defined cells.

Estimator and refusal rules

The Sonas method bins White rating minus Black rating in 25-point bins, drops fit bins with fewer than 30 effective observations, clips the predictor to [−460, +390], and runs a games-weighted least-squares line. The equal-rating estimate is its intercept. Other estimators use the raw pool or rounded difference bands of 25, 50 or 100 points.

A main-grid result needs at least 1,000 effective observations after its filters. An annual trend point needs 250. Fewer than three annual points means no trend. The reachable narrow corner is therefore labelled “undefined,” not converted into a percentage.