Beyond Accuracy: Measuring What Matters

The first round of the 2026 Candidates Tournament delivered three decisive games and one quick draw. But raw results only tell part of the story. How well did each player actually play – and more importantly, how hard were the positions they faced?

To answer this, I analyzed every game with Stockfish 18 at depth 20 and scored each position with the Elocator ensemble model, which rates position complexity on a 1-100 scale. A score of 70 means the position is more complex than 70% of positions in a calibration dataset of 35,000+ OTB games.

Instead of using “accuracy” (which treats a 5-CPL slip in a dead-equal endgame the same as a 5-CPL slip in a razor-sharp middlegame), I’ll focus on centipawn loss relative to position complexity. Losing 20 centipawns in a position with complexity 90 is more forgivable than losing 20 centipawns in a position with complexity 15.

The Scoreboard

Game Result White CPL Black CPL Avg Complexity
Caruana vs Nakamura 1-0 38.9 44.0 55.5
Pragg vs Giri 1-0 4.6 18.3 46.0
Bluebaum vs Wei Yi 1/2 3.1 1.2 20.2
Sindarov vs Esipenko 1-0 9.2 121.9 59.0

But these numbers need context. A CPL of 38.9 in a game averaging complexity 55 tells a very different story than a CPL of 3.1 in a game averaging complexity 20. Let’s dig deeper.

How Complex Were the Games?

The Sindarov-Esipenko game hit a complexity score of 100 – literally the top percentile in our calibration data. Meanwhile Bluebaum-Wei Yi peaked at just 47, never leaving comfortable territory before ending in a repetition on move 26.

The phase breakdown is revealing. All four games started with low-complexity openings (under 30), but diverged sharply in the middlegame. Sindarov-Esipenko and Caruana-Nakamura both reached middlegame complexity above 63, while Bluebaum-Wei Yi barely reached 28 before the draw.

The Sindarov-Esipenko endgame was the most complex endgame phase of the round at 85 – unusual, since endgames are typically simpler. This suggests the position remained deeply unclear even with reduced material.

CPL vs. Complexity: Who Handled Pressure Best?

This is the key chart. Each dot is a single move, plotted by the position’s complexity (x-axis) against the centipawn loss of the move played (y-axis). Dots clustered near the bottom mean clean play; dots shooting upward mean errors.

A few things jump out:

Pragg was surgical. His dots hug the bottom of the chart across all complexity levels. Zero inaccuracies, zero mistakes, zero blunders. An average CPL of 4.6 against a middlegame complexity of 65 is an extraordinary performance. Giri’s 18.3 CPL isn’t bad, but in this matchup it was enough for a full point.

Nakamura’s errors came in complex positions. His CPL spikes appear almost exclusively in the high-complexity region (60+), which is at least somewhat expected. But several of those spikes are large – the kind of errors that hand over a game at this level.

Esipenko collapsed. A CPL of 121.9 is brutal at any level, but it happened against a complexity backdrop of 59+ average. The chart shows his errors scattered across medium and high complexity, suggesting it wasn’t just one bad moment but a sustained breakdown.

Binning moves into low/medium/high complexity bands makes the pattern concrete. In low-complexity positions (complexity 1-33), everyone played well. In high-complexity positions (67-100), the separation is dramatic. Pragg’s high-complexity CPL is nearly zero. Esipenko’s is enormous.

Game-by-Game Breakdown

Caruana vs Nakamura (1-0)

An 83-move grind in the English Opening. The complexity stayed low through the opening but ramped up sharply around move 20 and stayed elevated through an incredibly tense middlegame and endgame. Both players made mistakes – Caruana averaged 38.9 CPL and Nakamura 44.0 – but Caruana’s errors were marginally less costly. The evaluation slowly drifted white’s way from the late middlegame, and Caruana converted a difficult endgame.

This was a war of attrition in high-complexity territory, and Caruana’s edge was narrow but persistent.

Praggnanandhaa vs Giri (1-0)

Pragg chose a sharp Closed Sicilian with an early f4-f5 break, and the complexity surged in the middlegame (peaking at 94). Despite the razor-sharp positions, Pragg played with machine-like precision: 4.6 average CPL with zero errors of any kind. That’s an exceptional performance at any level, let alone in a Candidates game.

Giri’s 18.3 CPL is respectable, but two blunders in the critical middlegame phase were enough. The game transitioned into a technical endgame where Pragg’s passed h-pawn decided matters.

Bluebaum vs Wei Yi (1/2-1/2)

The shortest game of the round at 26 moves. A Semi-Tarrasch where the complexity never rose above 47. Both players were essentially perfect (3.1 and 1.2 CPL respectively), but there was nothing to get wrong. Wei Yi forced a perpetual check, and neither player was ever in danger. A professional rest day.

Sindarov vs Esipenko (1-0)

The most complex game of the round, and it wasn’t close. From a quiet QGD Three Knights, the middlegame exploded around move 20 with complexity rocketing past 70 and eventually hitting 100. Sindarov handled the chaos brilliantly: 9.2 CPL with just one inaccuracy. Esipenko, facing the same complex positions, averaged 121.9 CPL with a blunder, two mistakes, and four inaccuracies. The evaluation swung decisively white after move 29 and never came back.

This was the most impressive result of the round. Sindarov – the lowest-rated player in the field at 2745 – navigated the most complex game with near-perfect play, while his higher-rated opponent fell apart.

Round 1 Takeaways

Complexity separates pretenders from contenders. When positions stayed simple, everyone played well. The real separation happened in high-complexity middlegames. This is what the Elocator model was built to measure, and Round 1 delivered a clear illustration.

Pragg’s performance was the standout. A 4.6 CPL in a game averaging 46 complexity (with middlegame peaks near 94) is a statement. If he plays like this for the rest of the tournament, the rest of the field has a problem.

Sindarov is for real. Handling the most complex game in the round with 9.2 CPL? Against a GM rated 2698? That’s not a fluke.

Caruana vs Nakamura was a slugfest. Neither played their best chess, but the positions demanded imperfection. The game’s average complexity of 55.5 with a middlegame of 63.4 meant errors were almost inevitable. Caruana just made slightly fewer of them.

Stay tuned for Round 2 analysis.

Analysis: Stockfish 18 depth 20, Elocator ensemble complexity model. All data and code available on GitHub.