Chess's 97%: An Empty Hall, a Clean Number, and the Missing Method Box

December 2026, Mumbai. The first Indian leg of the Global Chess League, and fewer people in the hall than there were chairs. Above the board the camera was fixed on, the screen repeated one number: 97% accuracy. Below it, a ticker of moves; above it, commentary calling the game "brilliantly precise." I asked the broadcast desk one question: where is the method box for this 97%? What is the sample size, which engine version, what depth, and compared against what? The answer came back — the graphic had been built, the box had not.

The hall was empty enough that evening that you could hear every piece touch its square. The empty hall made every touch sound like evidence. The real story was sitting right there: the number was clean, and I went looking for the process that had made it clean — or merely made it look clean. Those are two very different things, and the gap between them is the biggest unsolved problem in this trade.

Context: where this number came from

Statistics are not new to chess. Their public, daily use is. When ChessBase India was founded in 2026, the centre of Indian chess coverage shifted from the print desk to video and livestream. That channel now runs close to 2.9 million YouTube subscribers — an odd number for a game whose live halls have never filled with spectators.

My own working life is built from two different traditions. On one side, a newsroom like ChessBase India, where the daily job is to log a live game as fast as possible. On the other, desks like The Hindu and The Hindu Sportstar, where the source comes first and the sentence comes later. Reporting on the ground from the Olympics to World Championships taught me that the number printed fastest is the number checked least. Between these two traditions sits the place where chess data reporting gets stuck: speed journalism demands, verification takes time, and a broadcast graphic has time for neither.

The dated fact matters here. Tata Steel Chess India launched in Kolkata in 2026. That same year I had built a 64-match xG model for the Russia World Cup, logging 169 goals and 1,842 shot events. In November the six-player Kolkata rapid field arrived, and I ported the same logic onto chess: converting Elo into expected score, then comparing it against actual points. I was in that hall when the sheet first ran — and that was the first time I understood that a football tool does not climb onto chess without breaking something.

Core: one number mutating between Moscow and Kolkata

In football, xG answers one clean question: how many goals should these shots have produced on average? Chess has no direct equivalent. Expected score derived from Elo tells you which way a result between two players should have leaned — it does not measure the quality of chances. Broadcast "accuracy" is a third, separate thing again: it tells you how far a player's moves stayed from the engine's first choice.

So I followed the xG from Moscow to Kolkata and watched the definition change. In Moscow, the expected-score language that emerged was probability — a branch of mathematics. On my Kolkata 2026 sheet, the same word was an estimate with a sample size and error bars stapled to it. In a Mumbai 2026 broadcast graphic, the same word had become decoration with no box at all. Expected score in Kolkata was not the same sentence as xG in Moscow — and this 97% is not the child of either.

That Kolkata sheet flagged two overperformers and one player in collapse across a six-player field — before the final rounds, when nobody else at the desk was even looking that way. The desk asked me to write that he "deserved" the win. I refused, because "deserved" is not a unit of measurement. Since then the method box is mandatory in every data piece I file: sample size, model assumptions, and error bars arrive before the claim, not after it.

In 2026 that discipline hardened. With sports halls shut worldwide, the FIDE Online Olympiad ran out of server rooms. India and Russia were declared joint winners after server failures. I logged 1,248 games, tagged 37 disconnects and 12 lag spikes, and built a three-tab audit: clock, lag, result. I refused to file until every decisive game had a timestamp and a platform note. The crowd was gone, but the structural tell remained — and the tell was that writing a server error as a player error is the easiest lie available.

August 2026, the Baku World Cup, is where the method paid. In the quarterfinal, Rameshbabu Praggnanandhaa beat Magnus Carlsen in a classical game, and almost every report after it opened with the word "miracle." I logged the knockout classical games, recorded 1,842 positions, and marked the 43 critical positions where engine evaluation swung by more than 1.5 pawns. The Baku ledger turned a chaotic final into a handful of positions. Then I separated time-pressure errors from opening-prep errors — that split was the actual finding. Before the semifinal I sent the desk a one-page note: the story was not a miracle, it was prep, clock control, and conversion under pressure. The critical-position ledger did not judge; it located the pressure.

I carried the same frame into the 2026 world title match. In December 2026, D. Gukesh became the youngest world champion — that is a date, not a data point. Next to age records, workload numbers speak far louder: games per year, seconds burned on the clock, travel miles, school years deferred. Playing a run of games without collapsing is a process. The first number is free; the second is built with labour, and broadcasts always pick the first.

Contrarian: accuracy does not measure accuracy

Back to that 97%.

Engine accuracy measures how closely a player's moves matched the engine's preferences. It does not measure whether the move was good. Three things always hide in the gap between those two sentences.

First, the time decision. A player who finishes on 97% at move 40 may have chosen an engine-approved move at move 38 that burned his remaining time. The accuracy graphic goes up, the clock goes down. The next twenty moves are decided by that clock, and that portion is recorded nowhere in the graphic. This is where technical precision and the result part company.

Second, board assignment. In a team event, who gets sent out on board one is a coaching decision, not an engine decision. In a league format that decision is often worth more than the match itself. In Mumbai I put two teams' numbers side by side: one team's overall accuracy was higher than the other's, and it lost the match. The number said one thing, the result said another. Correlation is not causation — the least written sentence in chess data reporting, because writing it costs you a line on the graphic.

Third, and largest, money. India has solved chess popularity; it has not yet solved chess pricing. The Velammal pipeline in Chennai and the Tal Chess Club lineage are the structure that raised the generation after Viswanathan Anand. State cash awards, PSU payrolls, school tournament counts: these are receipts, and receipts can be read. The question standing in front of those receipts is what a live ticket is actually worth, and why the audience will not pay it. From the 2026 founding to the Mumbai leg of 2026, reach grew, the stage got built, and the hall stayed half full.

On federation politics my rule is simple: no paper, no report. Election fights and state-association suspension orders return every cycle and feel like a scoop every time. I write them only with a budget line, a suspension order, or a payment schedule in hand — something that says who was paid, who was suspended, and what the delta buys in grandmasters per decade. Everything else gets logged and left alone.

One thing keeps pulling me back. The accuracy figure is free, looks precise, and feels good to publish. A tool like the critical-position ledger is expensive — it eats hours, and its output is often boring. Sometimes it reports that the metric did not move at all. I still have to publish that boring finding, because if I only publish the startling ones, I am doing the same job the broadcast does, just in a better font.

Takeaway

Next time an accuracy number glows on a screen, ask two questions: what is the sample size, and which human decision fell outside this number? If there is no answer, it is a graphic, not evidence. Indian chess's next real data point will be the count of occupied chairs in the hall — not the count of pawns.