The Empty Board: A Lesson in Data Discipline for Professional Chess Analysis

Moscow, January, seven in the morning. Snow clung to the aluminium window frame in small patches and melted into streaks of water. I opened the output file my collaborators had sent overnight: a nine-field data frame, aligned as neatly as a score sheet for the pieces before a game. Article title. Article source. Article type. Domain label. Core viewpoints. Information points. Entities involved. Time sensitivity. Source quality.

I read each cell. Title: absent. Source: absent. Article type: unclassified. Core viewpoints: blank. Information points: an empty list, not a single entry. Entities involved: not extracted. Time sensitivity: not assessed. Source quality: undetermined. One signal survived: the domain label read “chess”.

I sat still for about two minutes. An outsider would have assumed I was angry about a technical fault. I was not. I had just touched the most dangerous thing in sports analysis: a properly formatted shell wrapped around nothing. Broken data files are usually loud. They throw errors, they crash, they push a red line onto the screen. This one was silent. It was formally valid. And that validity was precisely the problem.

I have spent nearly two decades writing about chess and football for the Russian market. I have handled thousands of broken data files, hundreds of truncated reports, dozens of articles where a player's name was extracted incorrectly. Never before had I received a file with the right structure, the right fields, the right format, and nothing inside. This morning was the first time.

CONTEXT: TWO LAYERS AND A VALVE NOBODY CHECKS

For this story to mean anything, I have to explain how the operation runs. Without that, this morning's incident is just a trivial technical fault worth a line in a server log.

A professional chess analysis pipeline runs in two layers. Layer one reads the source article — a tournament report, a player profile, a federation statement, a social media round-up — and extracts information points: player names, event names, round numbers, ECO opening codes, the move number of the turning point, average centipawn loss, engine match rate. Layer two takes that list of information points and runs eight analytical dimensions: game technique, player data, tournament system, competitive landscape, rules and governance, risk, public narrative, and industry transmission.

The critical link sits in one line of the operating guide: every analytical dimension must be anchored to the layer-one information points. No information points, no analysis. That is the safety valve for the entire system.

The valve had just blown wide open. And nobody heard a thing.

The “chess” label remained alone in the file, like a piece left on the board after both players have pushed back their chairs and left the room. It tells us the router assigned a category. It does not tell us that any chess content was read. Most likely the label itself was assigned by default, because no other field carries a trace of text having been processed. Its information value is close to zero.

I am used to working the opposite way. In 2026, when the pandemic wiped the global calendar clean, I had no matches to write about. Seven months without football, seven months of asking why without pause. I sat down and classified two hundred and fourteen goalless draws across five top European leagues between 2026 and 2026, sorted them into nine pressing patterns, and built a weekly-updated spreadsheet with twenty-seven data columns. From then on, I never wrote a conclusion without being able to point to the cell that fed it.

And yet this morning I sat in front of a file with no cells at all. And the file still looked like a complete file.

THE VALID SHELL

A data file can be structurally correct and empty inside. In sports analysis, the most dangerous thing is not wrong data — it is data that looks right.

That sounds academic. It is not academic in the slightest. It decides the entire value of a three-thousand-word analysis.

Picture a spreadsheet. Column A is player name, column B is Elo rating, column C is games won this month, column D is average centipawn loss. If every data cell is empty, the software still opens the file normally. It does not report an error. It displays a blank page with a header row. A reader skimming past will see structure. A reader paying attention will see there is nothing to calculate.

Our system behaves exactly the same way. The frame has all nine fields. The eight analytical dimensions have all eight headings. But when I walk into each dimension, everything collapses in the same fashion.

Game technique: no opening identified, no ECO code, no system, no move. Nothing to compare against an engine. No remaining thinking time to assess clock pressure or technical distortion under rapid and blitz formats.

Player data: no person named. No classical rating, no rapid rating, no blitz rating. No performance rating. No opponent to build a head-to-head record against. No age curve to place in a peer context. The question of whether a young player is approaching some round-number threshold on the live rating list cannot be answered at all.

Tournament system: no event. No tier. No format. No qualification path, no rival for a qualifying place, no championship-cycle stage.

Competitive landscape: the throne, the 2700-plus challenger tier, the rising-star tier, the reserve pipeline — all four levels empty.

Rules and governance: no rule system identified. No International Chess Federation, no continental federation, no national federation, no organiser, no online platform appearing anywhere for comparison. No controversy to model an escalation path for, and nothing to dismiss either.

Public narrative: no narrative label, no position in the heat cycle, no expectation gap to measure.

Industry transmission: no upstream trigger — no event, no player, no controversy, no reform — so no transmission path to draw, from youth training through online platforms to sponsorship and derivative markets.

Eight dimensions, seven entirely blank. The eighth — risk — is blank in every content category too, but it leaves exactly one real line: analytical-pipeline risk. High level. Status: already occurred. Impact: high. Mitigation in three parts: a minimum-yield threshold, a raw-text fallback, and a gate that blocks empty files.

That was the only live signal in the entire file. A signal about the file itself. Nothing about chess.

THE METRICS THAT SHOULD HAVE BEEN THERE

In a complete file, I would have concrete numbers to work with. Elo is a rating measuring a player's relative strength; the higher it is, the stronger the player. Live rating updates continuously from results in an ongoing event, unlike the official monthly list. Performance rating is the rating level corresponding to a player's results in a specific event, and it often diverges sharply from the official figure. Opening preparation is pre-game study of variations aimed at a particular opponent. A second is an assistant who helps an elite player prepare and calculate. A novelty is a new move that has never appeared in the databases. Average centipawn loss measures the average evaluation loss per move; lower is better. Engine match rate is the share of moves matching the engine's first choice. Over-the-board means in-person play, as opposed to online play. And in must-win formats, organisers use rapid or blitz tiebreaks to settle a drawn match.

A complete file on a Candidates Tournament game — the event that determines the world championship challenger — would carry the player's name, the ECO code, the move number of the turning point, and usually the second's name as well. For Ian Nepomniachtchi, who has won the Candidates twice, or for Magnus Carlsen, or for Ding Liren, every one of those figures can be looked up and cross-checked against public databases.

This morning my file had none of them. No name. No code. No number. No rate.

What is worth noting is that this absence was never flagged. No line in the file said “data unavailable”. The cells were simply empty, the way a spreadsheet is empty before anyone has entered the figures.

THE FABRICATION TRAP

This is the part that cost me sleep.

If I had handed that empty file to a collaborator this morning and asked for the eight dimensions, the result would almost certainly have been a fluent article. It would have had a player's name. It would have had an opening name. It would have had a perfectly plausible average centipawn loss figure. It would have carried a judgement on what percentage of the player's moves matched the engine's first choice. It would have closed on a line about the form curve and the risk of a rating slide.

All of it would have been wrong. And all of it would have read very convincingly.

I understand that instinct because I once lived inside it. In 2026, at sixty, I wrote a three-thousand-four-hundred-word analysis of RB Leipzig's gegenpressing system under Ralf Rangnick, using expected-goals data from thirty-four Bundesliga rounds. I argued the model would collapse against a deep-lying defence. The Russian online community tore into me and called me a reactionary. I did not argue. I spent six weeks rewatching every Leipzig match from that season, logging four hundred and twelve failed pressing situations, and published a correction with hard numbers in September.

The first article got pelted. Data never takes offence.

But the Leipzig story differs from this morning's in one lethal respect. In 2026 I was wrong because I over-interpreted real data. This morning the potential error sits somewhere else entirely: the real data never existed, and I could still have produced a perfect article.

A language model asked to analyse an empty file will tend to generate plausible-sounding chess content. That is its instinct, not malice. It is trained to complete text, not to decline tasks. In an industry where speed of publication is money, there will always be someone who wants to accept the fluent article rather than a line reading “analysis not possible”.

My tolerance for this kind of error is unusually low, perhaps because I learned chess before I learned writing. In chess, a position cannot be fabricated. The board lies very poorly. If you say the white king is on e2 while it stands on e1, any child can point it out in three seconds. Sports analysis has no such property. It lets a writer drop a number onto the page, and that number is only checked months later, by a stranger, at a moment when nobody remembers the original article.

THE FALSE-NEGATIVE SPIRAL

There is a second-order consequence more dangerous than the fabricated article.

Imagine this empty file is not blocked. It is processed, stored in the database, and tagged “analysed”. Six months later, a researcher queries that database: across chess coverage from this period, did any cheating controversy appear?

The database answers: no.

But the truth is that nobody read anything. The file never contained a single assertion.

In chess, this is not a theoretical worry. Cheating, outside assistance, federation transfers, the fairness of tiebreak procedures, the governance procedures of FIDE — these are real topics, with real controversies, real case files and real precedents. A database that says “no controversies” when nothing was ever read produces a wrong conclusion with statistical weight behind it. It will be cited. It will go into reports. It will become the basis for the next editorial decision.

The absence of a controversy in the data file is not evidence of its absence in the source article, if the data file never contained an assertion at all.

I call this false-negative contamination. It is quieter than a wrong article and far harder to fix. A wrong article can be corrected, retracted, challenged by readers. A starved database has nobody looking for it.

HOW TO PATCH IT: A YIELD THRESHOLD AND A LOUD-FAILURE RULE

People tend to assume the fix for a data problem is more data. It is not. The fix here is less silence.

We agreed on three mechanisms and began applying them this week.

The first is a minimum-yield threshold. Layer one must return at least three information points and at least one named entity — a person, an event, or a governing body. Below that threshold, the pipeline must fail. Fail loudly. A red error notice, a log line, an email to the person responsible. No valid-looking data frame is emitted in silence.

The second is a raw-text fallback. The source article's raw text must be stored alongside the structured output. When the extractor performs poorly, layer two can fall back to reading directly. I do not want to sit in front of a page of N/A markers again with nothing to cross-check against.

The third is capturing provenance metadata at fetch time: source URL, publisher, an absolute timestamp, and the access result. Those four fields must exist before any analysis begins. In this morning's file, all four were missing. No URL, no publisher, no timestamp, no note on whether the fetch succeeded or failed. Without provenance there is nothing to audit.

This is the part I have to own first. My spreadsheet has twenty-seven data columns on football matches, but not one column on the quality of the source being fed in. I tended the conclusion and ignored the doorway.

FOUR SIGNALS TO TRACK

Out of this incident, we built a dashboard of four metrics, updated weekly.

Re-fetch success rate. Count the articles returning three or more information points. If it falls below ninety per cent, the pipeline has a throughput problem and the analysis queue will back up within two to three weeks.

Empty-schema incidence. Count layer-one outputs with a complete frame but an empty information-point list. Any occurrence signals that the assertion layer has vanished from the system.

Entity-resolution yield. Named entities per article. Zero entities on an article that has already been assigned a domain label means the entity extractor or the upstream text is broken.

The ratio of structural validity to content validity. This is the metric I care about most. When the two sides drift apart, the downstream database is degrading and nobody is raising an alarm.

None of these four require sophisticated engineering. They require one accountable person to read them every Monday.

THE CONTRARIAN ANGLE

This is where I want to turn the board so others can see it.

The first reaction most people have to an empty file is to treat it as a failure. I would argue this empty file is more honest than most of the full files circulating in the industry.

People watch the players run. I watch the whole block of the formation shift. Chess analysis works the same way. People watch the chart, the average centipawn loss, the engine match rate. Nobody looks at layer one. Nobody checks whether forty information points were genuinely extracted, or whether two were extracted and the remaining thirty-eight were filled in by the writer while typing.

The biggest blind spot in sports analysis is that we audit conclusions and never audit intake. A chess analysis piece can be shared three thousand times across the Russian internet, complete with twenty-two redrawn position diagrams, and its foundation can be an extractor that recovered two out of forty data points. Nobody knows. The reader sees only confidence.

From the touchline, I see the whole match. And from the touchline of the pipeline, where nobody stands, I see the machine running empty.

There is one possibility I do not rule out: our evaluation systems partly reward formal completeness. A tidy nine-field file is easier to accept than an error notice. A fluent three-thousand-word analysis is easier to share than a line reading “insufficient data”. As long as the reward sits with form, there will be people filling the void with words.

WHAT I GOT WRONG

I believed the biggest problem in sports data analysis was the quality of conclusions. I spent years drawing my own charts, checking my own figures, cross-referencing pressing situation by pressing situation. I thought that was enough.

I was wrong about where to put the emphasis. A conclusion is only the top floor of a building. I tended the top floor and never went down to check the foundation. For years I never built a single gate on the intake side, because I assumed that was an engineering matter, not a writer's matter. This morning showed me that boundary is artificial. Whoever signs the analysis owns every layer of it.

THE SYSTEM DOES NOT LIE

I am sixty-nine. I still learn from twenty-five-year-old engineers, who rebuilt the gate this morning in four hours. The system does not lie, but it can only be heard when the data is thick enough. An empty file is not a lie. It is a silence, and the human ear always tends to fill a silence with what it wants to hear.

When football returned after seven months of shutdown in 2026, my first piece on how empty stadiums affected pressing rhythm prompted three Premier League clubs to email me for more. I mention that not to boast. I mention it to say that what gave that piece its value was not the prose. It was two hundred and fourteen matches classified in silence across seven months. The data sits in the foundation. It always does.

WHAT TO VERIFY NEXT ROUND

Next time you read a chess analysis with twenty-two redrawn diagrams — mine included — try asking one question: which door did the source text come through, and did anyone check that door?

I will answer that question for myself every week, using the four-metric dashboard. Any week the yield rate drops below threshold, I will write a line saying “analysis not possible” instead of a long piece. It is not a large commitment. But it is the only one I can make after this morning, and it is the one I can actually verify.