The problem that had no name until 1961
Rolf Landauer was not thinking about sport. He was at IBM in 1961, trying to answer a narrower question: why do computers get hot? John von Neumann had guessed that every logical operation carried an unavoidable energy cost. Landauer showed this was wrong in a precise and useful way. Computation itself — copying a bit, comparing two bits, moving a bit from one register to another — can in principle be done for free, with no lower bound on the energy spent. The cost appears only when information is destroyed. Erasing one bit, irreversibly, must dissipate at least kT ln 2 of energy as heat: about three zeptojoules at room temperature. Charles Bennett sharpened this in 1973, showing any computation could run reversibly if you were willing to keep the intermediate steps instead of discarding them, and in 1982 used the same logic to close off Maxwell's demon — the demon's ledger has to be wiped eventually, and wiping it repays every advantage the demon appeared to gain for free. In 2012 Bérut and colleagues measured the bound directly, in a single colloidal particle held in a double-well trap, watching a bit get erased at almost exactly the predicted cost.
The finding is narrow. It says nothing about the cost of reading, storing or reasoning over information — only about destroying it. That narrowness is precisely what makes it useful outside physics. Any system that takes in more than it can keep must, sooner or later, throw something away, and Landauer says that act is never neutral. It has a price, and the price is not an engineering inconvenience to be optimised out. It is structural.
The analyst's actual working day
A performance analyst at a professional club is, functionally, a continuous-intake system wearing a lanyard. The inputs never stop: tracking data at 25 frames a second from every match and most training sessions, injury and load reports updated daily, transfer and loan activity across two or three leagues worth watching, and opponent tendencies that shift week to week as a rival's own analyst adjusts their own game plan in response to what your team just did. None of this arrives as a corpus to be studied once. It arrives as weather.
The characteristic failure is well known inside the profession, and it is not a failure of collection. It is a failure of forgetting. A game plan gets built on the observation that the opposition's right-back overlaps into the half-space on 60% of their attacking sequences — true, and expensively established, across the last six matches of data the analyst had time to code. The plan ships. On matchday, the right-back holds his position throughout. It turns out the opponent stopped that pattern three weeks ago, after their own analyst flagged it as exploitable, and nobody on your side re-ran the tendency report before the meeting, because the six-match sample felt current enough to trust and re-running it would have meant discarding work already presented to the coaching staff.
This is not a data problem. The data existed. It is a retention-and-erasure problem: an old belief was never actively destroyed, so it persisted past its truth, and destroying it in time would itself have cost analyst-hours nobody had budgeted.
Why the older strategies don't fit
It helps to see why the analyst's job cannot be solved by either of the two easier intake postures.
One posture is to freeze a corpus and stop. This is what a Large Language Model does: it compresses a fixed training set into weights once, pays whatever forgetting cost that compression requires exactly once, and then holds still — no further learning, no further erasure, a knowledge cutoff rather than a decay curve. A club could do something structurally similar: commission a season-end report on an opponent, freeze it, use it all season. Nobody does this past the first matchday, because the opponent is not frozen. Their tendencies decay on a schedule measured in weeks.
The other posture is to forget almost everything by default and carry nothing forward. This is the Large World Model's move: a scene-bound system that discards nearly all of what it senses the moment the scene ends, cheap precisely because nothing persists to be revised. A single matchday debrief, treated in isolation and then dropped, behaves this way — accurate about Saturday, silent about the trend Saturday sits inside.
The analyst's actual job is neither. It is continuous intake with revisable belief: tendencies held as claims with a shelf life, not facts, each one tagged, at least implicitly, with when it was established, from how many sequences, and what would overturn it. That is the third posture on the intake axis — the one that has to hold provenance for what it keeps and pay to overwrite what it revises, because the streams never stop long enough for either of the other two strategies to apply. Sports analytics is not a special case of some general data problem. It is a compressed, fast-clocked instance of exactly the situation Landauer's principle describes: unbounded arrival, bounded attention, and a real cost every time something is thrown out.
What it would cost to fix properly
Suppose the club tried to solve this by never discarding a tendency, only appending new evidence for or against it — an opponent-tendency log that grows every week and never deletes an entry, on the theory that revision by addition avoids the whole erasure question.
Why erase anything? Keep every tendency ever observed, timestamped, and let the model weigh recent evidence more heavily. Nothing gets destroyed, so there's no cost to worry about.
This is close to how well-run recruitment and scouting databases already work, and it is a real improvement over ad hoc report-writing — provenance survives, and an analyst can always ask when a tendency was last supported. But it does not eliminate the trade; it defers it. The log grows without bound. The analyst's working memory, the coaching staff's attention in a fifteen-minute pre-match briefing, and the club's tolerance for reading a tendency report longer than the scouting budget allows, do not grow at all. Something still has to be cut for the briefing to be usable, and that cut is erasure by another name, made under time pressure rather than as a designed policy. What the append-only log actually buys is control over when the cut happens and what evidence is available to make it well. That is a genuine gain. It is not an escape.
A second objection, heard from the more technically minded end of an analytics department, is that the whole framing overstates the physics. The actual energy cost of storing or discarding a row in a tendency database is nowhere near Landauer's three zeptojoules — server power draw, cooling, and network overhead swamp it by many orders of magnitude, the same gap that separates a DRAM write from the theoretical bound. This is true, and it means nobody should justify a data retention policy by citing kT ln 2. But the argument does not rest on the number. It rests on the asymmetry the number illustrates: copying a tendency into a report is cheap and can be made arbitrarily cheaper; deleting it responsibly — checking what conclusions were drawn from it, whether a game plan still depends on it, whether removing it breaks an audit trail the head coach might want after a bad result — is not. That asymmetry shows up in the department's actual costs, in analyst-hours rather than joules, and Landauer is the reason it is not an accident of how the software happens to be written.
Where the third rung actually sits
None of this argues that a club needs an unbounded model of world football running at all times, nor that such a thing is being built or sold. The claim is narrower. Given that intake in this domain is already continuous and effectively unbounded — matches keep being played, transfer windows keep opening, injuries keep happening — the only real choice left is how the erasure is managed, not whether it happens. A department that treats stale-tendency risk as a first-class problem, with explicit decay rules on when a tendency is re-tested and explicit provenance on where each one came from, is not doing something categorically beyond what any analytics operation already attempts. It is doing the same job with the trade made visible instead of implicit.
That is what makes the third position on the intake axis terminal rather than merely fashionable. There is no fourth strategy waiting past "hold every stream as a revisable, decaying, provenanced belief" — because once intake is continuous and total, the only work left is deciding, correctly and cheaply, what to stop believing. Progress from here is not a new kind of observation. It is better decay curves, better audit trails, and cheaper revision. The analyst who gets burned by an abandoned tendency is not missing data. They are missing a policy for forgetting it in time.