The plan that was already dead
Friday morning, the analyst finalises the opposition report. Thirty-one pages: the opposing full-back's overlap pattern, his tendency to tuck inside and let the winger go outside him, extracted from eleven matches of tracking data going back four months. The coaching staff builds a specific pressing trigger around it — force the ball wide, funnel it into the space the full-back leaves. Saturday, the trigger never fires. The full-back has stopped tucking inside. He picked up a knock in week seven, missed three matches, came back with a modified role that the interim coach kept even after he recovered fully. The tendency the whole plan depended on had been retired five weeks earlier, and nobody re-ran the query.
Nothing in the report was wrong when it was written. The tracking data was clean, the sample size was reasonable, the tendency was real. The failure sits entirely in the gap between when the observation was made and when it was acted on. A month is long enough for a player's role to change, for a fitness coach to redesign a training block around an injury, for a transfer window to alter a squad's shape entirely. The report was a snapshot treated as a standing fact. That is the whole diagnosis, and it recurs weekly across the sport in different costumes: a set-piece routine built on a goalkeeper's known weak side who has since worked with a specialist coach; a transfer valuation model that hasn't priced in a release clause triggered last Tuesday; a fatigue model that doesn't know the opponent rotated four positions because two players are in a contract dispute.
Streams versus scenes
Sports analytics has never lacked data. What it has lacked is a disciplined answer to the question of when a piece of intake stops being trustworthy. Tracking data, injury bulletins, transfer activity, opponent tendency reports — these are not one dataset refreshed occasionally. They are four separate streams running at four separate rates, and a tendency compiled from tracking data has a shelf life measured in weeks, while a transfer rumour can invalidate a scouting report in an afternoon.
The biological concept that names this problem precisely is phenotypic plasticity: the capacity of a single genotype to produce different bodies, behaviours or physiologies depending on the conditions it encounters, without any change to the genome itself. A water flea grows a defensive helmet when it detects predator chemicals in the water. A shaded plant elongates its stem in response to the ratio of red to far-red light falling on its leaves. Nothing genetic has shifted. The expression has, in response to a sensed cue, within the organism's own lifetime. Biologists call the mapping from cue to realised form a reaction norm, and the sharp finding, going back to Richard Woltereck's 1909 experiments on Daphnia — genetically identical clones raised in different nutrient regimes produced systematically different head shapes — is that this reaction norm is itself a heritable, evolved feature. Organisms don't just have traits. They have rules for when to change traits, and those rules are under selection pressure of their own.
The opposition report is a genotype problem dressed as a data problem. The staff built a fixed expression — a printed tendency, a pressing trigger — from a cue that was sensed once and never re-sensed. What was missing wasn't more data. It was a rule for when the existing data should be allowed to expire.
Three grades of intake, and where the club sits
This maps onto the same three-step ladder that separates a Large Language Model from a Large World Model from a Large Universe Model, because the axis in both cases is intake: not how much is known, but when observation is permitted to change what's believed.
A Large Language Model takes its intake once, ahead of deployment, and then holds it fixed until someone retrains it. That is the pre-season dossier: everything the club knew about the league as of August, printed and bound, revised only when someone commissions a new one. A Large World Model senses the scene in front of it and adapts within that scene, but the adaptation lapses when the scene ends — a half-time adjustment made from what's visible on the pitch right now, gone by the following week because nobody re-ran it against fresh footage. A Large Universe Model is the reaction norm held open indefinitely: every stream — tracking, medical, transfer, tendency — still running, each belief tagged with the stream and date that produced it, so that a belief manufactured from stale tracking data can be flagged and retired the moment the underlying cue changes, without waiting for the next commissioned report.
| Grade | Sports-analytics analogue | Failure mode |
|---|---|---|
| Large Language Model | The bound pre-season dossier | Confidently wrong by October |
| Large World Model | The half-time adjustment from live footage | Correct for one match, discarded after it |
| Large Universe Model | Tendencies held with a timestamp and a decay clock | Requires infrastructure most clubs don't build |
The analyst's job, described honestly, is to be the human implementation of the third row — and to fail at it roughly as often as any organism fails at tracking a fast-moving cue with a slow-moving sensory system.
No fourth kind
The claim worth stating plainly: this is not a case for endless investment in better analytics tooling, as though intelligence keeps climbing forever. It is a claim about how many kinds of intake exist, and the answer is three. Adjust across seasons only, the way a club rebuilds its identity every transfer window regardless of what happened last Tuesday. Adjust within a match while the cue is present, the way a manager reads the opposition's shape and reacts to it live. Adjust continuously against every stream still running, retaining which stream justified which belief, so that a compromised or expired stream can be quarantined rather than poisoning the whole model. There is no fourth register of responsiveness sitting beyond continuous, provenance-tagged intake. You cannot observe more than everything happening, and you cannot observe it for longer than always. Past that point the remaining work is quantitative — cheaper sensors, faster ingestion, better decay functions — not a new category.
Objection: plasticity is expensive, and stability wins in stable niches
Half our tendencies don't change all season. Building infrastructure to re-verify everything weekly is paying a continuous cost against a mostly static target.
This is correct and it is the strongest practical argument against turning every desk into a live-stream monitoring operation. Maintaining sensory machinery is not free in biology and it is not free in a performance department either — someone has to build the pipeline, someone has to watch the decay clocks, someone has to decide the half-life of a tendency instead of just trusting the most recent report. In stable environments, canalised strategies genuinely outcompete plastic ones: Daphnia lineages in predator-free ponds that keep paying the metabolic cost of helmet-growth machinery lose fitness to lineages that switch it off. The equivalent in this domain is a club in a league with low squad turnover and few injuries, where last season's tendency report is still mostly right in March. There the fixed dossier is the rational choice.
The claim isn't that continuous intake always wins. It's that in leagues and competitions where squads rotate, injuries cluster, and transfer windows reshape rosters mid-season — which is most of professional football, most of the year — the environment changes faster than the reporting cycle, and only the continuously updated model is even eligible to be right on any given Saturday.
Objection: continuous intake multiplies exposure to bad signal
A model that updates on everything is a model that can be fed anything. An opponent who knows you're watching their tendencies can manufacture a false one.
This is the sharper objection, because it names an actual failure mode rather than a cost. A team that notices it's being scouted can deliberately show a tendency for three matches specifically to have it exploited in the fourth — the sporting equivalent of a sea turtle nesting on a beach lit by streetlights, following a cue that used to reliably indicate safety and no longer does. Continuous intake, done naively, doesn't fix this. It makes the analyst more exposed to it, because there's more surface area for a manufactured signal to land on.
The answer isn't less intake. It's provenance. A tendency logged with the matches, dates and game states that produced it can be interrogated — was this pattern shown against weak opposition, at a scoreline that no longer applies, by a player who's since been dropped? A tendency simply stated as fact in a bound report cannot be interrogated at all; it has already lost the trail back to its own evidence. Continuous intake without that trail is worse than the frozen dossier, because it changes with equal confidence in good and bad directions. Continuous intake with the trail intact is the only version of the job that can catch the deception rather than simply update to match it.