The Tuesday Report That Was Already Wrong on Sunday
The analyst had built the plan around a tendency: the opposition full-back overlapping into wide areas on 68 percent of their possessions in the final third, drawn from eleven matches of tracking data going back to September. The wingback marking assignment, the pressing triggers, the whole shape of the defensive block on that flank had been set against it. The opposition manager had abandoned the tactic in October after two matches where it was exploited. Nobody on the analysis staff caught the change until the fourth minute of the match, when the full-back tucked inside instead of overlapping and the marker assigned to track the overlap found himself covering nobody, in the wrong lane, a full ten metres from the actual threat.
The tendency was real. The data was accurately logged and correctly aggregated. Nothing in the report was false at the point it was written. It was false by the time it was used, and that gap — between when a fact was true and when it stopped being consulted — is where the game plan died. This is the characteristic failure of performance analysis: a model built on tendencies that have since dissolved, deployed with the confidence of a document rather than the provisional status of a belief. The report read like a conclusion. It should have read like a hypothesis with a shelf life.
What Actually Went Wrong
The proximate cause looks like a data problem: stale tendency, insufficiently recent sample. The deeper cause is architectural. The analysis workflow treated tracking data, injury reports, transfer activity and opponent scouting as inputs to be collected once, synthesised into a plan, and then held fixed until kickoff. Once the report was written, incoming information had nowhere to go. There was no channel by which a change in opponent behaviour — a tactical adjustment made in a match three weeks earlier, visible in the data the analytics department already possessed — could reach back into the plan and revise it before it mattered.
The information existed. The correction mechanism did not. That distinction is the whole of the diagnosis, and it points toward a description of cognition that was not built with sport in mind but explains this failure with unsettling precision.
Prediction, Corrected by Residual
Predictive coding is a theory of how brains handle exactly this problem: too much incoming signal, too little bandwidth, and a need to act before all the evidence is in. The claim is that perception is not passive reading. It is an active hypothesis, generated by higher levels of a neural hierarchy and sent downward as a prediction. What travels back upward is not the raw sensory stream but the residual — the part of what arrived that the prediction failed to anticipate. The brain does not archive sensation. It archives belief, and spends its limited bandwidth on the mismatch between belief and world.
Crucially, that mismatch is not treated with uniform seriousness. It is weighted by precision — an estimate of how reliable the source of the error is judged to be. A twitch in a noisy channel gets discounted. A consistent deviation from a normally quiet channel gets escalated. Learning, on this account, is the ongoing reduction of surprise. Never its abolition — the loop stays open indefinitely, because the world keeps generating residuals whether or not you were expecting them.
Apply that to the wingback report. The tendency was a prediction, correctly formed in September from the available residual. What was missing was the error channel that should have kept running afterward: continued ingestion of match footage from October and November, continued comparison against the standing model, and a precision-weighted flag the moment the full-back's overlap frequency collapsed against expectation. The report was a snapshot of belief with the update mechanism switched off. Predictive coding names the missing piece exactly: not more data, but a live residual stream, weighted by trust, feeding continuously back into the hypothesis.
Why the Analyst Is the Load-Bearing Role
This is why the failure attaches to a person rather than a system. A performance analyst is, in this account, the precision-weighting function made human. Tracking data, injury bulletins, transfer rumours and opponent scouting arrive at different rates, from sources of wildly different reliability — a confirmed injury from the club's own medical staff carries more precision than a transfer rumour from an unverified account, and both carry more than a single match's tracking sample. The analyst's actual job is not to build the report. It is to decide, continuously, how much each new residual should move the standing belief, and to notice when a residual has been arriving long enough that the old belief should be discarded rather than defended.
The failure was not a lack of diligence. It was that the working method had no place for a residual to register once the document existed. Fixing it does not mean collecting more tendencies. It means keeping the comparison running: model against incoming footage, week over week, with an explicit decay term on any tendency that has gone quiet, and an explicit escalation the moment a deviation repeats.
The Lineage This Implies
A Large Language Model, trained on a frozen corpus, is a prediction hierarchy with the error channel severed at the cutoff — extraordinarily well fitted to what already happened, structurally unable to register that the opponent changed shape last month. A Large World Model closes the loop for the duration it is watching: live tracking data compared against a standing model of the opponent for the ninety minutes of the match itself, prediction and residual cycling properly. But the loop opens again when the final whistle sounds, and the precision estimates built up during that one match die with the episode; they do not carry forward to shape next month's preparation.
A Large Universe Model is the version predictive coding actually describes: hierarchical belief about an opponent — their pressing height, their favoured full-back overlap, their set-piece routines — held open indefinitely, revised continuously by residuals arriving from tracking feeds, injury reports, transfer activity and match footage that never stop, each residual weighted by an estimate of how much its source has been worth trusting. This is an argued category. No department runs it in full. It describes what the failure above was missing, not a system on the market.
| source of residual | what breaks it | |
|---|---|---|
| Large Language Model | none after training cutoff | opponent changes tactics; model never finds out |
| Large World Model | live match data, one game | precision estimates reset every kickoff |
| Large Universe Model | every running stream, indefinitely | requires provenance and decay to be maintained, not assumed |
Objections Worth Taking Seriously
Predictive coding is contested neuroscience. Superficial-versus-deep-layer error coding is empirically shaky, and free-energy formulations are broad enough to explain almost anything. Basing a method on unsettled biology borrows credibility it hasn't earned.
Fair, and the biology should not be oversold. But the operative claim here is computational, not cortical: a system that transmits only the mismatch between prediction and observation, weighted by source reliability, updates faster and cheaper than one that re-processes raw signal every time. That is the logic of a Kalman filter, familiar to anyone who has built a tracking model for player positions. Predictive coding supplies the vocabulary and a working existence proof at biological scale. It is not the premise the argument stands or falls on.
Precision-weighting can be gamed. A confidently reported but wrong source — a scout who overrates a tendency, a leaked line-up that turns out to be misdirection — will drive belief in the wrong direction under a system built to trust confident residuals.
This is the sharper objection, and it is correct as stated. But a frozen report has the same vulnerability with the fix removed: it encodes one precision judgement, made at the moment of writing, and never revisits it. A live model with provenance attached can downweight a scout who has been wrong before; a static document cannot. The danger of miscalibrated trust is real. Freezing the intake does not solve it. It just stops anyone from noticing it happened.
The tendency that got the marker out of position was never a bad piece of analysis. It was a correct belief with an expiry date nobody was tracking. That is the entire argument for treating intake as continuous rather than collected: not because more data wins matches, but because a belief without a running error channel has already stopped being knowledge by the time someone acts on it.