Previewing the Unplayed Match — SRP-02
How a model of the match preview was built, tested against real writing, and corrected by a football result
A working paper on sports-writing analysis and the building of an AI analysis tool
#Summary
This paper records the development of a model of the match preview - the piece of sports writing that comes before an event rather than after it - and of an AI tool, SR3, built to analyse previews against that model. The work began as an offshoot of an existing tool for match reports and ended somewhere its starting assumptions did not predict. The model was reasoned out, rewritten once for plain language, then tested against six real previews and revised at each step. Two moments did more than the rest: a rewrite that exposed how much of the early work was jargon standing in for thought, and a football match whose result quietly refuted the analyst's overconfidence while leaving the tool's actual task untouched. The paper argues three things. First, the preview is not a smaller match report; it is the report's simpler ancestor, and the differences are structural, not cosmetic. Second, an analysis model is only worth what it survives, and the cheapest, most honest test is contact with real examples and, where possible, with the world. Third, the discipline that protected the work - separating what was proven from what merely read well - is the same discipline the tool itself is built to enforce, and the hardest place to keep it is in one's own prose.
None of the conclusions here are settled. They are recorded at the point the work paused, with the untested parts marked as untested. That is the honest state for the work to rest in, and it is the state the project's whole identity demands.
#1. Where it started, and the report it grew from
The match report has a tool. It reads a finished report and traces the reader's journey through it on three tracks: what the reader learns, what they are given to care about, and how hard each phrase works for its space. It does not rewrite; it diagnoses and hands judgement back as questions. It was the proven thing, tested on a handful of real reports, and it carried a working vocabulary - the live lead, the buried angle, the load-bearing cliche, attachment as a dial the reader slides up and down rather than a switch.
The preview is the report's neighbour and was the obvious next form. But the first question - is the preview just a report pointed forwards? - turned out to be the wrong question, and answering it properly took the rest of the work. The preview is writing about something that has not happened. That single fact, followed honestly, pulls almost everything else out of shape.
#2. The model, and what the report taught about the preview
The most useful early move was to take the standard practitioner account of why a report exists - to relive the match, to test your opinion against an expert's, to learn a fact, to grasp what the result means - and run each reason forward to the time before the match. The list did not copy across. It contracted, and the contraction was the finding.
Reliving has no forward counterpart - there is nothing yet to relive - but the need underneath it is not memory; it is the pleasure of inhabiting a match you are not at. The report serves that by replay; the preview serves it by rehearsal. Run forward, two of the report's reasons - living the match second-hand and reliving it - collapse into one, because the split between those who were there and those who were not depends on the match having happened, and before kick-off nobody has been. So the preview has fewer functions than the report, not because it is poorer, but because the event has not yet arrived to divide its readers. The preview is the report's simpler ancestor: the same thing, before the final whistle forks it.
From there the model settled, through discussion, into three jobs a preview does. It connects the reader to the match itself - making them care, building anticipation, setting up a better watch. It places the match in the wider context of the sport - the table, the season, the rivals, the judgement of teams and managers, the gambling read. And it gives the reader something to say - the social currency that lets a person talk to a parent, a stranger, a colleague, in the low-stakes way sport allows. A preview can serve one of these well and barely touch the others, and good ones usually serve more than one at once.
Underneath the three jobs sat a single recurring fault, which became the tool's central diagnostic: a strong fact presented on its own, never connected to a reason to care, judge or talk. Most often two facts a paragraph apart that, struck together, would have been the reason to watch. Naming this fault - and distinguishing it from the separate question of which angle the writer should have chosen - is most of what the tool does.
#3. The rewrite, and the lesson about jargon
The first full statement of the model was, on inspection, badly written - and badly written in a way that mattered, because the project's whole subject is writing that does not lean on shorthand the reader has to already know. The draft was thick with invented terms: things were “bound” or “unbound”, there were “offers” and “the existential offer” and “charge” and “liveness”. A reader who had not sat through the discussion would have lost the thread in the second paragraph.
The rewrite was instructive beyond tidying. The test that exposed the worst sentences was simple: try to edit the sentence. A sentence you cannot edit is usually one that does not say anything precise enough to edit. “It leaves one strong angle sitting unconnected on the page” resisted every attempt to improve it, because “unconnected” was a ghost of the jargon term “unbound” and never said what the thing was unconnected to. “Competent” resisted because it is empty until it is cashed out into which job, done how. The cure was to force every such word to name the specific thing it stood for, or be cut.
A distinction emerged here that shaped everything after. The vocabulary split into two kinds. There were industry terms - peg, angle, lead, colour - which a writer would meet in any newsroom and which the tool should use, because they are part of what the writer is there to learn. And there were the project's own coinages, which existed nowhere but the design notes, and which had to be said in plain words whenever they faced the reader. The test for a term was not whether it was technical but where the writer would meet it. The preview, being under-theorised, had forced the invention of more private vocabulary than the report ever needed - which is exactly why the preview tool was the more jargon-heavy of the two, and why the discipline mattered more there.
#4. What the examples did to the model
The model was reasoned out inside a single long discussion, by the same process that built the tool. That is the condition the project most distrusts: a model and a tool that agree because they were made together, marking their own homework with the answer key open. The only cure was contact with real previews. Six were used, and each one changed something.
#The form is several different objects
The founding assumption was that a preview is a kind of feature - a crafted, written-through article. The real examples said this is true of some previews and false of most. What is published under the word divides into at least four objects that must be judged differently: a single writer's article, with a voice and a real ending; a section-based compilation, effectively a fixture card grown to full length, which lists its material rather than connecting it and ends by running out of sections; a live data-run, written to a clock; and a betting card, the preview reduced to a price. The tool had to learn to name which kind it was looking at before judging, because holding a section-stack to a feature's standard - faulting it for not landing a crafted ending - is the same error as faulting a dictionary for having no plot.
#The ending claim, wrong twice
An early claim held that the preview, having no result, has no ending and should close on a question. A later correction replaced it with the opposite: that the preview is a feature and owes a crafted, widening close. The first three examples supported neither - they simply stopped, on a stat or a TV listing. Then a local-paper preview landed a real ending, widening out from the match to the long history of the fixture. The pattern was not print against web; it was authored article against assembled object. A piece a single writer composes through, start to finish, closes; a piece assembled from blocks ends when the blocks run out. The rule was not dead. It had been mis-scoped, and only the examples could have shown where.
#The mix shifts by level and by publication
A single standard would have misjudged half the pieces. The level of football sets what kind of material leads - tactics high up, people and belonging lower down. The publication sets the slant: a local paper sees the match through its town and its home support where a national stays neutral or sees it through the title race. These are two separate axes, and the tool had to hold both, or it would have marked a lower-league local preview down for lacking a tactical reading it was never trying to provide.
#5. The professional ceiling, and why student work was needed
Four professional previews, run through the tool, all yielded the same fault - the buried, unconnected fact - and this raised a real worry: did the tool know only one fault, and reach for it from habit? The answer turned out to depend on something about professionals. A working journalist does not over-commit to a wrong angle, does not make an empty claim of significance, does not break the structure. They have trained those faults out. The buried connection is the residue of assembling under deadline, and on professional pieces it is close to the only fault available to find. So the tool's range was untestable on professional work, not because the wrong pieces had been chosen, but because professional pieces cannot contain the faults the tool's range would have to catch.
Student pieces could. Given a student preview, the tool found faults the professional pieces never held: a worn phrase doing the work at the charged moment - “it will go down to who wants it more”, exactly where the writer's own reading was owed - and a headline that pointed one way while the body drifted another. Crucially, the tool named these as different faults rather than forcing everything into the buried-connection finding. Across two student pieces, run by two different AI models on the same tool, the same fault was caught the same way, which is the strongest evidence that the finding lives in the tool rather than in the model running it. The range held. The thing that could only be established on student work was established there.
#6. The match that judged the analyst
One student preview covered a fixture between the league leaders and the bottom side, whom a supercomputer gave a 99.1 per cent chance of relegation. The tool found its usual fault and proposed a buried collision - a returning defender reassembling a patched-up back line against a team that could not score. The human running the discussion then argued, from the other direction, that no angle could really matter in so lopsided a fixture: it was a dead rubber, and the tool was manufacturing drama that was not there.
The match was then played. The bottom side drew two-all, the leaders threw away a two-goal lead, conceded in the fourth minute of added time through a mistake between goalkeeper and centre-back, and finished with tempers flaring. The fixture both the tool and the analyst had, in their different ways, written off was the one worth watching.
It would be easy to read this as the tool being vindicated and the analyst embarrassed, and there is some of that. But the real lesson is sharper and cuts against both. The result must not be the judge of the preview, and the tool must never forecast. A preview that connected its material well was well-made whether or not the game bore it out; a buried angle was buried whether or not it later proved decisive. The analyst's error was not bad prediction - it was making a prediction at all, and letting an estimate of the result decide whether the writing had a job to do. The correction the match forced was a principle, not a patch: the whole value of the form is that the unlikely happens often enough to keep a reader watching, so the hardest mismatch is exactly where the craft of opening a door matters most, not least. The tool was instructed, after this, never to weigh an angle by how likely it is to come off, and never to imply a mismatch has no watchable angle. The match did not judge the tool. It judged the temptation to forecast, and that temptation had come from the human, not the machine.
#7. What this says back about the match report
Building the preview model exposed a gap in the report model it grew from. The report had been built on two engines - the reader's attachment, and the economy of the writing - with the social use of sport, the talking point for the pub, treated as a minor reader-type folded into attachment. The preview work made plain that giving the reader something to say is a separate function with a different beneficiary: it serves the reader as talker, not as watcher, and points outward to their standing in a conversation rather than inward to their own feeling. If that is a real, separate function for the preview, parity demands it be one for the report too - and the report is, if anything, the richer source of social currency, because it carries the verdict a reader takes into Monday morning. “We were robbed” is report-fed social currency. The report model is incomplete in the same place the preview model would have been without this correction, and the fix belongs across both tools, not just the new one.
The preview work also clarified what the report's craft actually is by contrast. The report is a performance: the writer executes moves - open on the live thing, pay off what you set up, do not bury the quote that is the story - well or badly. The preview's craft is different. It is completeness and selection: see every angle the fixture offers, work up the options, then choose which to build and which to connect. This difference has a sharp consequence for teaching, taken up below.
#8. What this says back about Andrews
The work began from the standard British textbook treatment of the preview, in Andrews' Sports Journalism, and was at first sharply critical of it. Andrews reserves the word “preview” for a narrow fixture-card data block - kick-off, referee, probable line-ups, a one-line forecast - and files the discursive, appetite-raising piece separately, under features, in another chapter. The early position here was that he had mislabelled the thing: that the scene-setter is the real preview and his fixture card is barely writing at all. The testing did not bear that out, and the honest correction runs in his favour.
His taxonomy was sounder than the merge that replaced it. The four production-objects the real examples forced - the single writer's article, the section-based compilation, the live data-run, the betting card - re-derive, with more granularity, exactly the distinction Andrews drew with two names. His fixture card is, almost precisely, the section-based compilation, and that object turned out to be a real, common, legitimate form, not a failure of one. He intuited that the assembled data-object and the crafted appetite-piece are different things doing different jobs and filed them apart. The model here merged them under one word, called the whole thing a feature, and was then forced by the examples to re-split into four. He was coarse but directionally right; the merge was tidy but wrong.
The student pieces sharpened the point further, and in his favour again. Both were attempts at the single-writer article - the one object that is genuinely a feature - and both produced the feature's surface over the fixture card's substance: continuous prose and a staked headline on top, an unconnected list underneath. Their central fault was precisely a peg-and-angle failure in Andrews' own terms: a headline committed to one line while the body drifted to another, the sprawl his compositional discipline exists to prevent. Andrews' advice - choose your angle, find your peg, hold the line - is exactly what the weaker writers needed and did not have. As teaching for a novice writing the piece, his apparatus is aimed straight at the fault the students made.
Where the criticism stands is narrower than first stated, and it is real. Andrews under-theorises the mechanism: he describes a preview's ingredients and its pegs, but never asks whether the writing connects any of it to a reason to care, so the buried-fact fault - the commonest thing wrong with real previews - is invisible to his account. That gap is the one the analysis tool fills. But the gap is a gap in theory, not in craft advice, and the distinction matters: his compositional discipline is sound, and only his silence about how a preview works on a reader is the limitation. The fair verdict is that he is thin as a theorist and sound as a teacher - and that “the extended preview is a feature” was this project's over-scoped claim, not his. He only ever called the scene-setter a feature, and the examples proved him right to keep the fixture card out of that category.
#9. What this says about building the tool
Several lessons about the AI work are worth recording, because they generalise beyond this project.
#A model is worth what it survives
Every confident conclusion reached inside the discussion was provisional until a real example tested it, and several were wrong. The founding “preview is a feature” framing was wrong; the “no ending” claim was wrong and so was its replacement; the analyst's “dead rubber” judgement was wrong. None of these errors could have been caught by more reasoning. They were caught by examples and, once, by the world. The cheapest and most honest validation available was contact with real material, and it was repeatedly deferred in favour of further elegant reasoning, which is the failure mode to name: a model that is internally coherent across its own documents feels confirmed, and is not. Coherence is not truth, and the more elegant the coherence, the more it should be distrusted until tested.
#Separate the tool's standards from the writer's
A recurring confusion was between the standard the tool holds the writer to and the standard the tool's own output must meet. They are different. The journalist writes to engage an uncommitted reader, and may use a clear cliche or a familiar trope to do it - those are the writer working, not erring, and the tool should read them generously, assuming the writer is trying to engage and may know their readership better than the tool does. The tool, by contrast, writes for a reader who is already engaged - a writer wanting their piece analysed - and so its own prose should be plain and specific, earning no licence for the engagement devices the journalist legitimately uses. Confusing the two led, more than once, to the tool being told to fix in the writer a thing that was actually fine, or to excuse in itself a thing it should not.
#The tool commits the faults it diagnoses
The most persistent problem was that the tool, in its own writing, kept committing the very faults it exists to catch. It used private jargon while being built to flag shorthand the reader does not share. It opened on the empty verdict word “competent” while being built to flag asserted feelings that give the reader no cause to feel them. It narrated its own process - “here is the close” - which is words doing no work for the reader, the exact thing it dings in writers. And it reached for the well-turned phrase in its own headings while telling writers to be plain. Each of these was caught only when a human read the output and asked the tool's own test of it. The lesson is that holding a tool to the standard it enforces is not automatic and is hardest in the tool's own voice, because fluent-but-empty writing reads as competent from the inside. It needs a mechanical self-check - name the empty words, forbid the self-narration - not a principle, because the principle was already there and was not enough.
#Distinguish the tool from the model that runs it
Running the same tool through two different AI models, on the same piece, separated what the tool reliably produces from what one model happens to add. The central findings - the buried fact, the cliche at the charged moment, the three functions - appeared in both, and so belong to the tool. The writing performance differed: one model wrote plainer and honoured the plain-prose rules more naturally; the other reached for atmosphere and had to be restrained. A specific over-committed claim that one model made did not reproduce in the other, which means it was the model reaching, not the tool instructing. This is worth doing deliberately: a tool's real, portable behaviour is what survives a change of the model beneath it, and the rest is disposition that the rules constrain but do not fully override.
#10. The next tool is a feature tool, not a preview coach
The report has two tools: one that analyses for a confident writer, and one that teaches a single move to a beginner. The plan was to build the preview's teaching partner on the same pattern - a preview coach for novices. The attempt to design it ran into an obstruction that, once understood, redrew the whole plan.
The beginner's report tool works because the report has discrete, nameable moves a novice can learn one at a time: open on the live moment, pay off what you set up, do not bury the quote. There is a list of techniques to teach one from. The preview, looked at as a preview, does not decompose this way. Its own craft is completeness and selection - survey every angle the fixture offers, then choose - which is a habit of thoroughness followed by a judgement, not a kit of techniques. So the conclusion was at first that the preview may have no teachable atomic moves at all, and that a beginner is best served by the analysis tool run kindly.
That was looking for the moves at the wrong altitude. The student pieces showed it: their fault was a peg-and-angle failure, which is a teachable move - and it is not a preview move, it is a feature move. A match preview, when it is the single writer's article, is a feature about a forthcoming match, and it inherits the feature's craft: pick your angle and hold it, find your peg, build intro to body to outro, land the ending, do not let the body wander from the line you opened on. Those moves are the same whether the feature is a preview, a player profile, a colour piece or a news feature about a hospital. They are general journalism, not sports-specific, which is exactly why Andrews' peg-and-angle discipline felt oddly generic when the project tried to treat it as preview theory: it was feature craft all along.
So the next tool is not a preview coach. It is a general feature-writing tool, which teaches the feature's moves, and which is far more broadly useful than anything sports-specific. The preview analyser then sits beneath it as a specialist, doing the one thing only a preview needs: surveying the angles a particular fixture offers and catching the unconnected fact. The feature tool teaches you to write features; the preview tool helps you work up this match's options. They are complementary, at different altitudes, and neither is strained into being something it is not.
This also reorders the work correctly: build the general tool first, then let it tell you what is specific. The preview analyser currently carries both kinds of rule mixed together - the genuinely feature-level (surfacing the angle, judging the ending, the voice and cliche work) and the genuinely preview-specific (the three functions, the unconnected-fact fault, the level-and-publication calibration, the no-forecasting rule), because it was built before any feature layer existed to hold the general parts. Once the feature tool exists, the preview tool can shed everything that is really feature craft and keep only what is its own. The same is almost certainly true of the report tools. The shared layer the project kept deferring is not “sports-writing rules”; it is feature and journalism craft, and the report and preview tools are specialisations of it. That layer is best discovered by building the general tool, not guessed at in advance - which is why the feature tool comes first, and the preview tool is rethought against it afterwards.
#11. Where it rests
The model and the tool are left at a defensible pause, not a conclusion. What is reasonably solid: the three functions of a preview; the division of the form into distinct objects judged by different standards; the central fault of the unconnected fact; the two axes of level and publication; the principle that the tool never forecasts and is never judged by the result; and the recognition that the next tool is a general feature tool, with the preview analyser refactored beneath it. What is still open, and marked so: whether caring before a match truly works like caring during one, which is the model's largest untested assumption; whether the tool's range, shown on two student pieces, holds across the fuller spread of what learners get wrong; and what exactly the preview keeps as its own once the feature craft is lifted out of it - a question that can only be answered after the feature tool is built.
The throughline of the whole exercise is a single discipline, applied to writing, to a model, to a tool, and finally to the people building them: say what is proven, mark what is not, and distrust most the things that read most cleanly. A football match enforced that discipline more bluntly than any argument could, by making a fool of a confident prediction in front of everyone. That it judged the analyst and not the tool is the point worth keeping. The tool had been built not to predict - and on the one day the world checked, that was the thing that saved it.
Working paper. Conclusions held provisionally; untested parts marked as untested. For internal reference and later development.