An audit found 415 cases where tactica. attributed part of a prediction to something that hadn't actually affected it. Fixing the explanation was straightforward. Deciding not to change the model was more important.
We found 415 cases where tactica. recorded a finishing adjustment as part of the explanation for a football prediction.
There was one problem.
That adjustment hadn't actually changed the prediction.
Not 415 incorrect predictions.
Not 415 bad expected-goals calculations.
But 415 explanations that claimed more about the cause of a prediction than the underlying calculation justified.
For a football intelligence product built around showing its reasoning, that's a serious distinction.
If tactica. tells someone why it thinks something about a match, the reason it gives has to be a reason the model actually used.
Ours wasn't always doing that.
And fixing it presented us with a choice.
The number existed. It just hadn't been used.
The problem came from a finishing contribution calculated inside part of tactica.'s expected-goals process.
In the affected situations, some of the player-quality evidence the model needed wasn't available.
The system had already calculated a finishing value.
But because that quality evidence was missing, the calculation followed its existing fallback path without applying the finishing contribution to the final expected-goals figure.
The prediction itself behaved as designed.
The explanation didn't.
It carried the finishing value forward as though it had contributed to the result.
That's how we ended up with 415 cases where tactica. effectively said:
This affected my prediction.
when the calculation showed:
No, it didn't.
There was an easy way to make the discrepancy disappear
We could have added the finishing contribution into the expected-goals calculation.
Problem solved.
The number would affect the prediction.
The explanation would become true.
And we'd have changed the model without evidence that doing so made the model better.
That's backwards.
Our audit had established something very specific:
The explanation was inaccurate.
It had not established:
The model should use this finishing contribution.
Those conclusions might look similar when you're staring at the discrepancy.
They're not.
One is evidence about how accurately we're describing an existing calculation.
The other would require evidence that changing the calculation itself improves the football model.
We had the first.
We didn't have the second.
So we left the model mathematics alone and corrected the explanation instead.
If it didn't move the prediction, don't say that it did
That became the simplest way of describing the new rule.
Behind it is a more formal distinction.
tactica. now separates information into different states.
Something can be applied: it actually affected the calculation.
Something can be computed or contextual: the system calculated it or knows about it, but that doesn't mean it changed the published prediction.
And something can be unknown: we don't possess enough evidence to make the claim at all.
That distinction matters because displaying more information isn't automatically the same thing as providing more transparency.
You can show someone twenty numbers and still mislead them if you don't distinguish between what the model knew, what it calculated and what it actually used.
Our standard needs to be simpler:
If something didn't move the prediction, tactica. shouldn't tell you that it did.
The same problem appeared elsewhere
Once we started being stricter about what tactica. was entitled to claim, the principle extended beyond finishing.
Take injury information.
Previously, missing injury evidence could effectively become Available.
Missing injury coverage could effectively become zero injuries.
Neither conclusion follows from the evidence.
If we don't have an injury report for a player, that doesn't prove he's fit.
So we changed the meaning.
Missing evidence is now Unknown.
Missing injury coverage remains unknown rather than becoming zero.
Positive evidence can establish an injury.
Absence of evidence cannot establish fitness.
It sounds like a small semantic distinction.
It isn't.
Imagine reading a Match Intelligence report and seeing a player described as available. You reasonably assume tactica. possesses evidence supporting that statement.
“Available because we have evidence he's available” and “Available because we haven't found evidence he's injured” are fundamentally different claims.
The product needs to know the difference.
A current team sheet isn't necessarily model evidence either
Lineups exposed another version of the same problem.
Suppose an official team sheet becomes available before you're reading a tactica. Case File.
That's a current football fact.
But its presence in the product does not automatically prove that the lineup influenced a prediction generated earlier.
Those two clocks matter.
What do we know about the match now?
and:
What did the model demonstrably know when it made this prediction?
are different questions.
W1 tightened that boundary too.
The same principle applies to bookmaker odds.
Odds remain downstream execution information. They didn't become an input into what tactica. thinks about the football match as part of this work.
Again, this isn't about displaying less information.
It's about being precise about the authority each piece of information has.
Then we had to prove we hadn't accidentally changed the model
Saying we only changed the explanation isn't enough.
We needed to demonstrate it.
So we compared the system before and after the correction across 313 cases.
Expected goals remained identical.
Home, Draw and Away probabilities remained identical.
BTTS probabilities remained identical.
Count projections remained identical.
Market projections remained identical.
With fixed inputs, recommendation membership was preserved too.
That doesn't prove tactica. predicts football perfectly.
It doesn't prove the probabilities are perfectly calibrated.
And it doesn't tell us anything about future profitability.
It proves something narrower, and that's exactly the point:
The W1 correction changed tactica.'s right to make the explanation without changing the underlying predictions across the tested comparison corpus.
That's what we intended.
So that's what we tested.
We didn't let ourselves declare victory
The correction then went through the normal release process.
Automated checks passed.
An independent review checked that the model mathematics hadn't changed, that the correction had happened in the appropriate part of the system, and that unrelated future work hadn't leaked into the change.
It was reviewed, merged and deployed.
But there was still one piece of evidence we wanted.
Production.
We could have manually forced tactica. to generate a convenient prediction and used that as proof.
We deliberately didn't.
We waited.
Then football generated the test for us
On 28 August 2026, tactica.'s normal production operation generated fresh Match Intelligence.
No specially constructed demonstration.
No hand-picked test fixture.
Just the system doing its ordinary job.
And it naturally produced both sides of the behaviour we wanted to verify.
In one case, the necessary player-quality evidence was missing.
tactica. still calculated finishing values.
But this time it correctly recorded them as computed, not applied.
The applied finishing contribution remained zero, and the published expected-goals values reconciled with the calculation that had actually run.
In another naturally generated case, the required quality evidence was present.
There, finishing was legitimately used in the calculation.
That's important.
We hadn't simply removed finishing information everywhere to make the problem disappear.
We had corrected when tactica. was entitled to say that finishing had affected the prediction.
That was the production evidence we were waiting for.
Transparency isn't the number of numbers on the screen
This work came directly out of the audit we wrote about in our first Journal article, Our model was profitable. We audited it anyway.
That audit started from a simple problem.
tactica.'s early observational Performance was positive.
But profitability couldn't tell us whether everything underneath that result was healthy.
This was one of the things we found.
And it sharpened something about what we mean when we call tactica. transparent.
Transparency isn't putting a complicated model output on a screen.
It isn't showing every intermediate number the system happens to calculate.
It isn't writing a convincing football explanation after the prediction has already been made.
The explanation has to correspond to what actually happened inside the model.
And when the evidence doesn't support a claim, the product has to be capable of saying:
We don't know.
That's less impressive than manufactured certainty.
It's also more useful.
Evidence earns change
The most important decision in this story isn't that we corrected 415 explanations.
It's what we refused to do when we found them.
We could have changed the model to make the existing explanation true.
We didn't.
Because finding an inconsistency doesn't give us permission to make whichever model change makes the inconsistency disappear.
The evidence has to tell us what kind of problem we've actually found.
In this case, it told us:
The explanation was wrong.
So we changed the explanation.
If future evidence tells us the model mathematics should change, we'll change those too.
But that evidence has to arrive first.
Evidence earns change.
Sometimes the difficult part of auditing a model isn't finding something wrong.
It's resisting the temptation to fix the wrong thing.