The finding that changed at 48 months

Fifteen assumption keys get argued over before a run. The horizon gets typed once, usually as whatever the function does when you leave the second argument off, and then never appears in the output. Marchmont's engine defaults to 24 months. Nobody chose 24 months; it was chosen in the source file by whoever wrote the signature, for no reason connected to the question being asked.

Treat it the way the growth rate is treated. It has a value, the value changes the answer, and the value belongs next to the answer. A figure quoted without its horizon cannot be reproduced by the person reading it, which means it cannot be disagreed with, which means it is not yet a finding.

One reader, two horizons. The variable that matters is the last one, and its name is doing real work: back returns zero when the balance does not come back inside this window, which is not the same statement as never.

Both rows are correct. The trough is the same month and the same figure in both, because extending a horizon cannot change a month the engine has already computed. Only the last field moves, and the whole headline hung off it.

The recovery reader returns 0 when it finds nothing. That is a convenient thing for a function to do and a dangerous thing for a reader to receive, because 0 is a perfectly good month number in a context where months are integers, and because "no recovery found" reads at a glance as "no recovery".

The fix is not clever. Return something that cannot be mistaken for an answer, or return the horizon alongside it so the reader has to look at both. A field called back with a value of 0 and no horizon next to it is a fact that has been stripped of the only context that makes it readable, and that stripping is where the draft went wrong rather than anywhere in the arithmetic.

The engine steps forward month by month from a solved opening sheet, and no month reads a later one. So months 1 to 24 of the 48 month run are the same values as months 1 to 24 of the 24 month run, and the trough at month 19 is the same trough in both. The horizon is a pure reporting choice: it changes what you can see and nothing about what happened.

That is worth confirming once by hand rather than believing, because it is the property that makes extending a window a cheap check. If extending the horizon changed earlier months, comparing two windows would not be a test of the reading, it would be a test of the model, and a much more expensive conversation.

Run it, run it again at one and a half times the length, and compare the sentence rather than the figures. The figures are supposed to change when the window gets longer, because there is more of them. The conclusion is not.

This is the cheapest test available and it is the one that would have caught the draft on its own. It also has a useful failure mode: if the conclusion is stable from 24 to 36 to 48 months you have learned something worth putting in the write-up, which is that the claim does not depend on where the window ends.

If the deepest month in a run is the last month of the run, the model has reported where you stopped looking. Marchmont's trough at 4% growth sits at month 19 inside a 24 month window, which is why the depth in the dead draft was right. At 8% growth the trough is -£7,885,358 in month 48 on a 48 month run, which is the same shape of unanswered question one order of magnitude further along.

The check is a comparison, not a judgement, so it can be an assertion in the bench rather than something a reader has to remember. Any run whose extreme month equals its horizon gets flagged and does not get quoted until a longer run agrees with it.

Draft the sentence you would publish if your claim were false, in full, before running anything else. For the dead headline the opposite is "a profitable company with a working capital gap comes back on its own if you wait long enough". Having written that down, the next question asks itself: what run would produce it, and has anybody done that run.

The value of writing it out is that a vague opposite is easy to dismiss and a specific one is not. "It might recover" invites a shrug. "It recovers in month 28 after eighteen months under" is a claim with a shape, and a shape is a thing you can go and test in one call.

Extending the window fixes a claim about time. It does nothing for a claim that was read off a single setting, and the canon set exists for that reason: five growth rates crossed with five collection settings, each run to both 24 and 48 months, fifty runs with the worst tie error at 2.24e-8.

Reading all ten rows is what stops the correction becoming the next error. The 48 month run at 4% says Marchmont recovers. It does not say that companies like Marchmont recover, and the rows underneath it say plainly that most of them do not.

Go through the draft claim by claim and mark each one with the run that supports it. The Marchmont draft had four claims in it, and two of them were never in danger.

Profit and cash diverge, hard, for a long time. Twelve month net income of £445,631 against free cash flow of -£297,261, a gap of £742,892, of which £664,607 is the working capital build. Unchanged by the horizon, because it is a twelve month total inside both windows.

The gap is real and has to be funded. The balance goes under in month 10 and stays under for eighteen months. Recovering in month 28 does not help anybody in month 19, so the funding requirement survives the correction completely.

The company does not recover. Dead. The correction run produced month 28.

Therefore borrowing is only a deferral. Dead, and it was the most damaging of the four, because it is the one that would have led somewhere. It was a conclusion built on the claim above it rather than on any run of its own, which is the pattern worth looking for: a claim with no run attached, sitting downstream of one that has.

Everyone agrees the corrected number goes out. What to do with the dead one is an argument, and it is not a soft argument about tone.

One position: publish the trail, including the claim that died and the run that killed it. A correction is the strongest available evidence that the work is being checked, and the near miss teaches more than the result does, because the result looks obvious once you have it and the miss shows how it nearly did not happen. Suppressing it means the next person runs 24 months and reaches the same place, and the audience learns nothing about what a window can do to a conclusion. On this view an analysis with no visible corrections is either lucky or not being checked.

The other position: a document carrying every abandoned claim is unreadable, and the abandoned claim is the part that gets quoted. "A profitable company runs out of cash and dies" is a better sentence than anything that replaced it, and printing it next to its retraction guarantees it travels on its own, stripped of the correction, into somebody else's deck. Readers act on documents, not on their footnotes. On this view the retraction belongs in a change log where the people who need it will look, and the published page carries the number people are meant to use.

Both sides have watched their own failure happen: one has seen the same mistake repeat because nobody recorded it, the other has seen a dead headline outlive its correction by years. There is no run that settles it, and the honest move is to say which risk you are choosing and why, rather than presenting a house style as a principle.

Whichever side of the argument above you land on, a correction has a shape that travels better than prose. Name the claim, name the run that supported it including its horizon, name the run that killed it, and name what is left. Four lines, each of them checkable by somebody who has the engine.

The detail that does most of the work is putting the horizon inside the dead claim rather than in a note about it. "Over a 24 month run, Marchmont never recovers" is a sentence that carries its own limitation wherever it goes, and it is much harder to quote out of context than "Marchmont never recovers, see footnote". The condition is not an apology for the claim, it is part of what the claim says.