The brief
Six blocks, in the same order, every time
A number on its own is not an answer. The brief says what happened, when it happened, which agent caused it, how wide the band is, what would have to be true for it to be wrong, and which line of which file the load bearing assumptions came from.
Written by the engine. A language model, if you have one configured, only rewrites the prose.
The order is fixed so that the uncomfortable parts cannot be edited out.
Every brief carries the bad case, the tripwires and the list of assumptions, in that order, whether or not the answer is good. A report format where the risk section is optional is a report format that will not have one when it matters.
The six blocks
What is in each one
The answer
The question you asked, then revenue and operating profit at the end month against doing nothing, then the customer count, then what the whole window is worth in cumulative operating profit. Four headline figures, each with the percentage against the baseline underneath it. If profit is up while customers are down, the block says so in that sentence rather than leaving you to notice.
What happens, and when
The worst month and how far below the do nothing line it went, and the month the scenario first gets ahead of doing nothing. If it never gets ahead inside the window, the brief says that is the most important sentence in it. Then a table of the path: revenue, operating profit, the baseline profit, customers and cash, sampled across the horizon and always including the final month.
Why it happens
The agent events, ranked by how often they happened multiplied by how serious they are. Each carries the month it typically lands, the agent that did it, what it did, and a frequency: the share of the replications in which it happened at all. If no agent crossed a threshold, the block says so plainly and tells you the result is the mechanical response on its own.
What could go wrong
The bad case and the good case for revenue, the lowest point cash reached anywhere on the path, and the share of replications that ended with cash below zero. When that share is more than a rounding error the brief calls it a plan with a failure mode. When the band on revenue is wide relative to the answer, it says the result is being driven by assumptions your record does not pin down.
What to watch, and by when
Tripwires. Each one names a signal, the month to look at it, what the model expects to see, and what would count as off track. They are set early in the horizon on purpose, because the point is to find out cheaply rather than to be proved right expensively.
What this rests on
The load bearing assumptions with their values, their bands, their origins and the quoted line where there is one. Defaults are sorted to the top, ahead of everything read from a document, because a default under a six figure decision is the single most useful thing the software can point at.
Block one
Against doing nothing, not against today
Every headline figure in the answer block is measured against the same twin doing nothing, run on the same seeds. That is a different and much more useful comparison than measuring against the month you started.
A business that was already growing will show revenue up in every scenario, including the bad ones, if you measure against the starting month. Measured against the baseline, the bad ones show up as bad. A business that was already shrinking will show revenue down in every scenario including the good ones, and the baseline comparison catches that too.
The cumulative figure sits alongside the ending figure for the same reason. A move that is worth a lot in month twelve but cost you for nine months before that is a different decision from one that paid from the start, and only the cumulative number separates them.
Block three
A mechanism with a frequency attached to it
This is the block that separates a simulation from a forecast. A forecast tells you revenue falls. This tells you that in month four a competitor matched part of your move, in month six an account with an eighteen percent share did not renew, and in month seven your own finance lead cut marketing because runway had fallen.
Every one of those carries a frequency. An event in ninety percent of replications is something to plan around. An event in twenty percent is a risk to price. An event in three percent is noise, and the brief ranks it accordingly rather than reporting everything that ever happened.
The first replication of a full run also records what the agents said and why, in their own terms, with the objective and the weight that drove it. It is not a language model writing dialogue. It is the decision rule, printed.
Block five
A forecast you cannot check is a wish
Anatomy
What a tripwire row says
| Field | What it holds | Example of the reasoning |
|---|---|---|
| Signal | The observable thing to look at | Monthly churn, new customers per month, competitor pricing, or the profit gap against doing nothing |
| When | The month to look | Set early in the horizon, because the cheapest moment to stop is before most of the cost has been spent |
| Expect | What the model says you should see at that month | Taken from the median path, not from the optimistic end of the band |
| Alarm | What would mean the model is wrong, and in which direction | Churn well above the expected level means the customer response is worse than modelled and the rest of the brief is optimistic |
The most significant agent events get tripwires of their own, with the month they typically land and a note that if the early signs arrive sooner, the timeline in the brief is running fast.
Block six
Defaults first, because those are the ones worth replacing
The last block ranks the assumptions that matter to this kind of answer, then sorts every industry default above everything that came from your record. It is the opposite of how a consulting deck orders its appendix, and it is deliberate.
Alongside it the brief states the arithmetic in words: of so many assumptions behind this run, so many came from something you uploaded or typed, so many were worked out from those, and so many are industry defaults doing a job nobody has given them evidence for.
That sentence is the honest summary of how much to trust the rest of the page, and it is generated from the ledger rather than written by anybody.
The language model
It rewrites the prose. It never touches the numbers
The brief is assembled by the engine from the run: the blocks, the figures, the events, the tripwires and the assumption rows are all computed. A complete brief exists and can be read with no key configured at all.
If a key is configured, the finished analysis is handed to a model with a narrow instruction: rewrite this as plain, direct prose, keep every number exactly as given, never add a number, never soften a bad result, lead with the answer, no em dashes, no jargon, around three hundred and fifty words.
It is given the finished analysis, not the run. It cannot recompute anything, it has nothing to recompute from, and the underlying blocks are stored next to the prose so you can always read what it was working from.
- The numbers are decided before the model sees anything
- The structure is fixed, so it cannot drop the risk section
- The stored blocks are the record. The prose is a presentation of them
Every brief is stored with the run that produced it and the seed that produced the run, so a brief you circulated three months ago can be regenerated exactly, and the changes since then are changes you made to the twin rather than drift in the software.
Get one you could put in front of a board
The test worth applying is simple. Hand it to the most sceptical person in the room and see whether they can find the assumption they want to argue about. In this format they can, because it is in block six with the quote next to it.