Documentation
Reading a brief
Six blocks, written by the engine with no language model involved. This page is how to read each one, and then how to argue with all of them.
A model can rewrite the prose if one is configured. It never supplies a number.
The six blocks
1. The answer
Revenue, operating profit, customers and cash at the end of the horizon, each with the percentage against doing nothing. Every figure here is a comparison against the same twin running the same seeds with no scenario applied, so the difference is the decision and not the noise. The lead sentence also gives cumulative operating profit across the whole window, which is often a different story from the month twelve figure.
Read the customers line next to the profit line. When profit is up and customers are down, you are looking at a move that improves the accounts while shrinking the base underneath them, and the second year is where that argument gets settled.
2. What happens, and when
The path, sampled at about six points across the horizon, with revenue, operating profit, the do nothing line, customers and cash. Above it, two sentences: the worst month and how far below the do nothing line it goes, and the month the scenario gets ahead. If it never gets ahead inside the window, the brief says that plainly, and that is the most important sentence it can produce.
3. Why it happens
Up to eight agent moves, each with the month, the agent by name, its class, what it did, and the frequency. This is the block that separates a simulation from a projection. A projection says revenue falls. This says a named account did not renew in month seven, and it happened in 62 percent of the runs.
4. What could go wrong
The bad case and the good case for revenue, the lowest point cash reaches anywhere on the path, and the share of runs that ended out of cash. The cash minimum is the one operators should read first, because a plan that works on average and dies in month eight is not a plan.
5. What to watch, and by when
The tripwires. Specific signals, the month to check them, what the model expects, and what would count as off track.
6. What this rests on
The load bearing assumptions with their value, band, origin and the quote from the file they came from. Industry defaults sort to the top. The lead sentence counts how many of the assumptions behind the run came from something you uploaded, how many were derived from those, and how many are defaults doing a job nobody has given them evidence for.
The band
What two hundred replications look like before they become three numbers
How to read a band
The p10 and p90 are percentiles of the replications at each point. One run in ten came in below the p10 and one in ten above the p90. It is not a confidence interval in the statistical sense, and it makes no claim about the real world. It is the spread the model produces when its own stated uncertainty is taken seriously.
Width is information, not a defect. A narrow band means the parameters driving this answer are well pinned down by your record. A wide one means the answer is being driven by things nobody has evidence for, and the fix is in the ledger rather than in more replications.
Two rules that will save you from most bad readings. First, if the band crosses zero, the model is telling you it does not know the sign of the answer, and no amount of median gazing changes that. Second, a band on a percentage change is not the same shape as a band on a level, so compare like with like.
The one thing the band cannot do is cover an assumption that is simply wrong. If your elasticity is genuinely minus 2.4 and the ledger says minus 1.3 with a band from minus 1.0 to minus 1.6, every replication in the run is wrong in the same direction. That is what the ledger and the sensitivity exist to catch, and it is why the last block of the brief is not optional reading.
Frequency
What it means when an event says 62 percent
Every event in the mechanism block carries a frequency. It is the share of the replications in which that event happened at all. The month shown next to it is the median month across the runs where it did happen.
So a competitor matching part of your move at 88 percent is close to a certainty inside the model, and its timing is the thing to argue about. A named account walking at 30 percent is a real risk with a specific name on it, and the honest way to hold it is as a risk rather than as a forecast.
Frequencies are not probabilities about your actual company. They are probabilities inside a model whose parameters you can read in the ledger. Treat a high frequency as a statement about the mechanism being robust to the uncertainty in your numbers, which is a genuinely useful thing to know and a different thing from a prediction.
- Above 80 percent: the mechanism holds across nearly every draw, so plan for it
- Between 30 and 80 percent: it depends on which draw you get, so the assumptions behind it are worth pinning down
- Below 20 percent: it is not filtered out of the brief because a rare event with a large severity is still worth naming, but do not build the plan around it
How to use the tripwires
A forecast you cannot check is a wish. Tripwires are the earliest points where the real world tells you whether the scenario is behaving like the model said, and each one comes with a month, an expected value and an alarm condition.
The standard four are monthly churn at about a quarter of the way through the horizon, new customers per month at the same point, competitor pricing two months later, and operating profit against the do nothing line. Any agent event with a frequency above 20 percent adds a fifth kind: watch for the early signs of that named move, and if they show up sooner than the model said, your timeline is running fast.
The way to actually use them
- Copy the four rows into whatever you review monthly, before you start the decision, not after
- Write the alarm condition down as a number, not as a feeling, because the whole point is to remove the later argument about whether things are going badly
- Decide in advance what you will do if one trips. A tripwire with no prepared response is just a way of being unhappy earlier
- Re-run the scenario when one trips, with the observed value put into the ledger. The band will narrow and the answer may change sign
The most valuable tripwire is usually the operating profit gap at the quarter mark, because it is the cheapest moment to stop. Most bad decisions are not bad at the moment they are made. They become bad at the moment they stop being reversible, and that is almost always later than the point where the first evidence arrived.
How to argue with it
The brief is not evidence. It is an argument with its premises listed, which means it can be attacked in specific places rather than dismissed in general.
Start at the bottom, not the top
Open the last block first. Count the defaults. If eight of the ten load bearing assumptions are industry defaults, you are reading a well formatted industry prior with your revenue number on it, and the correct response is to go and get two real numbers rather than to debate the median.
Check the quote, not the value
Every assumption with a document origin carries the line it was read from. Read three of them. Extraction gets things wrong: a total row read as an account, a percentage read as a ratio, a year read as a quantity. It takes two minutes and it is the highest yield check available.
Run the sensitivity
It moves each load bearing assumption to the ends of its own band, one at a time, and reports how much cumulative operating profit moves. The result is a ranked list of which numbers the answer is actually standing on. Usually two of them dominate and the rest are noise, which tells you exactly where to spend the next hour.
Change the seed
Runs are reproducible on purpose: same seed, same answer. Change the seed and re-run. The median should move a little and the story should not. If the mechanism block changes character between seeds, the replication count is too low for the question you are asking.
Attack the timing
The model is generally better at the shape and the order of magnitude than at the month. If the whole argument rests on something landing in month four rather than month seven, that is a weak argument regardless of how confident the chart looks.
Ask what is missing
The method page lists what is deliberately not modelled: general equilibrium, inventory, working capital timing, tax, seasonality beyond a flat market growth rate, and anything below department level. If your decision turns on one of those, the brief is not the right tool and no amount of reading it carefully will fix that.
The test of a brief is whether it changed what you did
If it did not, either the decision was already made or the model has nothing to add to it. Both are worth knowing in an afternoon rather than a quarter.