Documentation
Getting started
Nine steps from an empty account to a brief you can disagree with on specifics. The last two are the ones that matter, and most people skip them the first time.
You can do steps one to four with no files of your own at all.
The nine steps
Each one has its own screen in the app. You can stop after any of them, look at what it produced, change it, and carry on.
Make a twin
Name, industry and currency. The industry choice sets the starting priors for elasticity, churn, margin and contract length, all of which are visible in the ledger and all of which you can overwrite. Nothing here is a commitment: the industry can be changed later and the twin rebuilt.
Load the worked example
Harborline Components is invented: a contract manufacturer with 24 customers, 14.24 million dollars of revenue, 31 percent gross margin, 46 people, a largest customer at 18 percent, a top five at 54 percent, one single sourced marine grade billet and a price it has not moved since 2023. It arrives with sources, facts, a graph, agents and seven scenarios already in place. Use it to learn the screens before your own numbers go in.
Upload your first three files
A profit and loss or management accounts, tagged as financials. A customer list with one row per account and a revenue column, tagged as customers. Whatever you have on competitors, even a page of notes. Those three move the readiness score more than anything else you can do in ten minutes.
Check what was read out of them
Open each source. Every fact carries the quote it came from, so a wrong number is traceable in one click rather than arguable. A PDF with no text layer is reported as needing text instead of quietly producing nothing.
Rebuild
Rebuild re-runs extraction, rebuilds the graph, regenerates agents and refreshes the assumption ledger. Anything you have edited and locked survives it. Run this after every batch of uploads, not after every file.
Read the ledger before the answer
The ledger lists every number the simulation runs on with its origin: read from a document, worked out from something that was, estimated by a model, set by you, or an industry default. Defaults sort to the top. That ordering is the software telling you where its own weak points are.
Run the price scenario
Price up 20 percent ships as a template with its levers already set. A standard run is 200 replications of the scenario plus 200 of the baseline on the same seeds, which is 400 simulations, and takes about a tenth of a second for a twelve month horizon with around forty agents on the machine this was built on.
Read the brief, including the last block
Six blocks: the answer, the path, the mechanism, the risk, the tripwires, and what the whole thing rests on. Most people read the first and stop. The last one tells you how much the first is worth.
Fix the top default, then run it again
Go back to the ledger, take the first default on the list, replace it with a number you can defend, and run the identical scenario. If the median barely moves and the band tightens, you have bought certainty. If the median swings, you have just found the assumption the decision was actually resting on.
Step nine
What replacing one default looks like
Things that go wrong in the first hour
None of these is a fault. They are just the four places people lose twenty minutes.
- 1
A scanned PDF
The parser looks for a text layer. A scan has none, so the source is marked as needing text with a note saying so. Paste the numbers or upload the spreadsheet behind it. The app will not pretend a picture of a page is a page.
- 2
A customer list with no revenue column
Names alone give you named accounts with no size, so concentration cannot be computed and the lose largest customer scenario has nothing to remove. Any column called revenue, MRR, ARR, contract value, billing, spend or amount will do.
- 3
Everything tagged as other
The kind you choose at upload decides which reader runs and which part of the model the facts feed. A customer list tagged as other still gets read, but it will not become segments and named accounts.
- 4
Judging the first run
A twin built from three files is standing on defaults for half its parameters, and the brief says exactly which ones. The first run is a map of your own ignorance. That is useful, but it is not yet a forecast.
After the first hour
Three things worth doing in week one
Once the first brief exists, the useful moves are narrow and specific rather than more uploading.
- Run the sensitivity on your main scenario. It moves each load bearing assumption to the ends of its own band and reports how much the answer moves, so you learn which two numbers to go and find
- Run a sweep on price. Twenty five settings at eighty replications is two thousand full simulations and returns a curve with a best point rather than a single guess
- Compare up to eight scenarios against one baseline on one set of seeds, so the differences between them are the decisions and not the noise
Next: what to upload
Thirteen kinds of file, what each one gives the model, and the exact column headings the extractor looks for.