The things people ask, answered plainly

It reads the documents you already have, builds a model of your company from them, gives every customer, competitor, supplier, salesperson, executive, employee, investor, regulator and partner its own objectives, and then runs the year forward hundreds of times with different draws from the uncertainty in your own numbers.

What comes back is a median with a tenth and a ninetieth percentile around it, a list of what the agents did and how often, and a brief naming the assumptions the answer is standing on.

A spreadsheet gives you one path and has nobody in it. It has no competitor who matches part of your move in month four, no salesperson who starts discounting when they are behind, and no customer whose contract happens to end in the month you send the price letter. Those are the reasons a price rise rarely goes the way the spreadsheet said, and they are what this adds.

No. The simulation is arithmetic and runs with no API key at all. Given a seed it produces the same answer every time.

A language model does three narrow jobs at the edges, all optional: reading prose the rules could not parse, rewriting a finished brief into better English, and answering questions in Ask. None of them touches the simulation.

Its accuracy is the accuracy of its assumptions, so they are all listed with bands and origins. It is good at shape, order of magnitude, timing, which actor causes the damage and how wide the uncertainty is. It is bad at predicting a market, at one off events, at anything your files do not contain, and at long horizons.

There is a whole page on this, so read it before you pay for anything.

Then the model says so. It runs on industry defaults and the readiness score tells you how much of the answer is standing on them. The missing files are listed in the order to fix them. A twin built from one profit and loss and a customer list is already more useful than an argument.

A standard run is 200 replications of the scenario and 200 of the baseline, which is 400 simulations, and takes about a tenth of a second for a twelve month horizon with around forty agents on the machine it was built on. A 25 step sweep at 80 replications is 2,000 simulations.

Yes. The seed is stored with the run, so the same twin, the same levers and the same seed produce identical numbers. That is what makes it possible to change one assumption and know that what moved was the assumption.

No. No backtest against known outcomes has been done or published. When it is, it will be on the changelog with the numbers, including the cases where the model was wrong. Until then the calibration is untested and this site says so everywhere it matters.

Yes, on Enterprise. The whole application runs as a private deployment on a server you control, and nothing leaves your building unless you set up a language model key, a hook or a connection yourself.

An ordinary web server with a database, which most hosting already is. Nothing else: no queue, no separate service, no graphics card. It is set up with you as part of Enterprise.

There is an API that the application itself uses, authenticated as a signed in user, and feed keys that let a finance system post files and actuals to one twin. There are no general API tokens for third party use yet, which means reading a twin from outside the browser is not supported today. It is on the roadmap and listed on the status page.

On a private deployment, yes: the source is on your server, readable and not obfuscated. Read the extraction rules, read the engine, and check that the claims on this site are true. That is a deliberate property rather than an accident.