A twin needs it for the same reason a person would. You cannot decide whether to tip without knowing what the app suggests and what the fee already came to. The environment is how a twin knows the things a person in that situation would already know.
It matters that every twin in a simulation gets the same one. When two twins disagree, you want the disagreement to be about who they are, not about what each of them happened to have looked up.
What is in it
Facts arrive in layers, from general to specific:
About the world, at the date you set. What has happened to prices over the last two years. What people have got used to.
About this group. Facts that hold for your population but not everyone. What the delivery market looks like in Florida metros, and who operates there.
About the situation. Facts specific to the scenario itself. What the checkout screen shows. What the default option is.
The thing you are testing, when there is one. A competitor cutting price fifteen percent. Your new pack size on shelf. This is kept separate from the background on purpose — it is the change you are studying, and mixing it into the scenery is how a comparison stops being attributable.
What twins are told, and what they are not
Every fact carries a source, and you can see all of them. But not every true fact gets given to a twin, because a fact you hand someone is a fact you have asserted they know.
Tell a twin that the average delivery fee across Florida metros is $3.99 and you have invented a person who knows a statistic almost nobody knows, and who will reason like an analyst rather than like someone at a checkout screen. So facts are rated by how widely known they actually are. Ones a normal person would not know are held back, and ones people half-know are given the way people hold them: "delivery usually adds about four dollars before you get to the tip."
The precise, sourced version of every fact is still there for you to read. It is just not what the twin gets.
Where the facts come from
Some are written by hand, which is the right approach for the handful of facts a scenario really turns on. Some are researched, in which case every fact carries the source and the passage from it that supports the claim. Some come from you — your prices, your launch dates, your product specifics — which override anything general on the same subject, visibly.
Facts we cannot point at a source for are marked as assumptions rather than quietly presented as established.
Getting it right
Two checks matter more than the rest.
Nothing from after the date. A fact only applies if it was already true at the moment your scenario is set in. A twin reasoning about last August that knows what happened in November will produce a confident, wrong, untraceable answer. Facts are dated, and anything that had not happened yet is excluded.
No contradictions. Two facts that cannot both be true will split your population for a reason that has nothing to do with your question. They are caught before the simulation runs, not discovered in the results.
Seeing and changing it
You do not fill in an environment when you set up a simulation. It comes from your population and your date. But every result shows exactly what its twins were told, with sources, and you can inspect it, correct a fact, or add your own before running.
This is worth doing at least once. "63% will tip" is a number. "63% will tip, given these eleven things we told them were true, of which nine are cited and two are assumptions" is something you can actually argue with — and an argument about a specific fact is the kind that improves the answer.
One rule when you are comparing runs: all sides of a comparison share the same environment. If the background could move between runs, the difference between them would not be attributable to the thing you changed.