A sandbox is an isolated SET OF CHANGES you can publish or discard. It is not a separate environment, and it is not a backup. Those two misreadings cause most of the trouble in this module.
NOT AN ENVIRONMENT. You are still in the same instance, against the same data. A sandbox captures configuration changes, not transactions — so anything you do to actual records while inside one is real and stays real when you discard.
NOT A BACKUP. Discarding a sandbox reverses what the sandbox holds, and nothing else. It cannot undo a change somebody made outside it, and it cannot undo a published sandbox.
WHAT IS CAPTURED: page personalisations, flexfield definitions, Application Composer objects and fields, and the other configuration the tooling explicitly places in a sandbox context.
WHAT LEAKS AROUND IT is the part to internalise, because it is the source of every surprise here. Value set values, reference data, scheduled processes you submit, and anything you change through an ordinary setup page while a sandbox is active are environment-wide — visible to everybody, and unaffected by discarding.
So the working rule: a sandbox contains the SHAPE of things, not their CONTENT. Adding a field is in the sandbox; the data typed into it is not. Test that boundary once, deliberately, in an environment you own — change one thing of each kind, discard, and see what came back. Five minutes, and it settles a question that otherwise gets guessed at under pressure.