Deterministic replay
Every record carries the code version that produced it, and reproduces byte-identically.
The guarantee
Every derived record stores the code_version that produced it. Computation is a pure function of stored inputs and never of wall-clock time, so reprocessing a range at a given version reproduces the original output exactly.
When a version changes we backfill, and old and new coexist until the new output is verified. Nothing is silently recomputed underneath a customer who already reported from it.
Why it constrains us
This forecloses convenient shortcuts: no in-place mutation, no reading the current multiplier because it is easier to hand, no deriving a value from now(). Those are exactly the shortcuts that make historical data irreproducible, which is why the constraint is worth accepting.
Correctness
If you reported a figure from our data three months ago, you can reproduce it today. If you cannot, that is a bug on our side and we want to hear about it — it is the single guarantee this whole architecture exists to provide.