Rounding a central value can push it outside its own unrounded min/max
A pydantic model with a low <= expected <= high validator rejected its own input on 8% of real production records, while passing every fixture.
The construction was:
Ranged(low=min(h), expected=round(sum(h)/len(h), 4), high=max(h))
When the observed values are all equal or nearly so, round(mean, 4) can land just ABOVE the unrounded max(h) (or below min(h)). Example from the real data: low = high = 2.75795841217041, mean rounds to 2.758, which is greater than high, so the validator raised.
This never fires on hand-written fixtures because those use already-short decimals (2.75, 3.0) where rounding is a no-op. It fires on measured float data, where the full binary expansion is present. Our golden examples all passed; 5 of 61 real records crashed.
Fixes, in order of preference:
1. Do not round one member of a tuple that has an ordering invariant. Round all three or none.
2. If you must round for display, clamp: expected = min(max(rounded, low), high).
3. Round low down and high up (math.floor/math.ceil at the same precision) so the interval only ever widens.
General rule: any invariant of the form a <= b <= c is broken by rounding b independently of a and c. Grep for round( near a validated interval, percentile set, or sorted triple.
Also worth taking from this: a crash rate of zero on curated fixtures tells you nothing about real inputs. Running the pipeline over a few dozen real cached records took 62 seconds and surfaced three distinct crash classes that the whole fixture suite missed.