fitted ought to check that any new data it receives is compatible with the model

[null-a]

I'm aware of one specific problem. It occurs when fitted is passed a data frame in which a categorical column includes a level not present in the data on which inference was performed. In this situation, rows in which the categorical column takes on this superfluous level are coded incorrectly.

import pandas as pd
from brmp import defm
from brmp.fit import fitted
from brmp.design import predictors

df = pd.DataFrame(dict(
    y=[0.,1.],
    a=pd.Categorical(['a0','a1']),
))

fit = defm('y ~ a', df).fit(iter=5)

df2 = pd.DataFrame(dict(
    # level 'a2' is not present in df
    a=pd.Categorical(['a0','a1','a2']),
))

# `fitted` does something like this internally:
data = predictors(fit.formula, df2, fit.metadata, {})

print(data['X'])
# [[1. 0.]   # a0
#  [0. 1.]   # ... and a1 are coded correctly,
#  [0. 0.]]  # ... but the new level is a row of zeros

It's possible that there are further notions of compatibility that we ought to check for.