Where does black box collide?

In cybernetics a black box is any system an investigator can act on and observe but cannot open: knowledge of it is built entirely from the record of inputs applied and outputs returned, never from its internal mechanism. Ashby's An Introduction to Cybernetics sets out the black box as a general method of inquiry, showing how repeated probing lets one infer the system's behaviour and its possible internal states without ever seeing inside. The referent is an epistemic stance toward a system (study it through its input-output relation) and it has nothing to do with a software test plan or an uninterpretable predictor beyond the shared image of sealed internals.[1]

Unit

The mapping from input history to output, inferred without internal access

Recovery means

Not applicable

Measured by

Characterised by the input-output protocol (the sequence of stimuli applied and responses recorded) from which behaviour and state structure are deduced

In software engineering, black-box testing is a strategy that derives test cases only from a program's specification and external behaviour, deliberately ignoring its internal code and structure. Myers's The Art of Software Testing frames it against white-box (structural) testing and develops the input-partitioning and boundary techniques that make specification-driven testing effective. The referent is a concrete software-quality methodology (a way of choosing and judging tests), distinct from a cybernetician's method of inquiry or a statistician's opaque model, joined only by the idea of treating the thing under study as sealed.[2]

Unit

The set of test cases derived from the specification, scored by fault detection

Recovery means

Not applicable

Measured by

Assessed by specification coverage and the defects caught per test suite, independent of code-internal coverage

In statistics and machine learning a black-box model is a predictor (a random forest, a boosted ensemble, a neural network) that achieves accuracy through a form too complex to be read as a simple explanation of how inputs drive outputs. Breiman's Statistical Modeling: The Two Cultures contrasts this algorithmic, accuracy-first culture with the interpretable data-model culture, and defends black-box predictors where nature itself is complex. The referent is a class of opaque but predictive models and the interpretability problem they raise, unrelated to a test methodology or a cybernetic probe except through the common theme of hidden internals.[3]

Unit

Predictive accuracy of a model whose internal decision process is not interpretable

Recovery means

Not applicable

Measured by

Evaluated by out-of-sample predictive accuracy, with interpretability pursued separately through post-hoc explanation

In science and technology studies, black box refers to a fact or machine so settled that its inner workings and history are no longer questioned, only its inputs and outputs. This meaning comes from Latour (1986) in Science in action: how to follow scientists and engineers through society.[4]

Unit

Recovery means

Measured by