Now we start being in a context of science, supposing we have to assess something in absence of the blueprint. But that is complementary to engineering, where we decide the blueprint - making it much easier to tell and plan "is" aut "seems".
And in a context of science, you cannot prove me whether your cousin is intelligent, until we open her and check the wiring - but I can be convinced that she is not intelligent, if her behaviour so shows.
Plenty of clear hints come from machines and, more unfortunately, humans - I listed a few recently elsewhere. This that does not know that you cannot ring the doorbell of Caius Julius Caesar though it read a biography; that which would indicate the swan right there as white in spite of being told differently; that which goes full delirious.
An intelligent entity has to be able to know entities and assess their relations with foundations. It is not difficult to reveal when this is not happening, and the literature contains a number of "codified" examples: one is that of understanding the pronoun in the sentences "The trophy will not enter the case: it is too big" and "The trophy will not enter the case: it is too small", which can betray the absence in the engine of a capability for (sub)world modelling.
Prof. Patrick Winston said that intelligence is "that you have never run with a bucket full of gravel but you know what it entails regardless", because you can reliably "tell yourself a story". That is a requirement: the "intelligent" entity has to know what "bucket", "gravel", "running", and all possibly related concepts, and be able to refine them sensibly combining them, and produce statements that are solid at the state of development of its ontology.
So, before going into finer details, let us have a system that definitely builds an ontology, made of virtual scenarios in which entities are known progressively, and let us see how it reasons about it. Because when this is missed, then Intelligence is missed in spite of any mockery of actual intelligence. A photocopier do not make one an artist, and capabilities for imitation are a risk, not a goal: they define a warning that makes the "real thing" trickier to discriminate. And since we already know that by design the "real thing" is not there, failure under some field condition is already foreseeable. If you cannot tell "looking like" from "being", then you have a problem - because what "is not" will probably reveal so.
You have to build the system so that the core is foreseen.
And if instead the "foundation" of the outputs comes from crunching the most likely fitting response, that is directly the opposite of intelligence - which understands /why/ something works, and does not just record what "works". It is the opposite because intelligence by definition "reads in", investigates, while the other fully "delegates out".
Did not you mention Science? Well, to be intelligent an entity has to be a Philosopher and a Scientist. And if it expresses like one, if it is a black box your duty is to check thoroughly with renewed effort that you are not tricked, that there is substance behind it - far from being contented.
--
Edit: you want a "moving goalpost"-like test? The Resnik-Halliday-Krane Physics textbook contains a number of exercises: the AI has to be able to solve them correctly and properly - "properly" more than "correctly". And then, for all disciplines, there will be similar exercises - from "Why would the radical school of Austrian Economics fail in the Kholomity region, and why would it succeed", to "Why would implementing radical school of Austrian Economics measures in the Kholomity region be immoral, and why would it be moral"; "legal/illegal" etc. Make it pass examinations for degrees. But it remains key that no tricks are employed: it has to build instances of subworlds, made of organically developed concepts, including laws, and make its tests and reasoning on them, and draw conclusions. Tricks are always possible, and "copying from the neighbour or learning answers etc" cannot count.