Whatever the clarifications will be, I suggest:

• either removing "for real" words in the task description, highlighting link to "k-means" algorithm and explicitly stating in the problem that k-means clusterization should be used here;
• or remembering Kotelnikov theorem and removing test cases which violate it - i.e., where sampling rate is lower than twice the transmission rate; the words Clearly 10 samples per second is enough resolution for this speed should be removed as well with the example around them, as not only that is not clear, that is plain wrong.

There is a test where "1001" (full string) should be decoded as "EE". Why not "I"?

00 could be treated as either 1 unit in length or 3 units. Why round up, instead of down, as in dot/dash ambiguity?

Ah, sorry.

• Add JSON to the list of frozen modules. With the redefining of JSON.load credentials object becomes totally under control, and then it's easy way to all tests pass.
• The term "Metaprogramming" should be defined more clearly in the description.
• I didn't quite understand the tests. Do they check that any admin password is good, not just 'h4xx0r3d'?

I suggest to explicitly state that if a train is at the station at the start — it shouldn't wait.

It should be mentioned explicitly in the problem description that all calculations are to be made with floats, not integers.

It is reasonable to assume by default that "5 / 3 = 1", especially given that Ruby works in that manner.

The following test case should be added (or similar):

fn f1 a1 a2 => a1 * a2
fn f2 a1 a2 a3 => a1 * a2 * a3
f2 f2 1 2 3 f1 4 5 f1 6 7
# => 5040

This will check that function composition is parsed correctly, i.e. above expression is parsed as f2 (f2 1 2 3) (f1 4 5) (f1 6 7).

Without that explicit test wrong programs will fail at random tests, leading to tedious and exhausting debugging.