Every two weeks, the difficulty adjustment algorithm compares the number of blocks over the previous two weeks with the expected number of blocks, given a target block interval of ten minutes, and adjusts the difficulty up if there's too many blocks or down if there are too few blocks.

My question is, what clock/calendar does the algorithm depend on to determine block times?

Would the developers trust time stamps in the block headers, or internet time servers? AFAIK, the only reliable information in a block about the block time is the nonce count, but it's not possible to calculate the block time without also knowing the miner's hash rate (block_time = nonce_count / hash_rate).

In any case, the DAA assumes that the next 2016 blocks will be the same as the last 2016 when they never are and that's why actual block times are all over the place, anywhere between 0.3 seconds to 3 hours.