I am looking for references on conjectures or heuristics concerning cancellations of the Möbius function in very short intervals, namely how small is it believed one can take $H$ so that
$$\sum_{X \le n \le X+H} \mu(n) = o(H)$$
holds for every $X$ (rather than almost every $X$), especially if there are conjectures that go beyond $H = O(X^{1/2+\epsilon})$ from the Riemann Hypothesis.
In the case of short sums of the Von Mangoldt function (that is, counting primes in small intervals) the expected heuristic is known to fail for $H = (\log X)^A$ and conjectured to be true for $H = X^{\epsilon}$. I am looking for something similar for the mobius function.
Thanks in advance