A Universe of Sorts

First idea: take random assignments, and use this to estimate total number of models.
Next idea: take deterministic partial assignments, then extend these partial assignments randomly. This will use the randomness on a 'smaller domain', and exploits linearity of expectation.
Next refinement: A partial assignment is just a predicate $A$ that we adjoin to our formula. Therefore, can we randomly pick such an $A$ , such that if we can estimate $|M \models \phi \land A|$ , we can estimate $|M \models \phi|$ ? Yes, pick symmetric $A$ that cut down the number of models by a factor of 2 in expectation. For example, pick $A$ to be something like $x_i = x_j$ .
More generally, can pick $x_i = x_j$ with probability 1/2, and $x_i = \neg x_j$ with probability 1/2, and this will give us the same guarantee.
This can be written as $x_i \oplus x_j = b$ , where $b$ is a random bit.
So, pick $\{ A_k \}$ of the form $x_i \oplus x_j = b$ , and then estimate $|M \models \phi \land_k A_k|$ for the full set $A_k$ by random sampling. This gives us an estimate of $|M \models \phi|$ as $2^k * |M \models \phi \land A_k|$ !
Why does this work so well in practice? I don't know! I should read Kuldeep Meel's paper to find out.

§ Randomized SharpSAT