The method not very interesting. The result is the 20,568 product/inverse candidates of size (18,3) satisfying the basic constraints.
We used a Perl script called attack, which collects a set of countermodels and filters candidates w.r.t. those countermodels. Attack takes two filename as arguments, reads a stream of candidates from the standard input, and writes filtered candidates to the standard ouput. For example,
    attack head interps < candidates > candidates.filtered
The file "head" is a Mace4 input file, and "interps" (which must not
exist when attack is called) is the set of structures that is 
constructed during execution of the command.
We have 20,568 candidates in cand0, and we start attacking.
    attack head.2-3  interps.2-3   < cand0 > cand1  
 
    attack head.r9   interps.r9    < cand1 > cand2 
 
    attack head.r29  interps.r29   < cand2 > cand3 
 
    attack head.da10 interps.da10  < cand3 > cand5 
Next we ran a simple Mace4 job, without a candidate, to get all nonassociative
diassociative inverse loops (is that what they're called??) of size 12.
We get 16 of them; we could run isofilter to get rid of the isomorphic ones
(leaving 2), but that takes a half an hour and is not necessary here.
    mace4 -n12 -m -1 < da.in | get_interps > interps.da12 
 
    modfilter interps.da12 false_in_all < cand5 > cand6 
 
    attack head.r41 interps.r41    < cand6 > cand7 
Then we ran some Otter jobs on the candidates.  One of them seemed close
to being a single axiom, deriving all but associativity, so we ran some Mace4 jobs, including
a lot of group-like properties that were derived.  A size-16 countermodel was found, which eliminated
that candidate and 5 others.
    mace4 < da16.in | get_interps > interps.da16 
 
    modfilter interps.da16 false_in_all < cand7 > cand8 
Thirty-six candidates remain.
(Some of those can be derived from others, and
others can be completed, showing they are not single axioms.)
Six more candidates were eliminated, four of the remaining candidates were shown to be equivalent, and several insights arose.
From the 36 candidates, if we remove the four that complete*, the six with finite counterexamples, and three that are redundant because they are equivalent to another, we are left with 23 candidates.