Here are the slides for the talk I presented last week at the ACS meeting in Washington. It describes my understanding of the Daylight toolkit as deduced by John:
Over at the NextMove blog, Peter Shenkin brought up the biphenylene case, which (to my mind) illustrates alternative approaches to reading aromatic SMILES. Consider the SMILES string c12ccccc1c3ccccc23. Some toolkits may read this, work out that only the two six-membered rings can be aromatic, and then make sure that the double bonds are not placed in the four-membered ring. I refer to this approach as dearomatisation, an approach that Open Babel used to use. It involves ring detection, 4n+2 counting and so forth. Apart from taking some time, an obvious problem is different aromaticity models may be used by the reader and writer, thus leading the reader to drop aromaticity from a particular ring, typically by setting those bonds to single bonds and adjusting hydrogens, resulting in a different structure than intended.
In any case, this is not the approach used by the Daylight toolkit, which did not consider 4n+2, or even detect cycles. The approach is described in the talk above so I won't repeat it here. For the SMILES above, I believe that it would generate one of two Kekulé forms depending on the atom order; one with the two double bonds in the four-membered ring, and one with two benzenes. It's for this reason that Daylight would never generate that SMILES for biphenylene ("don't generate aromatic SMILES that you can't kekulize"), but always write a single bond symbol for the bonds connecting the phenyl rings (e.g. something like c12-c3c(-c2cccc1)cccc3). When written that way, kekulization always gives the desired form.
Following up on a comment by Rajarshi, while differences in aromaticity models are a problem for 'dearomatisation' algorithms, they are not a problem for the kekulization algorithm used by Daylight. So long as the structure is kekulizable (and appropriate single-bond symbols are used) then it can read in any structure without loss of information no matter what aromatic model is used.