Given that the di-y-lid is either TD2 or TD3, there only appears to be one mechanism consistent with the experimental and theoretical data. Oi atoms diffuse together normally, to form dimers. These are faster diffusing, with a barrier to diffusion of 1.7 eV. They migrate, and normally encounter a third Oi. In this case they may form a trimer, but if they do so it is only briefly stable. The trimer itself is also capable of diffusion, and so can move to another Oi. At this stage the 4Oi atoms experience sufficient lattice strain to switch them into the di-y-lid structure, and are responsible for TD3. This is not mobile, and so can only grow through further dimer addition (Oi migration is assumed to be too slow for growth purposes). Hence TD formation slows down at this stage, and TD3, the di-y-lid, becomes the primary thermal donor.
In parallel with this process, the trimer is able to restructure into either TD1 or TD2 with a barrier of 1.2 eV. These two defects are also able to restructure into their electrically inactive shared isomer, which is different from the standard trimer structure, and probably not stable at room temperature. The low stabilities of TD1/2 are Fermi level dependant, so as the Fermi level rises with the addition of more thermal donors, first the TD1 species tend to transform to TD2, and eventually these are also unstable and transform back to trimers to rejoin the main thermal donor formation path.
We now need electrically active trimer structures for TD1 and TD2. The Snyder-Stavola 3O model is a candidate, assuming there are a few modes which are weaker and not yet observed by experiment (this was the case with the oxygen dimer, where theory predicted modes which were not initially observed by experiment).
The results of initial dispersion treatments and isothermal anneals  suggest that TD1 forms in the initial anneal stages along with the trimers, but rapidly converts to either trimers or TD2. TD2 is stable for a longer time. This suggests that there is only a small transformation barrier between the trimer and TD1, with a larger barrier to TD2; however once formed, TD2 is more stable. There are presumably several schemes that could fit this information. One possibility is if the trimer is stable in the `Manx' form, as our preliminary results suggest. This can then become a +2 TD1 structure by pushing one of the oxygens into a y-lid; in this case TD1 is a y-lid with Oi atoms in both back bonds of the core Si. This requires little structural rearrangement. Later the Oi atoms are able to migrate around to lie in the same plane as the y-lid, forming the Snyder-Stavola structure as TD2.
Watkins has shown that the reorientation time for TD1 is roughly half that of TD2 . This is consistent with this model, since the TD1 model proposed here can reorient by the y-lid O atom dropping back into a bond centre, and one of the other bond centred Oi atoms moving into a y-lid configuration. However, for the Snyder-Stavola TD2 structure to reorient requires several hops for the Oi atoms at the ends of the defect, and hence would be slower. The TD1 model here would only be a double donor if both the trivalent oxygen and the core trivalent Si atom both lost an electron. There is no obvious electrostatic compression mechanism exerting on the empty p-type orbital of the core Si, so it is not obvious that such a +2 donor state would be shallow. Experimentally however TD1 has a much deeper gap state than TD2. As TD1 and TD2 are isomeric, it is notable that during the reorientation experiments, transformation into the thermodynamically more stable phase is not observed, i.e. the TD1 species do not all transform into TD2. Presumably this must be because the reorientation does not bring the atoms into a position where they can hop directly into TD2 sites; the model described here would fulfill that criterion.