Heuristics developed in discussion amongst Dan, Woodley and Emily when comparing analyses on 12/10/14:

1. Make the biggest possible island with good dependencies. (MOAR DEPENDENCIES) (This conflicts with the initial impulse to make the syntactic tree as close to the right shape as possible.)
2. Pick the right lexical entries when available. (Using x_-_bridge_le is tantamount to asserting that none of the options provided by the grammar is appropriate in its semantics.)
3. Don't tell lies. (Don't include semantic dependencies that are known to be incorrect.)
4. Don't use fragments in bridging. (The grammar should not make these available as options, but if they are, avoid them.)
5. Given two candidate analyses with the same number islands, prefer the one where the additional dependencies captured are rooted in (i.e. semantically headed by) open class words.