I love taking metaphors until they break 
I was definitely thinking of it in the most common definition of the word, which is as an assembler. I.e. taking a high-level desired behavior and ‘compiling’ that into more granular, lower-level incentives that aim to generate, in aggregate, the desired low-level behavior. For the core algo, which actually generates scores that are usually meaningful enough to be actionable (though typically only within a single repo/Discourse instance, as activity is different in nature across repos), it’s a bit fuzzy. Because the semantics are fuzzy. We’re looking at what’s been valued in the past and saying, “More of that please…”. It remains to be seen if we’ll actually get more of that in the wild, but let’s say for the sake of argument that it works well enough. This seems more like AI almost, which learns a model by being fed “successful” data. It “compiles” code much the same way it can generate target images using other images.
I suppose one can also say that new Issues, PRs, etc. are a more specific set of semantics guiding behavior. In this case the language is English (or other spoken language). But also, in my experience in OSS, the language of doing. It is concrete actions, guided by Issues, etc., or self created but still valuable contributinos, that generates 
s, PR reviews, and other actions that confer meaningful cred, driving the incentives.
When Initiatives come online, they will act more like high-level programming languages. Giving more specific semantics (desired behaviors).
These seem kinda the same, as both compile high-level languages to low-level languages. One just does it ahead of time in a more optimal manner, and the other does it on the fly. Maybe Initiatives are more like compiled languages, and the base SC algorithm is like a run-time compiler, the other “developers” entering “commands” in the form of Issues, comments, likes, etc.