Before the map, a vocabulary

Osmanthus is usually described by its output: a platform that maps how specific neurotransmitter receptor combinations produce distinct emotional and cognitive states. The less visible work sits underneath that sentence. Before a state can be mapped it has to be described, and before it can be described there has to be a vocabulary stable enough to describe it with.

The vocabulary problem

Subjective experience arrives as prose. The sixty thousand human reports of psychoactive drug effects that Osmanthus synthesizes were written in ordinary language, by people reaching for whatever words were nearest. Two reports can describe the same effect in entirely different terms, or use the same term for entirely different effects. A model trained on that surface would mostly learn the noise.

The instinct of an earlier era was to treat this as a reason the domain could not be modeled at all: subjective reports dismissed as anecdote, too soft to compute on. The reports were never the problem. The missing piece was a representation layer between the prose and the model, and building one is documentation work before it is machine learning work.

The correction predates the platform. Years of community documentation work went into formalizing the subjective effects of psychoactive substances into a granular, named taxonomy: an index of effects precise enough that a term reliably points at the same experience across thousands of reports. That index functions for us the way a coordinate convention functions for a cartographer. It makes no claims about the territory; it makes the observations comparable.

From reports to coordinates

With a controlled vocabulary in place, the harmonization problem becomes tractable. Language models resolve the surface variation of raw reports into structured observations, and each observation can then be set against biochemical data: the receptor-level profile of the compound that produced it. Repeated across the corpus, the pairing is what turns psychoactive pharmacology into a searchable design space for subjective effects rather than a pile of anecdotes.

None of this flattens experience into numbers. A report keeps its texture; what changes is that its claims about perception, mood, and cognition now land in named, comparable dimensions. The same discipline that lets two field biologists trust each other's notebooks lets two decades of psychonaut documentation function as scientific instrument reading.

This is the representation layer, one of the three components state design requires, alongside biological control variables and controlled human validation. It is deliberately the least glamorous of the three. The map is not the experience. It is the thing that lets the next experiment be specified instead of stumbled into.