The fourth pillar was already there in the schema design. The _meta field says “hypothesis — designed to be deformed by the inference store.” That’s evolutionary epistemology. We just hadn’t named it. The word was already there. The pillar was unnamed.

Text below is LLM output with inferential guidance (except this).

What evolution means in this context.

Hypotheses enter sparse. The inference store applies selection pressure. What doesn’t hold up against real data gets displaced — revised out of the schema, replaced by structure that the actual inferences support. Schema versions, operator frameworks, author weights: all of it is subject to that pressure. No special protection for any source or framework. Weak structure loses.

This is not a metaphor. It is the literal mechanism the pipeline is designed around. The schema says explicitly that it is meant to be deformed. The inference store is the deforming force.

What this resolves about schema design.

There is a real tension in building a sparse social schema before the inference store is large enough to validate it. The RISC discipline — build sparse, resist premature complexity — can feel like excessive caution. Why not add Schütz, Goffman, Henrich, Bourdieu? The frameworks exist. The connections are plausible.

The evolutionary framing gives the cleaner answer. The discipline isn’t “only add what’s proven.” It’s “add sparse hypotheses and let the data select.” The consequence of adding a framework prematurely is not catastrophic — the inference store will displace it if it doesn’t fit. The consequence of adding it too thickly, as a foundation rather than a hypothesis, is that it becomes harder for the data to dislodge. The RISC principle is about keeping the schema shapeable, not about keeping it small.

Thin sourcing (Benedict at culture, Machin’s affective field values in the social container model) is intentional under this framing. These are hypotheses entering the coordinate space. If they’re wrong, they’ll be selected out. The schema carries a _status: "stub" marker on the parts that haven’t been validated — processing code knows not to treat them as equivalent to the validated dimensions.

The four pillars.

The project has been described as local-first, privacy-preserving, no external taxonomies, analogical orientation. Those are the original four non-negotiables. Evolution is different in kind — it’s not a constraint on the system’s behavior, it’s a description of the system’s epistemology. How the system learns. How it improves. How weak structure gets replaced without manual intervention.

The inference store is not just the output of the pipeline. It is the selection environment. What the pipeline produces feeds back as selection pressure on the schema the pipeline uses. That loop is evolution. It was designed in from the start — but it needed to be named to be used intentionally.

The practical implication.

When reviewing input from other models, the question changes. Instead of “is this sourcing solid enough to add?” it becomes “is this sparse enough to be shapeable?” A hypothesis that enters as a stub — clearly marked, with _inferences: [] ready to receive evidence — is not a liability. It’s a hypothesis in the correct state. The inference store will do the rest.

Part of the Dimensional Fields project — local-first, privacy-preserving, human-centric.