The Reciprocal Corpus Program

— A method for learning from the enormous body of community-made 3D character work — the way a model learns from a corpus — without taking anyone's assets, and giving the resulting analysis back to the people who made the work in the first place.

The idea, and why it's fair

Communities around character-creation tools — from open avatar makers to games whose studios have since closed — have produced millions of characters. Each one is a point in that tool's parameter space: a vector of choices about proportion, feature, and form. Taken together, that's a map of what people actually build and find appealing.

We want to learn the shape of that map to steer our own procedural generator — but we do not want, and do not take, anyone's meshes, textures, or files. The distinction is the whole program:

That last point is the reciprocity. An open analysis of a community's collective output is a gift back to that community. It's the CC-attribution spirit made mutual: we learn from the aggregate, and the aggregate learns from itself through our lens.

The pipeline — one substrate, many adapters

SOURCE → adapter (parse; extract FACTS; retain ZERO asset bytes) → normalized parameter / statistic rows (provenance + license tag) → distribution study (taxonomy + fits + coverage gaps) → { generator target distributions } the private win + { open research brief + data points } the public give-back

Every source, however different, reduces to the same shape: a parser that turns it into rows of measured facts, feeding one shared study stage. A rigged model, a parametric character card, and an academic body dataset all land in the same normalized frame.

The phases

Phase 0 — the substrate foundation

A single measured-facts store (one row per source-and-fact, with provenance and license tags), and the extraction contract with a gate that asserts zero asset bytes survive. The honest guarantee is mechanical, not a claim.

Phase 1 — source adapters in progress

One parser per corpus shape: rigged glTF/VRM models (mesh, skeleton, material, UV census); parametric character cards (the parameter block is the data — no mesh involved); and open academic body datasets folded in as ground-truth anchors. Each adapter emits only numbers.

Phase 2 — distribution studies the data product

Per source: a taxonomy of parameters; distributions over the community's vectors (what recurs, what clusters, where coverage is dense or sparse); and a normalization that maps every tool's space onto one common frame so they can be compared at all. The output is a coverage report with a number on each gap.

Phase 3 — generator wiring the private win

The measured distributions become target distributions for our own generator, which must first cover the community's measured span, then exceed it on axes a fixed asset library can't follow — per-character regeneration, physically invariant motion, no authored files.

Phase 4 — the give-back the point

Each study is published as an open brief with attribution to the source community, and its data points — the taxonomies and distributions — released as open data. This page is the first of them: the methodology itself, given back before any specific study.

Guardrails — mechanical, not promised

UNVERIFIED — what this page does not yet establish

Sources & foundations

Lineage — Forks off the surface-floor and rig-floor measurements, which established that a reference corpus can be measured into a capability target without taking its assets. Parent: character asset floor, 2026-08-03.