nishi code wiki / research / biological oracles

Biological plausibility as a build gate: the unoccupied niche

Published 2026-08-02 · lineage: forks off the perceptual-realism oracles brief, which established that image fidelity and motion plausibility are gateable while presence is not · sibling of the craft SOTA roadmap

The finding that matters. Every ingredient for an automated biological-plausibility oracle exists — open ground-truth datasets, published reference constants, permissive licences, scriptable tooling, and CI hooks in shipping engines. Nobody has assembled them. Three independent search lines (animal anatomy, tree generation, terrain) converged on the same conclusion: the generative-graphics literature evaluates realism with distribution distances, geometric agreement with an artist's rig, or human panels — never against quantitative biological reference values. A May 2026 survey states it outright: "the absence of an agreed-upon production-readiness evaluation suite makes it difficult to quantify systematically the production gap." Every open validator (glTF-Validator, glTF Asset Auditor, usdchecker, Omniverse) checks format and budget conformance. Not one asks whether the thing is a plausible organism. And the industry-standard tree tool explicitly declines to enforce botanical rules by design.

1. What we measured in our own world

This brief is written after wiring the first two of these checks into our own build gate, so the numbers below are ours, not aspirations:

Check	Reference band	Our measurement	Outcome
Stem density, medieval-highland identity	500–800 trees/ha at DBH ≥ 10 cm (>3,000 forest plots)	372 → 643 trees/ha	the first reading failed: our forest was barely half a real one. The generator was changed, not the band
Hypsometric integral (Pike–Wilson identity, one O(n) pass)	mature/equilibrium 300–600‰	594‰	we had declared the world "youthful" by intuition; the probe disagreed and the declaration was corrected

Both readings changed a belief. That is the argument for cheap biological oracles in one line: the forest we thought was dense was half-density, and the landscape we called young is mature. Neither was visible by eye.

2. The cheapest defensible checks, with their constants

Oracle	Reference value	Cost
Phyllotaxis divergence angle (apex primordia)	137.5°, robust to ~0.001°	O(n)
Branch-diameter exponent (da Vinci → Murray)	α ∈ [1.5, 3.0], measured 1.5–2.5 across four centuries of art that reads as realistic	O(n)
Tree slenderness / crown ratio	h/d < 80; live crown ratio ≥ 0.36	arithmetic
Branching angle, species-class aware	conifer ±9°, broadleaf ±22° (measured reconstruction-vs-field error)	O(n)
Mean leaf inclination	global 41.47° ± 9.55°, per plant functional type	O(n)
Quadruped limb allometry	chord length ∝ M^0.31, diameter ∝ M^0.35, across 32 species spanning 0.02–3500 kg	O(bones)
Gait duty factor / Froude transitions	walk >0.5, run <0.5; walk→trot at Fr 0.3–0.5, trot→gallop 2–3	arithmetic
Köppen climate classification	pure threshold rules; the published map agrees with 22,078 real weather stations at 80%	arithmetic
Terrain: Hack's law, concavity, fractal dimension	h ∈ [0.55, 0.62]; θ ∈ [0.40, 0.60]; D ∈ [2.00, 2.30]	O(n) to O(n log n)

Sources: Sci. Rep. 5:15358 (phyllotaxis) · arXiv:2402.13520 (branch-diameter exponent) · Biewener, J. Exp. Biol. 105:147 (limb allometry) · ESSD 17:1347, 2025 (leaf inclination) · Beck et al., Sci. Data 5:180214 (Köppen) · Tallo: 498,838 trees, CC BY 4.0

3. The calibration traps — where a naive oracle would be wrong

A perfect generator is less realistic than a slightly noisy one. Real sunflowers show exactly-Fibonacci parastichy counts only 74% of the time (82% including structured variants). Enforcing 100% would reject reality.

The golden angle does not survive to maturity. On mature beech, spruce and pine, measured branch divergence angles are statistically uniform, not clustered at 137.5° — bud mortality and branch shedding erase the pattern even when the underlying bud angle is constant. Enforce it at the apex; never on surviving branches.

A tolerance cannot be tighter than its reference. Joint-angle goniometry has a 1–6° (median 3°) intra-observer floor; body-mass-from-bone-circumference carries 25.6% mean prediction error; crop-model inter-comparison shows a 24–38% irreducible relative error. Demanding better generates false failures.

One published oracle has been falsified. The classic speed-from-stride-length trackway equation was tested against ground truth in 2025 and overestimated speed by 1.17× to 4.74×; the authors call it "inaccurate, if not outright misleading." Use Froude and duty-factor consistency instead of absolute speed inference.

4. Licensing is the hidden constraint

Most of the attractive datasets are non-commercial. The clean, commercially-usable stack is smaller than it first appears: Tallo (498,838 trees, CC BY 4.0), Köppen-Geiger maps (CC BY 4.0), Whittaker biome polygons (MIT), SoilGrids (CC BY 4.0), Copernicus DEM, Hydrography90m (CC BY 4.0), Smithsonian 3D (CC0), AVONET, and the open tooling (pyflwdir MIT, OpenSim and MuJoCo Apache-2.0, glTF-Validator Apache-2.0). Blocked or restricted for commercial use: the largest crop model, the leading yield-gap atlas, the bare-earth DEM most accurate under canopy, the dominant vertebrate CT repository's default terms, and the one existing perceptual tree-realism dataset.

UNVERIFIED / declared gaps

The hypsometric stage cutoffs are contested. One sweep sourced the 0.60 / 0.35 boundaries through secondary literature; a second could not verify them in the 1952 primary. We therefore treat a 1% miss at that boundary as non-binding and say so in the config — a build must not turn red on a threshold its own reference cannot pin down.
No comparative-anatomy oracle exists. There is no open, machine-readable table of per-taxon limb-segment proportions, joint-axis orientations, or muscle attachment sites. Building one from open CT meshes is tractable but is original work, not adoption.
No benchmark compares procedurally generated terrain's drainage against surveyed rivers. The only real-vs-real reference point is that the best global hydrography dataset overlaps a national survey by 46% at a 100 m buffer — so an oracle demanding 70% would demand something no published dataset achieves.
Not yet measured by us: everything in §2 except stem density and hypsometric integral. Those two are wired; the rest are adopted-but-unrun, and are recorded as such rather than claimed.
Animal and plant assets do not yet exist in our worlds — the fauna oracles above are therefore forward-looking, and honest reporting requires saying so rather than listing them as coverage.