nishi code wiki / research / measurement constitution
The measurement constitution: metrology across every domain, down to pore scale
Published · lineage: forks off ruler-craft (the classical projective ladder) and generalises it — operator: the house as real as the person as real as the tree, and micro-detail down to a mole, a Montgomery gland pattern, stubble.
Freshness declared. Verified to the author's May-2026 horizon plus the corpus; the May→July 2026 tail is UNVERIFIED. Every numeric band below is a commonly-cited range pending the citation-archive pass, not a pinned figure.
The reframe this brief exists to deliver: our worlds look fake because everything in them is PERFECT. Real walls are not plumb, real planks are not flat, real skin is not smooth, real areolae are not evenly stippled. Realism is measured IMPERFECTION — and metrology is precisely the science of specifying imperfection with numbers. That makes the discipline's whole outline directly load-bearing for us: the same framework governs a face, a house and a tree; only the tolerance tables change.
1. The operator's paper: three shape methods, and why the answer is a BATTERY
The Class-III mandible literature exists to ask exactly our question — do cephalometric, Procrustes/GPA and Euclidean-distance (EDMA) analyses of the same specimens agree? The durable finding of that genre: they agree on gross verdicts and disagree on localisation, because each carries a different artifact.
| method | strength | its characteristic artifact | our use |
|---|---|---|---|
| Cephalometric (named angles/ratios between named landmarks) | clinically interpretable and transferable — a surgeon speaks it | assumes the chosen measurement set is the right one; blind to anything not enumerated | the medical INTERFACE. Nasolabial angle, Goode ratio, nasofrontal angle, gonial angle: this is the layer a cosmetic-surgery client and clinician can read. Our rhinoplasty targets are already cephalometric-class. |
| Procrustes / GPA (optimal superimposition, PCA of shape) | whole-shape, anchor-free, gives principal axes of variation | the “Pinocchio effect”: one localized deviation smears residual across all landmarks | the honest global verdict. Ran here 2026-08-03 and immediately overturned an anchored result — see §2. |
| EDMA (full inter-landmark distance matrix, coordinate-free) | no superimposition choice at all; localises to specific distances | loses spatial intuition; many-comparison inflation | the localiser. 17 landmarks → 136 distances: 13× the redundancy of a hand-picked measure set. |
| TPS / deformation grids | decomposes the difference by SPATIAL SCALE (affine vs principal warps) | sensitive to landmark noise at small scales | emits the cut list in carve order — the marble doctrine as mathematics. |
Adopted rule: run all four; require agreement; treat disagreement as a finding. Robustness is not a better single method — it is methodological triangulation, exactly as the paper's genre demonstrates.
2. What the battery already caught (our own instrument, tonight)
Our pointing machine anchored on the two eye corners (IPD) and reported the jaw as not significant. Procrustes superimposition — anchor-free — ranked jaw 16.3 mm, contour 13.8 mm as the LARGEST residuals in the face. ⇒ an anchored measurement hides error at its anchor: pinning the eyes forced the residual outward and systematically under-read the jaw, the very region the operator's eye kept calling wrong. The operator's judgement and the anchor-free mathematics agreed; the anchored instrument was the outlier.
3. The metrology outline, mapped onto our domains
| discipline | what it actually gives us | domain reach |
|---|---|---|
| VIM — the vocabulary | Say precisely what you mean: accuracy (closeness to truth) = trueness (bias) + precision (scatter); error ≠ uncertainty; resolution ≠ discrimination threshold. Our 8.9 mm was reported with a spread of 0.0 from ONE sample — perfect precision, unknown trueness. The vocabulary alone would have flagged it. | every ruler |
| GUM — uncertainty budgets | Every number carries a budget of named contributors combined in quadrature, reported as expanded uncertainty (k=2). Ours: detector bias, IPD prior ±3%, residual pose, pixel quantisation, model-fit residual. This is the single biggest rigor upgrade available to the pointing machine. | every ruler |
| Traceability | Every measurement chained to a reference standard. Ours is currently a statistical prior (mean IPD), not a traceable metre — a known-size object in frame promotes a scene from inferred to traceable scale. State which regime each measurement is in. | person, corpus |
| GD&T — geometric dimensioning & tolerancing | The answer to “why does the world look fake”. GD&T specifies form (flatness, straightness), orientation (perpendicularity, plumb), location (position), and profile — with TOLERANCES. Real construction has published tolerance bands (walls out of plumb by millimetres per metre, floors out of level, timber bowed and crooked within grading limits). Our voxel world holds every surface to zero deviation, which is a value no real building has ever achieved. | house, furniture, tools, roads |
| Surface metrology (ISO 25178 areal) | One parameter family — Sa, Sq, Sz, Sdr, Str — describes ANY surface: skin, bark, plaster, sand, worn stone. Adopting it means one texture ruler across every material instead of per-material taste. | skin, bark, plaster, sand, stone, metal |
| Dimensional metrology | The calibration hierarchy and the discipline of gauge R&R: how much of observed variation is the PART vs the GAUGE. Ours is unmeasured — a repeatability study of the landmarker on identical renders is a one-hour job that would bound it. | every ruler |
| Forensic metrology | Measurement that must survive adversarial scrutiny: validated method, documented uncertainty, chain of custody. This is the standard our “provably the best” scoreboard must meet — not a number, but a defensible number. | the scoreboard itself |
| Time metrology | Beyond percentiles: Allan deviation characterises stability ACROSS timescales — it distinguishes white jitter from slow drift, which a p99 cannot. Directly applicable to frame pacing and to the sim tick. | frame smoothness, determinism |
| Metrication / unit discipline | One symbol, one scale, always. This exact failure bit us: two different q8 conventions shared one name (world q8 = 0.39 cm vs model q8 = 0.061 mm, 64× apart) and a published claim was wrong until the arithmetic was redone at implementation. | engine-wide |
| Historical metrology | The anthropometric canons and the mason's transfer methods — the ancestry of the whole programme, and the source of the proportion priors we fit against. | person, architecture |
4. Micro-detail: moles, Montgomery glands, stubble, pores
The key reframing: micro-detail realism is a STATISTICS problem, not an art problem. You do not place a mole; you match a spatial point process plus a size/colour distribution, and let the operator veto individuals. That makes “a real mole here” gateable: the count, the size histogram, the spatial pair-correlation and the colour must all sit inside published human ranges.
| feature | the ruler | the gate (what must match) |
|---|---|---|
| Nevi / moles | Dermoscopy convention (10×, polarised); lesion diameter (the clinical 6 mm threshold is the familiar landmark); public lesion archives (ISIC, HAM10000) as appearance oracles; 3D total-body photography (Canfield VECTRA WB360-class) maps lesions onto a body model — the licensed-twin capture path for real placement | count per body region inside published adult ranges; size histogram; NON-uniform spatial distribution (sun-exposed vs not); colour inside measured nevus gamut |
| Montgomery tubercles | Counted, sized structures (commonly cited ~4–28 per areola, 1–3 mm); areolar diameter distributions | a hardcore point process on the areolar annulus — minimum spacing, no clumping; count + diameter inside published bands. A perfectly even ring is as wrong as a random scatter. |
| Stubble / body hair | Phototrichogram / TrichoScan conventions: density (hairs/cm²), shaft diameter (terminal ~50–100 µm vs vellus <30 µm), growth rate (scalp ~0.3–0.4 mm/day), anagen fraction | density and diameter per region inside published bands; stubble = sub-mm shaft protrusion — a normal/displacement + anisotropic shading problem, NOT geometry |
| Pores & microrelief | ISO 25178 areal parameters on skin replicas or fringe-projection systems (PRIMOS-class, micrometre vertical resolution; Antera/VISIA-class clinical imaging give pore COUNTS and scores) | Sa/Sq inside measured facial bands; pore diameter distribution (tens to a few hundred µm, largest on the nose); furrow network anisotropy (Str) — skin is directionally textured, not isotropic noise |
| Skin tone & colour | CIE L\*a\*b\* + ΔE00; ITA° (Individual Typology Angle) — a single objective number classifying tone from L\* and b\*; melanin/erythema indices; cross-polarised capture separates surface specular from subsurface colour | ΔE00 < 2 per region vs reference; ITA° inside the target class. This retires “warm tan” verdicts permanently — tone becomes a number with a class boundary. |
5. Adoption order (free precision first)
| # | act | why it is next |
|---|---|---|
| N1 | Uncertainty budget (GUM) + gauge repeatability study on the pointing machine | every existing number becomes defensible; costs one repeat-measurement run |
| N2 | Procrustes as the default verdict, EDMA as localiser, cephalometric set as the medical interface | already proven to overturn an anchored result — the jaw finding |
| N3 | ITA° + ΔE00 per region replacing eyeball colour verdicts | pure math on pixels we already render; kills a retracted verdict class |
| N4 | ISO 25178 Sa/Str as the one texture ruler — skin first, then bark, plaster, sand | one standard, every material; makes “too smooth” a number |
| N5 | GD&T tolerance tables for the built world — plumb, level, flatness, straightness with published construction bands | the house/tree realism ask: specify imperfection instead of shipping perfection |
| N6 | Point-process gates for nevi / tubercles / follicles | micro-detail becomes generative + gateable rather than hand-placed |
| N7 | Allan deviation on frame times beside the percentile tails | separates jitter from drift, which percentiles cannot |
UNVERIFIED / declared gaps
- May→July 2026 tail unverified (search budget exhausted). Most volatile: neural skin-appearance models, dermatology dataset releases, clinical 3D-imaging product capabilities.
- All numeric bands here are commonly-cited ranges, not pinned figures — the citation-archive pass must attach a source and measurement context to each before any becomes a gate threshold.
- Construction tolerance tables are jurisdiction- and trade-specific; the GD&T-for-worlds rung needs a sourced table per material class before it can gate anything.
- Gauge repeatability of our landmark detector is UNMEASURED — until N1 runs, every millimetre we report has an unknown share of instrument variation. Stated plainly rather than hidden.