nx_charjudge_census.nx

nx_charjudge_census.nx -- THE CHARACTER REFEREE AS A STANDING INSTRUMENT: judge an ENTIRE named corpus in one call, one line per image, with the partition reconciled -- AND (2026-08-19, GR1 of /compare/graphics, Phase 0 of the past-MetaHuman roadmap) THE JUDGE PANEL `cjc_panel`: a multi-lens referee that can see what the single statistics judge is proven blind to. WHY IT EXISTS: the judge was a single-image CLI, so grading a corpus meant one invocation per file. That cost is exactly what turns "measure the population" into "measure a sample", and a tooling gap is never a licence to sample. The whole render corpus is now one call. WHY THE PANEL EXISTS (measured 2026-08-16, both directions): nx_charjudge scored a GREEN CLAY figure crop 257 (the operator saw gumby) and a smooth photoreal-class splat render 0 ("smooth = no detail"). Its axes (composition, palette, contour, face) are hue- and semantics-agnostic BY CONSTRUCTION, so every fidelity claim graded through it alone is ungradeable. The panel COMPOSES incumbents, never re-implements: stats = ./nx_charjudge.elf (the statistics judge, forked -- the artifact that ships) skin = nx_color_kovac_skin_pixel (Kovac 2003 published skin gamut, over the figure / centre) sym = ./nx_image_beauty.elf path (bilateral symmetry Q1000, forked) percept = ./nx_percept.elf judge (RELATIVE perceptual position vs the 10 human-confirmed commercial frames minus the junk controls; admitted only while its own ladder reproduces the human ordering) vlm = UNOBSERVABLE (no sovereign VLM oracle exists on the estate -- measured 2026-08-19 by tree grep + local probe; the slot abstains, it never acquits) THE LAW THE PANEL IS BUILT UNDER: AN UNCALIBRATED CLASSIFIER REPORTS NUMBERS, NEVER VERDICTS. `panel` prints every lens value for every image always; it prints a GOOD/BAD verdict ONLY while an admission receipt written by `calibrate` says FP=0 and FN=0 on a LABELLED corpus with both classes present -- and the receipt is bound to the exact floors it was measured at, so editing the conf silently un-admits. Floors are DATA (knowledge/cjc_panel.conf), never literals here. A lens with no floor in the conf is REPORT-ONLY (printed, not voting): that is how a lens proven blind is retired from the vote by evidence. Composite = AND over voting lenses (the MIN-headline doctrine: no lens may be compensated by another). WHY IT FORKS rather than looping in-process: chj_judge allocates several w*h*8 working buffers per frame, so an in-process sweep multiplies every one of them by the corpus size -- measured, that is gigabytes of address space on a box already sitting at 727 permil swap. A child returns its entire working set to the OS on exit, so peak RSS is ONE frame regardless of corpus size. Forking also means this census measures the artifact that ACTUALLY SHIPS rather than a second compiled copy of the same lib: if the promoted binary is stale or missing, that surfaces here instead of being hidden by an identical rebuild. The panel keeps that design: the parent forks ITSELF (`lens <path> <conf>`) per image; the child decodes once (nx_img_to_rgb: PNG/JPEG/GIF/WebP-lossless/BMP/...), computes the in-process lenses on a bounded fit-resampled frame, forks the three sibling judges, and prints ONE anchored line.