First-Divergence Consequence Analysis for Masked Visual Generators

Munsik Kim · ORCID 0009-0008-9350-2435

Technical preprint; not peer reviewed. First public release: v1.0.0.

Paper PDF · Zenodo record · Korean README

FDCA couples a reference masked decoder and an approximate decoder, isolates the first approximation-induced split, and decomposes endpoint risk into exact split incidence and split-conditional consequence. The visual study uses an audited branch-independent Halton schedule, coordinatewise conditional maximal coupling, event-addressed random-number generation, realistic W8 round-to-nearest quantization, and a common approximate suffix after the natural seed.

Results at a glance

Checkpoint	Contrast	rho 0.5	rho 0.3	Ratio	Paired 95% CI for difference
H-MaskGIT-T	Conditional LPIPS	0.03013	0.01013	2.98	[0.01719, 0.02287]
H-MaskGIT-T	Incidence-weighted LPIPS	0.00393	0.00221	1.77	[0.00137, 0.00207]
H-MaskGIT-S	Conditional LPIPS	0.03076	0.01038	2.96	[0.01714, 0.02381]
H-MaskGIT-S	Incidence-weighted LPIPS	0.00411	0.00214	1.92	[0.00150, 0.00248]

Headline values are operational LPIPS/DINO outcomes for two checkpoints in one model family, not human perceptual ground truth or architecture-universal guarantees. See compact evidence.

Quickstart (model-free)

python -m venv .venv && . .venv/bin/activate
pip install -r environment/requirements-model-free.txt
make verify test reproduce

Repository map

• fdca/: theory, coupling, stable-set/first-divergence, RNG, adapters, perturbation and analysis
• experiments/: T0 through P3S purpose, preregistration, gate and negative findings
• public_results/: compact row/block evidence, tables and figures
• paper/: v1.0.0 PDF and LaTeX/BibTeX source
• docs/: architecture, stack, scope, experiment and claim ledgers
• scripts/ and tests/: public validation and model-free reproduction

Full GPU boundary

Tier-B regeneration requires official checkpoints, VQ and metric assets listed by exact revision/size/SHA-256 in the external asset registry. They are not redistributed. Full trajectories, datasets, Conda environments, external clones and private logs are absent.

Negative and unsupported claims

P05R's valid influence envelope was vacuous; P2A remained incomplete; the P1M profile largely reproduced rho ordering; W4 was stress-only. No claim is made about human judgment, production controllers, persistent FP32-vs-W8 deployment, other schedules, coupling independence, or models outside this two-checkpoint H-MaskGIT family.

Citation

@article{kim2026fdca, title={First-Divergence Consequence Analysis: From Quantization Onset to Perceptual Impact in Masked Visual Generators}, author={Kim, Munsik}, year={2026}, doi={10.5281/zenodo.21965747}}

Licenses

Original code is MIT. Paper, figures, tables, release documentation and included audits are CC BY 4.0. See license scope and third-party notices.