Research release 0.1 · 7 August 2026

Fractal-addressed storage became a measurable codec experiment.

FIC/5 keeps one recursive mathematical image geography, then lets each addressable 64×64 region choose the best reversible representation. The frozen balanced candidate passed its declared public targets and then a separately opened H2 holdout while reconstructing every tested crop exactly.

Jonathan H. Lindley + ChatGPT

Read the paperSee the address modelReproduce itOpen live codec lab

Frozen public photographic median

Total package0.658×34.2% smaller
Crop payload0.967×3.3% fewer bytes
Cold crop0.975×2.5% fewer bytes
Decode time0.705×29.5% faster in recorded Vercel run

Previously sealed H2 holdout

Total package0.512×48.8% smaller
Crop payload0.830×17.0% fewer bytes
Cold crop0.848×15.2% fewer bytes
Decode time0.419×58.1% faster in that run

What is new here

Recursive logical geography

Every image region has a stable Base-4 ancestry. The fractal supplies locality and addressability, not magical compression.

Local reversible encoding

Each block tries nine reversible pixel transforms before compression and stores a decode-aware winner.

Exact access + fallback

Requested crops touch only relevant blocks, outputs are hash-verified, and a conventional tile pyramid remains legal when FIC loses.

Claim boundary. This result is against the exact-deduplicated complete 64/128/256 tile-pyramid baselines implemented in this experiment. It does not establish superiority over PNG, WebP, AVIF, JPEG XL, JPEG 2000, or every possible tile system.

Research package

Paper

Architecture, algorithm, public result, H2 holdout, negative results, limitations, and next cycle.

Open paper →

Reproduce

Commands, evidence files, frozen candidate identity, and what must remain unchanged.

Open guide →

Media / replication

Plain-language description, claim-safe summary, citation, and independent replication checklist.

Open kit →