ZPE-Image

What This Is

Deterministic sparse-stroke image encoder. Geometry-layer encoding for glyphs, flow graphs, mazes, and structural skeletons; natural-image coverage stays out of scope.

Two routes ship: a primary sparse-stroke route and a narrower hole-bearing bundle route for exactly three hole-bearing forms. Both are CI-falsified.

Strongest intrinsic result (CI-anchored, no external baseline): 5.75× byte reduction on the five accepted sparse figures versus the internal quadtree-enhanced fallback, with 100% accept and reject rates and perturbation floors clearing documented thresholds.

Codec Mechanics

Key Metrics

Source: proofs/artifacts/fresh_falsification_packet.json, validation/results/fresh_falsification_check.json

Repo Identity

Readiness

Honest Blocker

We do not claim photo, texture, gradient, or broad natural-image coverage.; We do not claim that the narrower hole-bearing route covers the full sparse set.; We do not claim that bounded acceptance on this pack equals general image coverage.

What We Prove

• The primary sparse-stroke route accepts glyph_a, fork_tree, loop_spine, maze_turns, and serpentine, and rejects all 7 retained mixed and natural-image negatives on the same verification pack.
• Under four perturbation types (dilate_1, salt_pepper_1pct, shift_x1, shift_y1) the worst-case reconstruction IoU is 0.632 and worst-case skeleton F1 is 0.741, both clearing their documented thresholds (0.62 and 0.74 respectively).
• On accepted sparse figures the geometry-sparse-stroke encoder uses a mean of 1,362 bytes (20-bit packed) versus 7,839 bytes for the quadtree-enhanced fallback — a 5.75× byte reduction within this bounded scope.
• The narrower hole-bearing bundle route accepts exactly glyph_a, loop_spine, and maze_turns; fork_tree and serpentine are correctly outside that subset. Bundle projection IoU = 1.0 and skeleton F1 = 1.0 under all perturbations for all three accepted cases.

What We Don't Claim

• No photo, texture, gradient, or broad natural-image coverage.
• The hole-bearing route does not cover the full sparse set.
• Bounded acceptance on this pack does not equal general image coverage.
• No external codec comparison (JPEG/WebP/AVIF/JPEG-XL) — baselines are internal only.

Verification Status

Proof Anchors

Repo Shape

Extended Metrics

Rows retained from the previous expanded ## Key Metrics table. The public product page uses the first four rows only.

Quick Start

python3 -m venv .venv
. .venv/bin/activate
python3 -m pip install --upgrade pip
python3 -m pip install '.[dev]'
zpe-image-verify --output validation/results/fresh_falsification_check.local.json
pytest -q

Upcoming Workstreams

This section captures the active lane priorities — what the next agent or contributor picks up, and what investors should expect. Cadence is continuous, not milestoned.

• External comparator integration — Active Engineering. Add gzipped-SVG and PNG-of-render-at-fixed-quality comparators to the proof artifact; report both, let reviewer pick. Removes the no-market-reference-frame gap on the current 5.75× CR claim (currently measured only against an internal quadtree fallback).