Franken Drone Geometry Reconstruction — Development skill for Claude Code

franken_drone_geometry_reconstruction (`fdgr`)

[](./LICENSE)

**An agent-native, evidence-grade operating substrate for turning owner-authorized drone observations into metrically honest semantic digital twins.**

**Current implementation boundary — 2026-09-04:** FDGR has a 28-member, dependency-closed safe-Rust reference workspace. The executable path covers immutable recorded-media custody, bounded native ISO BMFF/sample evidence, canonical timelines, clock/calibration/scale semantics, deterministic keyframes and correspondences, epipolar and relative-pose adjudication, graph topology, component-local orientation, relative edge-baseline gauges, component-relative camera-center initialization, translation-only robust pose refinement, structural bundle-problem compilation, and a separate image-domain/seed-provenance/held-out-independence audit. It does **not** yet claim live DJI acquisition, owned FFmpeg execution, native feature or motion-candidate generation, joint pose-landmark bundle adjustment, metric camera pose, sparse/dense reconstruction, semantic resolution, cloud recovery, a complete agent runtime, or production qualification.

FDGR is designed for a person who manually flies a compact drone such as the DJI Flip around a property while an agent helps decide what to observe, what to compute, what can be trusted, what remains unknown, and what should happen next.

The product is not “a mesh from a video.” It is a **multi-version evidence and constraint database for physical reality**, operated through one coherent cognitive control loop.

The system from the driver’s seat

After every success, progress event, refusal, or error, a competent agent should be able to answer:

1. What is established at the exact current anchor?
2. What changed since the anchor I understood?
3. Which uncertainty matters to the active objective?
4. What work is active, bloc