GeoAlibi

GeoAlibi — A change needs an alibi. Satellite images can appear to show new construction or damage because of misalignment, lighting, shadows or seasonal differences. GeoAlibi helps investigators challenge alternative explanations before supporting an apparent change. The workstation combines change proposals, registration diagnostics, appearance checks and bounded perturbation tests with optional Meta-powered interpretation and building segmentation. It displays supported, artifact and review candidates, discloses regions outside the current inspection batch, and exports evidence packets containing imagery, measurements and checksums. Linked inspection batches preserve the original full-scene context. On eight frozen full LEVIR test scenes, pretrained U-Net proposals achieved 78.66% pixel F1 versus 8.08% for the RGB baseline. The verification stage accepted zero pixels because its registration gate flagged all eight scenes unreliable; supported recall was zero and precision undefined. These are small-subset diagnostic results, not full-benchmark accuracy. The release includes a reproducible offline demonstration, evidence-verification tools and 220 passing tests. Its contribution is an inspectable record of what survived testing, what remains unresolved and what evidence is needed next. Independent alignment review and corrected real-image Meta validation remain outstanding. For judges: run the actual local application and inspect exported evidence using https://github.com/ADI238IND/GeoAlibi/blob/main/docs/JUDGE_QUICKSTART.md . The submitted video remains an offline procedural walkthrough, not a live application recording or live Meta inference. An isolated fixed SIFT registration follow-up on the same eight observed scenes passed 0/8 real pairs and 8/8 same-image controls; it is research-only and does not improve supported recall. Report: https://github.com/ADI238IND/GeoAlibi/blob/research/stable-background-registration/docs/SIFT_OBSERVED_EIGHT_FOLLOWUP.json