Conservators call the gap a lacuna. Museums hold millions of them — paint losses, torn photographs, worm-eaten manuscripts. Machine-learning tools can now fill those gaps convincingly, and that is exactly the problem: a convincing fill, presented seamlessly, is a guess wearing the costume of the object.
Lacuna is a lab built on the opposite stance. It runs several reconstruction methods over the same loss, shows you where they disagree, and stamps every result with its method, parameters, source, and content hash — so the claim can travel, but never masquerade.
No seamless "restored" output exists here. A reconstruction is presented as what it is: one method's story about missing content, flagged so a downstream reader can never mistake it for the surviving object.
Run diffusion, exemplar synthesis, and learned models over the same loss and they tell different stories. Lacuna renders that divergence as a heatmap with a score — the honest measure of how much any fill should be trusted.
Method, parameters, model weights, source object, license, content hash — bound into a portable, content-addressed record and kept in an open ledger. Facts divorced from their documentation history stop being facts; we keep the history attached.