Five working demos from along the way, each one still runnable.
These pages were built during the research to answer one question each, and they survived because the
answer still matters. Every asset each demo needs travels inside it or beside it in this bundle, so what you
load here is what the page was when it settled the question. Press a card to mount the demo right here, or
open it standalone.
The packed single-file splat viewer from the August rerun: nine exterior
reconstructions with their measured arcs, in one page that needs nothing but a browser tab. The site-wide
viewer on the Capture page is its grown-up sibling.
The day the project learned what makes a generated clip reconstructable. R2's
rejected generation beside its rotation-dominant control, the r6 dolly candidates with their solve
verdicts, and R3, the first clip the arbiter fully solved, beside the control that made it work.
Honest result2.4 MB
The registered views worked. The sparse splat did not.
The two-track benchmark review: a registered 29.88-degree warp and a
deterministic-parallax branch against the video models, scored on anchor fidelity, structure and house
fill. The warp won its lane and its splat stayed rejected, and this page shows both halves.
The plate that put the front-arc and full-turn reconstructions side by side at
eight shared bearings, the study behind section 09k of Missing angles, preserved as its own page.
The R5 walkthrough: coherent appearance pins propagated camera to camera out to
59.75 degrees, the f240 frame quarantined with its pose left honestly unknown, and the viewer bug the
attempt flushed out. The clearest end-to-end telling of the measured-orbit pipeline.