Benchmark + autoresearch harness for the flattening stage#68

Draft

mvdoc wants to merge 35 commits into

devfrom

benchmark-autoresearch

Commits on Jun 10, 2026

Add benchmark/autoresearch harness for the flattening stage (Phase A)

mvdoc
and
claude
committed
Add Tutte/LSCM flip-free init probe (Phase B primary experiment)

mvdoc
and
claude
committed
Add unit tests for Tutte flip-free init and area scaling

mvdoc
and
claude
committed
Add fast flatmap plotter for visual verification of experiments

mvdoc
and
claude
committed
Add general autoresearch experiment runner

mvdoc
and
claude
committed
Document experiment.py runner and plot.py in benchmark README

mvdoc
and
claude
committed
Add curated findings: Tutte init verified (n=4), phase-removal ablations

mvdoc
and
claude
committed
Add optimizer-swap probe (L-BFGS/CG/Adam on the same energy)

mvdoc
and
claude
committed
Findings: optimizer-swap is a dead end on this energy; multiscale is the lever

mvdoc
and
claude
committed
Add spectral (manifold-harmonic) multigrid probe

mvdoc
and
claude
committed
Findings: multigrid + Adam optimizer experiments

mvdoc
and
claude
committed
Add speed levers to experiment runner (line-search points, iters, smoothing cap)

mvdoc
and
claude
committed
Findings: how to speed up the FreeSurfer optimizer (~2.5x, near-free)

mvdoc
and
claude
committed
Findings: validated 3.6x speedup (fast_ultimate, n=4)

mvdoc
and
claude
committed
Findings: energy/quality angle - true-geodesic metric + correction recalibration

mvdoc
and
claude
committed
Findings: true-geodesic metric, local-vs-global, distance-optimal output scale

mvdoc
and
claude
committed
Findings: ground objective in Fischl 1999 (metric distortion, no area term)

mvdoc
and
claude
committed
Bench: global true metric + target-scale sweep; correction ~1.10 wins on global

mvdoc
and
claude
committed
Findings: target correction re-judged on global metric (cross-hemi sweep)

mvdoc
and
claude
committed
Findings: broadened correction validation (8 hemis/7 subjects) - not a robust default

mvdoc
and
claude
committed
Findings: executive summary + refreshed next-ideas (autoresearch wrap-up)

mvdoc
and
claude
committed
Validate fast config on 9 hemispheres: not overfit, ~3.4x at equal-or-better quality

mvdoc
and
claude
committed

Commits on Jun 11, 2026