v0.1.0 — initial SLN+DREDge checkpoint

Joint spike localization + motion drift correction (SLN+DREDge)

Snapshot of working code + delivered figures at the time of NeurIPS 2026 submission prep.

Headline result — dataset1_p1 (NP 1.0, 2.48M spikes)

method	apply ρ̄	apply H̄	Δρ̄ vs MP+DREDge
MP raw (no MC)	0.267	6.27	−0.301
MP+DREDge canonical	0.568	8.25	0
CNN-SLN all-spike postdredge ep20	0.663	8.14	+0.095 ← best
TR-SLN all-spike postdredge v3 ep2	0.641	8.19	+0.073

See figures/comparison/aggregate_3method_comparison.png for the hero image and docs/COMPARISON.md for the full scoreboard.

Contents

• 31 Python source files (src/) — models, training, apply, 13 visualization scripts
• 88 final visualizations (figures/) across 7 method variants + cross-method comparisons
• docs/COMPARISON.md — scoreboard with full viz coverage matrix
• docs/scoreboard.csv — per-axis Pearson ρ (global + local frame)
• docs/PROBLEM_STATEMENT.md — problem framing + architecture details

Excluded (per repo policy)

• .mp4 localization movies (~370 MB; regeneration recipe in README)
• .pt model checkpoints (rerun training pipeline to reproduce)
• 1958×1958 pairwise NCC .npy matrices (regenerable via make_xy_pairwise_correlation.py)

Key findings

1. Data scale > architecture. Both 500K-subset models (CNN, TR) underperform MP+DREDge; both all-spike models beat it by +0.07–0.10 ρ̄.
2. z gets an implicit gradient through shared-encoder representation drift, even though z has no direct gradient in the loss.
3. One outer DREDge iteration is enough — after 20 inner SLN epochs, a second DREDge round produces a motion field 99.7% correlated with the first (1.18 μm RMS difference).

Verification

All 9 core modules import cleanly and both models (CNN-SLN 402K params, TR-SLN 580K params) forward correctly on synthetic input. End-to-end smoke test on a 1k-spike synthetic dataset passes for all viz scripts (depth raster, spatial entropy, pairwise NCC, aggregate projections).