[Stanback]

Summary

• Empirical Course Conversion Model (turnmark-empirical-2026):
  • Built and evaluated a 6,320,277-race US competition dataset (national-checkpoint-15) across 213,105 source-scoped swimmers and 4,721 meets (2010–2026, 13 regional/national archives), paired with a deduplicated national registry of 1,813 physical aquatic facilities and 2,175 certified pool basins (92.2% meet-to-facility resolution).
  • Trained a shrinkage-regularized, strictly monotonic Tier 1 Huber Pace-Curve Spline (154 cells + pooled fallbacks) embedded in crates/turnmark-core via compact data/empirical.json (10.6 KB).
  • Across 25,855 held-out 30-day same-swimmer transitions (3,223 disjoint test swimmers):
    • Reduces overall MAE from 2.9362s (NCAA) and 3.5329s (Colorado) to 2.6658s (−9.21% vs. NCAA, $z = -15.57$; −24.54% vs. Colorado, $z = -22.92$), matching an unconstrained monotonic GBDT ensemble (2.6648s) to within 0.0010s.
    • Beats both NCAA and Colorado across all 14/14 canonical events and all 6/6 course directions, including 1650/1500 Free (−3.32s / −10.6% vs. NCAA, −19.73s / −41.4% vs. Colorado), 1000/800 Free (−1.88s / −11.7%), 500/400 Free (−0.82s / −11.6%), 200 Fly (−13.7% at 30d / −20.6% at 60d), 400 IM (−12.0% at 30d / −16.6% at 60d), and 200 Breast (−21.0% vs. NCAA).
    • Collapses NCAA's speed-quartile bias slope (Q1: −0.554s to Q4: +0.587s) to near zero (Q2: −0.022s, Q3: −0.034s, Q4: −0.048s).
    • Isolates same-facility bulkhead transitions (2.1942s MAE / 1.36s Med AE across 5,925 pairs) vs. cross-facility travel (2.8261s MAE) and high-altitude transitions (3.6674s MAE across 1,287 pairs at $\ge 3,000\text{–}6,900\text{ ft}$).
• Hybrid Local Lakehouse & Research Report:
  • Dual-store SQLite B-tree (0.011 ms warm athlete-history lookup) + Iceberg-partitioned Parquet warehouse (18.8× smaller, 2.32 ms warm DuckDB analytical scan) in research/conversion-model/lakehouse.py.
  • Full paper-grade report in research/conversion-model/REPORT.md and aggregate receipts in evaluation-checkpoint-06.json & acquisition-checkpoint-06.json.
• First-Principles Method Comparison UI Across Web, Embed Widget, MCP Apps, iOS, & Android:
  • Exposes Turnmark Empirical alongside NCAA (default) and Colorado Timing across all SDKs and platforms while keeping NCAA as the default rule.
  • Adds Course & Turn Geometry (lengths · turns · ±walls), an Interactive Method Spectrum Track, Formula Decomposition & Signed Deltas vs. NCAA, and an inline SVG Pace-Divergence Curve & 3-Point Sensitivity Strip (Faster −8%, Your swim, Slower +8%).
• Benchmark Selector Ordering Polish (Closes Polish benchmark selector ordering across web and native apps #43):
  • Groups USA Swimming · Youth, USA Swimming · Futures, Junior Nationals (USA Swimming · Junior Nationals on native), NISCA · All-America, and USMS · Masters under Time standards, followed by World Aquatics points under Points, across Web, iOS, and Android.