[Stanback]

Summary

Replaces the pace-curve regression behind the turnmark-empirical-2026 method with Turnmark Equivalence, a model-based estimate of the equivalent time in another course. Each (event, category, course) gets a linear map in log time onto one SCY-anchored shared scale, fit from 8.51M swims (national-checkpoint-17). It estimates the equivalent time, not "what will this swimmer swim next"; that question is the separate predictor POC below.

Model (research/conversion-model/evaluate_and_train.py, receipt evaluation-checkpoint-08.json, method in REPORT.md)

• Deming (errors-in-variables) lines per cell, with the noise ratio estimated from same-course repeat swims, weighted by swimmer.
• Event lines shrink toward pool geometry and a flat multiplier; category cells shrink toward their event line.
• Weighted least squares across the three course pairs gives one map per course, so A→B→A and A→B→C agree before rounding.
• Evidence is counted over swimmers, not rows. The shift variance is swimmer-clustered and the slope variance is a disclosed approximation. Repeating a swimmer's rows changes nothing (tested).
• No nonlinear stage. An earlier revision fitted one to flatten a pair-mean diagnostic that trends under unequal course noise even for the true map, so it was removed.
• Validation was used to look for remaining curvature; test drove no choices.
• data/empirical.json is schema 3: two-knot tables, lowEvidenceCells and preciseLatentRanges (14 KB).

Test split (39,314 rows, 4,710 held-out swimmers)

Method	MAE
NCAA	3.041s
Flat multiplier	2.858s
Percentile, same input	2.788s
Equivalence	2.771s

• Main evidence: −0.270s vs NCAA (CI −0.302 to −0.240) and −0.087s vs the fitted flat multiplier.
• Against percentile maps: not distinguishable after Holm correction.
• Errors by ability: a separate swim at least 31 days from both pair swims is a less-coupled diagnostic, not an unbiased one. On it, NCAA’s fastest-minus-slowest spread is −2.8% and this map’s −0.86%. A mid-range bend remains, which is a lead for a known-truth experiment, not a fix to ship.
• Arithmetic: round trips are within 0.02s and there are no monotonicity violations.
• Known limits:
  • Same-course repeats can't measure swimmer-specific course aptitude, so the noise ratio is an estimate.
  • The slow tail converts fast on the later-season and regional holdouts.
  • 1650/1500 is the one event NCAA wins.
  • Out-of-range times extrapolate.

Engine and clients

• turnmark-core adds knot-map conversion and the EmpiricalOutOfRange and EmpiricalLowEvidence caveats; out-of-range wins when both apply. Elevation is computed in whole hundredths. The parity fixture has 1,008 cases.
• Web, iOS, Android, SDK and MCP:
  • label the method "Turnmark Equivalence" and describe its result as an estimate;
  • show both caveats ("Outside the model’s fitted range: this estimate extrapolates beyond the available fit.").
• The MCP resource is now r7.
• canonical_event_key no longer buckets SCY 400 free or 800/1500 yd free with the 500/400 m events.

Also in this PR

• Review fixes to engine arithmetic and evaluation design.
• Canadian directory: clubs, facilities and university swim programs, with no contact data.
• research/conversion-model/predictor/: a forward-only "likely time" predictor POC, not shipped.
  • Features require age recorded at the source swim, and a test checks that target-only fields can't change them.
  • 3.32s MAE vs 4.97s for NCAA.
  • Its tests aren't in CI.
• Widget Wasm test: now asserts the standards payload is absent rather than a size ratio.
• MCP spectrum labels stay inside the track at large text.

Test plan

• Python research tests (101: known-truth recovery under unequal noise, duplication invariance) and predictor tests (16)
• Independent reviewers reproduced the full evaluator run exactly and audited every report number against the receipt
• cargo fmt, clippy, workspace tests
• npm test (187), check, engine, mcp, sdk:package, Playwright (55), production Playwright (22), CI-change tests (84)
• iOS app suite locally (28/28)
• Android instrumented tests: the history tests fail in CI (under investigation); not runnable locally
• iOS benchmark UI tests fail in CI; they were already failing on this PR before this revision

🤖 Generated with Claude Code