Date: 2026-04-24 (computed), 2026-04-26 (TOSS promoted to primary) Effective: 2026-04-20 (models computed in parallel); 2026-04-26 (TOSS becomes primary) Branch: claude/fix-ranking-league-balance-HdQrV Author: Engineering, after multi-round design review with site owner and a designer-supplied alternative spec.
Starting with the 2026-04-20 publish window, the rankings system computes three Power Index variants in parallel. Starting 2026-04-26 (Week 4, the first Sunday-cadence snapshot), TOSS is the primary Power Index — it drives the State Rank, Class Rank, Power Index column, head-to-head tiebreakers, league standings, and the playoff simulator. The pre-2026-04-26 RPI-based formula is retained as Legacy for comparison, accessible via the Model dropdown above the rankings table. QWS continues as the experimental B of the A/B test, also accessible via the dropdown, and remains under evaluation for possible adoption in 2027.
The weekly publish cadence also shifts from Saturdays to Sundays starting 2026-04-26 so that Saturday match results are included in that week's snapshot.
The site owner's read of the season-to-date rankings: the Power Index was over-rewarding teams that ran up dominant flight scores against weak-league opponents. Specifically, the Flight-Weighted Score (FWS) — half of the Power Index — is computed without any reference to opponent strength, so an 8-0 sweep of a cellar team contributes the same as an 8-0 sweep of a state contender. Teams in genuinely competitive leagues, who can win 5-3 matches but rarely 8-0, were being penalized in the depth half of the formula even when their dual-match record signaled they were stronger.
The owner explicitly did not want to back-date any change ("would not do this retroactively"), did not want to remove FWS entirely (it captures a real signal — roster depth), and did not want to introduce an explicit "league strength" metric (OSAA doesn't use one). The fix needed to be implicit, schedule-aware, and forward-only.
Two distinct design approaches surfaced during review:
1. TOSS (engineering-proposed) — leave APR alone (it already includes strength-of-schedule via OWP and OOWP, plus the Pass 2 league-depth adjustment). Make FWS opponent-aware by multiplying each match's FWS contribution by the opponent's APR relative to the median, clamped to a [0.75, 1.25] band. Smallest blast radius, surgical fix to the specific distortion the owner called out, and naturally captures the transitive "if you beat a team that played strong opponents, you get credit" property because APR contains OOWP.
2. QWS (designer-proposed, modeled on the ITA college tennis ranking) — replace RPI-style APR entirely with a quality-weighted wins formula: each win earns points equal to the opponent's Power Index × 100, each loss costs a flat 50 points, normalize per team and iterate to convergence. Larger structural change, mirrors a proven college-tennis ranking, eliminates known RPI cascading-weak-opponent pathologies. The flat 50-point loss penalty was a point of concern (it doesn't differentiate a close loss to the #1 team from a 0-8 blowout against the cellar) but expected-value math still incentivizes scheduling up because strong-opponent wins are worth more.
After comparison, the owner chose to ship both in parallel rather than pick one blind. The two formulas attack different halves of the Power Index: TOSS adjusts the FWS half, QWS replaces the APR half. Running both alongside the unchanged Current model through end of 2026 lets us see — on live data — which approach better matches the eye test, and decide for the 2027 season with evidence rather than intuition.
For every team in the 2026 rankings, processed_rankings.json now carries:
power_index, apr, normalized_fws, rank, class_rank — Current (baseline, unchanged)power_index_toss, adjusted_fws, schedule_multiplier, rank_toss, class_rank_toss — TOSSpower_index_qws, apr_qws, qws_iterations, rank_qws, class_rank_qws — QWSHistorical seasons (2021-2025) and the three already-published weekly snapshots (public/data/weekly/2026-04-04.json, 04-11.json, 04-18.json) keep only the baseline fields and are byte-for-byte unchanged.
A new "Model" filter in the rankings toolbar swaps the State, Class Rank, and Power Index columns to display the selected model's values. Default is Current. Switching to TOSS or QWS:
_toss / _qws fields[TOSS] / [QWS] label so it's obvious which model is in viewFor each dual match the team played:
m_i = clamp(opp_APR / median_APR, 0.75, 1.25)
adjusted_fws = sum(match_fws_i * m_i) / sum(m_i)
power_index_toss = (APR * 0.50) + (adjusted_fws * 0.50)opp_APR is the Pass-2 APR (post-league-depth-adjustment), so the multiplier already encodes the transitive opponent-of-opponent signal. Unknown opponents (e.g. Idaho schools that occasionally show up in border-school schedules) fall back to APR = 0.5, matching the existing OWP/OOWP convention. The weighted-average normalization (÷ sum(multiplier), not ÷ count) keeps adjusted_fws on [0, 1] so it composes cleanly with the existing 50/50 PI split.
Behavior in practice:
(5/8) × ~1.25 = ~0.78 credit (vs ~0.62 raw)(8/8) × ~0.75 = 0.75 credit (vs 1.0 raw)(3/8) × ~1.25 = ~0.47 credit (vs ~0.38 raw) — losses to strong teams are softened, not punished extra0 × ~0.75 = 0 credit — no change (zero is zero)For each team across the 2026 season:
total_matches = wins + losses + ties
quality_points = sum(opp_PI * 100 for each WIN, plus 0.5 * opp_PI * 100 for each TIE)
loss_penalty = 50 * losses + 0.5 * 50 * ties
raw_apr_qws = (quality_points - loss_penalty) / total_matches
apr_qws = clamp((raw_apr_qws + 50) / 150, 0, 1)
power_index_qws = (apr_qws * 0.50) + (normalized_fws * 0.50)Iterated: seed opp_PI = opponent Win% on iteration 0, then use the previous iteration's power_index_qws for subsequent iterations until max |Δ PI| < 0.01 or 5 iterations. On live 2026 data, every team converges in 3 iterations.
Weeks 1-3 (2026-04-04 / 04-11 / 04-18) shipped on Saturdays. Beginning week 4 (2026-04-26), weekly snapshots publish on Sundays so any Saturday match results land in that week's file. Week numbering is continuous across the transition:
| Date | Day | Week # |
|---|---|---|
| 2026-04-04 | Sat | 1 |
| 2026-04-11 | Sat | 2 |
| 2026-04-18 | Sat | 3 |
| 2026-04-26 | Sun | 4 |
| 2026-05-03 | Sun | 5 |
| 2026-05-10 | Sun | 6 |
get_week_publish_date(), get_prior_publish_date(), and all_week_publish_dates() in scripts/generate_weekly_rankings.py handle the transition. The legacy get_week_saturday() / all_week_saturdays() names are kept as aliases for backward compatibility. Running the script with no arguments during the cadence transition window (Apr 19-25) intentionally rolls *forward* to 2026-04-26 instead of overwriting the prior Saturday snapshot.
ADJ_MODELS_ENABLED=0 python generate_site.py produces a processed_rankings.json byte-identical to the pre-change main branch. The new models are entirely additive when the gate is off. (The gate is also automatically off for any run before 2026-04-20.)
The three pre-existing weekly snapshots in public/data/weekly/ show no diff from main. Historical seasons 2021-2025 emit only the baseline fields — no power_index_toss / power_index_qws keys at all.
- Current ↔ TOSS: 0.97 - Current ↔ QWS: 0.95 - TOSS ↔ QWS: 0.98
Both fixes agree on the big picture. Where they diverge — and they do — that's the comparison data we built this for.
The expected ordering at the top of 2026 Boys (Lincoln, Jesuit, Grant, OES, Catlin Gabel) holds across all three models with only minor reshuffles. The biggest QWS drops are mid-pack 6A/5A teams whose records were padded by weak league bottoms (Sam Barlow -49 spots, Ridgeview -40, Redmond -32). The biggest QWS rises are teams that played tough schedules and earned losses but had quality wins to show for it (Springfield +21, Century +20, Thurston +20). These are exactly the patterns the formula is designed to surface.
After one week of parallel data, the decision to promote TOSS and keep QWS experimental came down to three things:
1. Honors the "don't penalize scheduling up" constraint more cleanly. TOSS's per-match opponent multiplier rewards partial FWS earned against strong opponents even in losses. QWS's flat 50-point loss penalty doesn't differentiate a 3-5 loss to the state's #1 team from a 0-8 blowout by the cellar — both cost 50. La Salle Prep boys (4-5 in 5A-1 NW Oregon) drop from #22 → #40 under QWS specifically because of this; TOSS barely moves them (#22 → #20). That's a narrative problem we'd rather not own mid-season. 2. Smaller, more defensible change. TOSS keeps the OSAA-compatible RPI APR intact and only adjusts how flight scores count. Explainable in one sentence to coaches. QWS replaces APR entirely with an iterated quality-weighted-wins formula — a structural change that's harder to defend mid-season without more data. 3. Fixes the obvious cases either way. The biggest distortions (Ridgeview boys #16 → 62, Sam Barlow #32 → 66, The Dalles girls #11 → 42) are corrected under both models. We don't need the more aggressive model to surface the clear bugs.
QWS isn't wrong — structurally it's arguably the cleaner long-term answer because it eliminates RPI's cascading-weak-opponent pathologies entirely. But the flat loss penalty needs tuning before it can be primary. A 2027 candidate is a variant where the loss penalty scales inversely with opponent strength (e.g., 100 × (1 - opp_PI) so losing to the #1 team costs ~20 points and losing to the cellar costs ~100). We'll evaluate that through end of 2026 and decide for next year.
1. Watch through end of season. Each Sunday snapshot will preserve its TOSS-primary ranks plus power_index_legacy and power_index_qws for comparison. By end of the 2026 season we'll have ~5-6 weeks of side-by-side data. 2. Decide for 2027. With live evidence in hand we can keep TOSS, adopt QWS (possibly with a tuned loss penalty), or refine TOSS's clamp band. No code change happens until we have data to argue from. 3. Coach feedback. With the Model dropdown live on the main page, coaches can verify the switch themselves. We expect questions; the methodology page explains all three formulas.
generate_site.py — TOSS and QWS computation, server-side alt-model rank/class-rank, model selector dropdown, dynamic rank/PI/class-rank columns, methodology banner.public/methodology.html — new "Alternative Models (A/B Test)" section with formulas and JSON field references.scripts/generate_weekly_rankings.py — Saturday-to-Sunday cadence with continuous week numbering and backward-compat aliases.CHANGELOG.md — short summary entry pointing here.AAR-power-index-ab-test.md (this file) and rendered public/aar-power-index-ab-test.html.public/data/processed_rankings.json — regenerated artifact carrying the new fields.