Per-division ridge regression on 22 features: the 14 abilities plus combo features (Sprint&Acceleration, min(Mountain,TT), Hill&MidMountain, Cobbles&Flat, TT&Prologue, endurance average, and the rider's best and second-best discipline scores). Abilities only count above a division-specific threshold (PT 65, PCT/CT 70) and are raised to a power (2–2.5), which captures that points come from high-end stats, not averages. The target is square-root points, squared back for display. Predictions are then top-end calibrated against actual 2025 points (band medians raw→actual, monotone; full-range for PT at 75% strength, restricted for PCT at 50%, none for CT where no bias was found) — this fixes the systematic under-prediction of 300+ point riders (was +65 to +130 points, now near zero). Trained on the full 2025 season (all divisions, riders without points count as 0). The 2021–24 PCT seasons are included at 15% weight each (PCT model only), with MidMountain imputed for pre-2025 databases. Out-of-sample R² (5-fold CV, points space): PT 0.81, PCT 0.72, CT 0.54.

Tested on 2025 residuals (actual minus stat-based projection). Internal competition dominates: the team's #1 sprinter overperforms (×1.19) while sprinters behind a better teammate underperform (×0.89 as #2, ×0.84 as #3+; permutation p<0.001, n=169). Puncheurs show the same but weaker (#1 ×1.16); climbers/TT/GC riders none.

Lead-out trains help the #1 sprinter only: among 59 clean #1 sprinters, points rise ≈3.9% per point of train quality above 72 (train quality = mean of the two best teammate lead-out scores). A weighting grid search found the helpers' engine-relevant profile is LO = 0.12·Flat + 0.48·Sprint + 0.40·Acc — Sprint and Acceleration count almost equally, Flat barely, and the Sprint-vs-Acc balance within a helper is irrelevant (corr +0.07). The boost estimate carries wide uncertainty (bootstrap CI includes zero in the relative form) — treat it as an informed lean, not a law. Separately,

ACC-heavy sprinters (Acc ≥ Sprint+3, sprint level 72+) beat projections ×1.12–1.14 consistently across all three divisions (n=69) — a real engine effect the ability model cannot express.

Acceleration's effect differs sharply by rider type — but a full within-puncheur screen revealed the real driver: Sprint is the dominant hidden stat for puncheurs (Spearman +0.48 with residuals, positive in every division; Sprint 69+ puncheurs deliver ×1.40 of projection, Sprint ≤64 only ×0.82, and Sprint 69+ with Flat 72+ hits ×1.47 — Lutsenko, Van der Poel, Valter, Vansevenant are the archetype). ACC 76+ without the sprint (Sprint <69) shows nothing (×0.91): the apparent ACC effect was a Sprint proxy. Hilly finishes come down to reduced-group sprints, and the model's linear pricing misses that a puncheur's Sprint converts placings into wins.

Mountain, MidMountain, Hill, Stamina, Recovery, Cobbles, Downhill and Fighter all show no residual effect within puncheurs; TT/Prologue are marginal (p≈0.03–0.05, not acted on). For classics riders and climbers

ACC shows no residual effect at all; high-ACC GC/TT hybrids slightly underperform (small samples — read as "odd profiles the model overrates", not as ACC hurting).

The full within-climber screen (n=129) found three independent effects: Stamina (74+ → ×1.13; the real damage is at the bottom: ≤69 → ×0.77; strongest in PT where mountain races are longest, weak in PCT), Sprint again (65+ → ×1.19, ≤60 → ×0.89 — reduced mountain groups also finish in sprints; Areruya 1319→1739 is the poster case), and a smaller independent Resistance effect (75+ → ×1.09 even without high Stamina). Mountain, MidMountain and Hill variation within climbers shows nothing — terrain is priced correctly; the finishing and endurance package is what the linear model misses.

The classics screen (n=117) flips that: their hidden value is terrain versatility — MidMountain 68+ classics riders deliver ×1.22 (MM ≤63 → ×0.90; Mountain and Hill show the same but overlap heavily, so MM carries the rule); own Cobbles beyond the archetype adds nothing.

The sprinter screen (n=169) found extreme top-stat convexity: among team-#1 sprinters, Sprint 79+ delivers ×1.26 while Sprint 76–78 manages only ×0.92 — the engine is near winner-take-all at the bunch-sprint top (Farantakis 791→1670, Aniolkowski 510→1263). The raw 76–78 group looks even worse (×0.76) but that is mostly the competition effect: only 24% of them are their team's first sprinter. Hill on top of high Sprint is the mirror image of the puncheur-Sprint finding, but with a different risk profile: for Sprint 75+ riders, Hill adds nothing hidden up to ≈72 (bands 1.00–1.03 — the model prices it correctly), while the hybrid corner (Sprint 72+ and Hill 72+, n=39) delivers ×1.19 on average — but median ≈0 and only 54% positive: a lottery-ticket profile whose average is carried by breakout seasons (Yates +313, Ackermann +226, Démare +219 vs projection). No context rule is applied since the typical rider is neutral; the asymmetry vs the robust puncheur-Sprint effect makes sense — a puncheur's Sprint converts placings to wins in every race he rides, while a sprinter's Hill only pays when the team actually deploys him on hilly race days.

The GC screen (n=149) resolves the earlier anomaly into a clean monotone gradient: the more complete the endurance package, the more the model overrates it — Resistance 77+ GC riders deliver ×0.88 (with Stamina 76+ too: ×0.86; Pogacar's 1531-projection → 896 season is the emblem), Resistance ≤73 "lighter" GC profiles deliver ×1.20. GC points hinge on lumpy three-week outcomes, so stat-stacking inflates projections faster than points.

TT riders (n=131) echo the classics versatility pattern: Mountain 67+ → ×1.14, Mountain ≤60 → ×0.86 (weaker evidence, p≈0.03–0.08, applied with extra shrinkage); their own TT stat beyond the archetype adds nothing. The same residual screen across Stamina/Resistance/Fighter/Downhill/Flat (30 stat×type tests, so p<0.01 is the bar) found three more real effects: climbers with high Stamina overperform (top vs bottom quartile +71 pts, p=0.003; Resistance similar at +58, p=0.008 — the two overlap heavily) and puncheurs with high Flat overperform (+61 pts, p=0.006) — a puncheur needs Flat to be there when the hill comes; his ACC to finish it.

Fighter and Downhill showed nothing anywhere. One anomaly: GC riders in the top Stamina/Resistance quartile underperformed heavily (−90/−131) while the continuous correlation is ≈0 — driven by a few elite stage-racers whose seasons are binary (win GC or bust), not by the stats hurting. These archetype effects were tested as model adjustments and rejected by cross-validation (flat-to-worse out-of-sample in every configuration — the effect sizes are selection-inflated). They therefore live as a labelled engine context layer in the rider detail view (conservatively shrunk multipliers), and deliberately do not feed expected points, wages or VORR.

Practical reading: when buying a puncheur, Sprint is the first-class hidden stat (aim 69+), with Flat 72+ as the positioning complement — ACC only matters via the sprint; for sprinters the market splits at Sprint 79: pay up for 79–80, avoid paying sprint-money for 76–78; Acc-over-Sprint profiles are a cheap bonus; for climbers avoid Stamina below 70 at all costs, pay for Stamina 74+ and Sprint 65+, with Resistance 75+ as a bonus; for classics riders buy MidMountain 68+ (versatility) rather than more Cobbles; never pay extra for Fighter or Downhill on anyone; be sceptical of paying full price for "complete" GC packages (their projections flatter them) and prefer TT riders who can climb.

Climbing trains don't exist. The same analysis for climb leaders (Mountain 79+, n=67) found no benefit from climbing support: teammate climbing quality (top-2 Mountain, with or without endurance weighting) correlates ≈0 with the leader's over/underperformance — if anything slightly negative, because climber-stacked squads are the "complete GC" teams that underdeliver.

There is also no internal competition penalty among climbers (leaders with a better climbing teammate: ×1.00 vs ×0.96 without — nothing, unlike the sprinter ×0.89/×0.84). Mountain points are individual in this engine: hilltop finishes shred the groups, so domestiques cannot deliver a climber the way a train delivers a sprinter. Team-building consequence: never buy climbing domestiques for the leader's sake, but feel free to stack several scoring climbers on one squad — they don't cannibalise each other.

The same screen for puncheur leaders (Hill 77+, n=116) and cobbles leaders (Cobbles 75+, n=81) found no support/train effect for either (teammate punch or cobbles quality correlates ≈0-to-negative with leader performance). Competition differs: puncheurs show none when ranked by raw Hill, but the earlier by-overall-quality split (#1 ×1.16 vs #2+ ×0.98) still applies; cobbles leaders show a mild penalty — team's best cobbler ×1.04 (median +42) vs riding behind a better one ×0.94 (median −20). So the sprint train remains the only team-composition effect that adds points; everywhere else, squad-building is purely about the individuals.