S16 · Every champion’s 26.19 / 26.20 patch sensitivity
All 173 champions × 5 roles · Fixed S16 history, target patch26.20, field and shared S15 coefficients · No refit
These are model sensitivities, not observed balance effects. The dataset contains no completed26.19/26.20 games. Frozen-model extrapolation remains; large gains are not independently validated conclusions.
The frozen Rocketbelt20→10 haste component has a confirmed sign error affecting13 champions directly. A separate corrected-sign diagnostic is shown without replacing the frozen forecast.
Overall CSV (V1/V2) · Role CSV (V2) · Top25 PNG · Validation
Interpretation
No patch: zero both patches’ numeric, mechanism and system/item encodings while keeping every other input. Frozen: both enabled. Each patch contribution averages its marginal effect under both addition orders, so the two columns sum to the total; this is nonlinear model attribution, not an observed chronological change.
An unchanged champion can move because of other champions and probability normalization. Pick, ban and presence are whole-event proportions. A change from10% to15% is+5 percentage points.
Sorted by absolute combined change in the selected metric. Click a champion for roles, encodings and diagnostics. Extrapolation and role-revival flags diagnose V2; the haste error concerns shared encodings used by both models.
| Champion | No patch | Frozen | Combined / pp | 26.19 / pp | 26.20 / pp | Pick Δ / pp | Ban Δ / pp | V2 / encoding flags |
|---|
Four model scenarios
| Metric | None | 19 only | 20 only | Both | Δ / pp |
|---|
Five role pick rates
| Role | None | Frozen | Combined / pp | 26.19 / pp | 26.20 / pp | Observed last45 days |
|---|
Show patch encoding magnitudes
Values are100×weighted log changes before old-patch exposure adjustment. They are not probability percentage points or measured champion strength. Role relevance, ability importance and item exposure include declared assumptions.
| Role | 19 numeric | 19 mechanism | 19 system | 20 numeric | 20 mechanism | 20 system |
|---|
Patch-feature extrapolation diagnostic
Flag rule: distance from the training mean exceeds5 saved standardization scales. The scale has a0.025 floor; this is not an empirical standard deviation or significance test. The zero-recent-picks revival flag requires a forecast≥2% and patch-driven increase≥2pp.