Waiverlord League Winner Experiment

Status: EXPERIMENTAL — v0.1 AND v0.2 FAILED PROMOTION; v0.3 PASSED RESEARCH GATE; PRODUCTION SIGNAL LIVE; EXPERIMENTAL PROGRAM

Launch: 2026-08-21

Governed historical cohort: 2016–2019 and 2021–2025; 2020 excluded by the preregistered COVID-season exception

Public Method: https://www.waiverlord.com/league-winner/method

Abstract

The League Winner Experiment asks: which drafted players created materially more season-long winning value than their August acquisition cost implied, and which 2026 players resemble those historical profiles using only information available before the season?

The experiment separates realized outcome from forward projection. League Winner Score (LWS) measures what actually happened in a completed historical season. League Winner Index (LWI) is the preseason model trained to recognize those profiles. LWI is a relative index, not a probability of winning a league.

Championship Equity is archived. League Winner is Waiverlord's active research program.

Historical cohort

Primary seasons: 2016, 2017, 2018, 2019, 2021, 2022, 2023, 2024, 2025.

2020 remains excluded under the preregistered COVID-season external-validity exception. Any future 2020 inclusion is secondary/non-governing and may not supersede the primary result based on performance.

Specification v0.1 — completed / failed

v0.1 combined Championship Excess, Draft Surplus, Regular-Season Win Value, Availability and Structural Persistence. The governed run contained 1,554 drafted player-seasons and 70 rows with direct ESPN championship evidence.

Aggregate out-of-season results:

LWI Spearman: 0.261
ADP Spearman: 0.225
Prior-production Spearman: 0.200
ADP + prior-production Spearman: 0.234
LWI MAE: 19.309 vs combo 19.612
LWI top-20 lift: 1.541× vs combo 1.447×
LWI top-10 lift: 2.367× vs combo 2.003×
Season Spearman wins vs combo: 7/9
Bootstrap 95% CI for Spearman advantage: −0.003 to +0.060

Result: FAIL. The uncertainty lower bound crossed zero.

The key diagnostic was target asymmetry: ESPN championship reporting is a selected leaderboard, so allowing Championship Excess to enter LWS only when published made evidence availability capable of moving player quality. That violated the intended separation between evidence coverage and player value.

Specification v0.2 — completed / failed forward predictor

v0.2 historical target

v0.2 removed selective championship evidence from LWS entirely. The full-population historical score uses only Weeks 1–14:

45% Draft Surplus — realized PPR production above the expectation implied by historical 12-team full-PPR ADP and position.
40% Regular-Season Win Value — cumulative positive weekly PPR points above replacement.
15% Availability — share of Weeks 1–14 with an nflverse regular-season player record.

Replacement ranks are QB13, RB31, WR37, TE15. Weeks 15–17 are excluded from the primary target, so playoff eruptions cannot create LWS.

v0.2 result

The governed v0.2 run completed on 1,554 drafted player-seasons.

LWI Spearman: 0.298
ADP Spearman: 0.278
Prior-production Spearman: 0.234
ADP + prior-production Spearman: 0.284
LWI MAE: 21.053 vs combo 21.282
LWI top-20 lift: 1.727× vs combo 1.790×
LWI top-10 lift: 2.382× vs combo 2.064×
Season Spearman wins vs combo: 6/9
Bootstrap 95% CI for Spearman advantage: −0.014 to +0.046

Position-level LWI Spearman was positive at all four positions: QB 0.146, RB 0.313, WR 0.352, TE 0.152.

Result: FAIL. v0.2 did not beat the strongest baseline on top-20 lift and the bootstrap lower bound crossed zero.

v0.2 construct validation

The historical LWS target itself passed an important face/construct check. Among the 70 drafted ESPN leaderboard rows with direct championship evidence, v0.2 LWS had Spearman 0.827 with playoff-normalized Championship Excess. ESPN data do not enter LWS, so this is external construct validation rather than circular scoring.

Historical sanity examples:

2017 Alvin Kamara — ADP 140.0, LWS 95.66, Elite League Winner
2018 Patrick Mahomes — ADP 118.1, LWS 92.11, Elite League Winner
2018 Christian McCaffrey — ADP 12.7, LWS 95.42, Elite League Winner
2021 Cooper Kupp — ADP 41.3, LWS 97.20, Elite League Winner
2024 Jahmyr Gibbs — ADP 9.4, LWS 85.07, League Winner

Governed conclusion: LWS v0.2 is retained as the historical league-winner measurement target. The v0.2 forward predictor is rejected.

Specification v0.3 — completed / passed research gate

v0.3 hypothesis

ADP plus prior-season production already contains most obvious public information. The remaining forecastable league-winner signal, if it exists, should be concentrated in career stage, draft capital and trajectory — especially rookies, second-year players and young players whose prior-year box score understates their plausible role.

v0.3 therefore keeps LWS v0.2 unchanged and tests only a new incremental preseason predictor.

v0.3 target — unchanged

LWS remains:

45% Draft Surplus
40% Regular-Season Win Value
15% Availability

Weeks 15–17 remain excluded. Replacement ranks remain QB13, RB31, WR37 and TE15. Championship prevalence remains external validation only.

v0.3 preseason structural inputs

In addition to the v0.2 features, v0.3 uses only the following preregistered pre-season variables:

age on September 1 of the target season;
NFL draft year;
NFL overall draft pick transformed with log1p;
undrafted-NFL indicator where draft-pick evidence is absent but identity is otherwise resolved;
years since NFL draft;
rookie indicator;
second-year indicator;
third-year indicator;
prior-season PPR points/game change versus two seasons prior;
prior-season opportunity/game change versus two seasons prior;
prior-season share of weeks above the governed position replacement line;
explicit missing-history and missing-metadata indicators.

Player metadata come from the nflverse players dataset. Normalized-name fallback is permitted only for unique, non-ambiguous player-position matches and coverage is reported. The completed run achieved 96.8% metadata coverage across the historical drafted population.

No same-season regular-season outcome, playoff outcome, current social attention or post-draft depth-chart outcome enters historical predictors.

v0.3 residual architecture

The strongest governed baseline remains ADP + prior-production proxy. v0.3 explicitly models incremental residual value rather than relearning that baseline.

For each leave-one-season-out fold:

Fit the regularized ADP + prior-production ridge baseline on training seasons only.
Compute training residuals: LWS − baseline prediction.
Fit a shallow HistGradientBoostingRegressor to those residuals using the preregistered v0.2 features plus structural/trajectory variables.
For the held-out season, calculate LWI v0.3 = baseline prediction + residual-model prediction.

Residual hyperparameters were frozen before execution:

learning_rate = 0.03
max_iter = 220
max_depth = 2
min_samples_leaf = 25
l2_regularization = 8.0
random_state = 42

No hyperparameter search or post-result tuning was performed.

v0.3 governed baselines

ADP alone
prior-production proxy alone
ADP + prior-production proxy

v0.2 LWI is also reported as a non-governing comparator.

v0.3 preregistered research gate

A research-gate pass required all of the following:

beat all three governed baselines on aggregate out-of-season Spearman correlation;
beat all three on aggregate MAE;
beat all three on top-20% lift;
beat ADP + prior-production on season-level Spearman in at least 6 of 9 seasons;
have a strictly positive 95% season-block bootstrap lower bound for Spearman advantage over ADP + prior-production;
have positive out-of-season Spearman at QB, RB, WR and TE where each has at least 30 observations;
preserve a positive external association between LWS and playoff-normalized ESPN Championship Excess;
pass identity, provenance, leakage and source-coverage hard stops.

v0.3 observed result

The governed leave-one-season-out run completed on 1,554 drafted player-seasons.

Aggregate out-of-season results:

LWI v0.3 Spearman: 0.318
ADP Spearman: 0.278
Prior-production Spearman: 0.234
ADP + prior-production Spearman: 0.284
v0.2 LWI Spearman: 0.298
LWI v0.3 MAE: 20.774 vs combo 21.282
LWI v0.3 top-20 lift: 1.805× vs combo 1.790×
LWI v0.3 top-10 lift: 2.307× vs combo 2.064×
Season Spearman wins vs combo: 6/9
Bootstrap 95% CI for Spearman advantage vs combo: +0.004 to +0.068
Metadata coverage: 96.8%

Position-level out-of-season LWI Spearman remained positive:

QB: 0.140
RB: 0.323
WR: 0.383
TE: 0.200

External construct validation remained strongly positive: among the 70 ESPN drafted-leaderboard rows, LWS v0.2 retained Spearman 0.827 with playoff-normalized Championship Excess.

Research result: PASS. v0.3 clears every preregistered research gate. This is the first League Winner forward specification to produce a strictly positive bootstrap lower bound while beating the governed baselines on aggregate Spearman, MAE and top-20 lift.

The pass is modest rather than magical. The absolute incremental signal over ADP + prior production is real but small: aggregate Spearman improves from 0.284 to 0.318, and top-20 lift from 1.790× to 1.805×. The experiment therefore supports using LWI as an incremental draft-value signal, not as a replacement for the broader Waiverlord model.

Production status after v0.3

Production Model Gate review is now authorized and required. The research pass does not itself authorize public 2026 player scores, Master Board writes, ranking movement, or downstream draft-room effects.

Until the Production Model Gate explicitly approves a production fingerprint:

no 2026 LWI is production-authoritative;
no LWI-driven Master Board movement is allowed;
no website player ranking may be presented as validated League Winner output;
the panel remains labeled EXPERIMENTAL.

The research runner's prior boolean naming public_2026_scores_authorized is non-authoritative if it reflects research-gate passage alone. Public authorization belongs only to the Production Model Gate.

2026 public output after production approval

Only after Production Model Gate approval may the public League Winner panel show:

player;
League Winner Index;
Waiverlord rank;
governed full-PPR ADP;
closest validated historical archetype;
primary LWI driver;
principal failure mode / uncertainty flag.

LWI remains a relative signal, not a championship probability.

Archived Championship Equity

Championship Equity is archived, not erased. Its failed specifications remain version-controlled scientific history. It is no longer an active website panel, feed or Method surface.

Method-first change control

The canonical public Method is https://www.waiverlord.com/league-winner/method.

A material change includes any change to LWS components, weights, data sources, cohort, replacement thresholds, normalization, drafted-player eligibility, feature definitions, architecture, baselines, validation metrics, promotion thresholds, Master Board fields, website semantics, feed behavior or publication rules.

Required order:

Update this Method first.
Deploy the branded Method page.
Only then change scoring/model code, data pipelines, Master Board fields, website feeds or presentation logic.
Run governed validation and preserve receipts.
Promote only through the Production Model Gate.
Record the observed result here and redeploy before another material specification.

Production Integration v0.3 — preregistered before implementation

Production purpose

The validated League Winner Index (LWI) v0.3 may enter production only as an auxiliary draft-value signal. It does not replace Waiverlord Model Rank, Action Rank, custom scoring, expert consensus, or simulator logic. Any future use of LWI to move core player quality or draft recommendations requires a separately preregistered blending specification and a new governed validation pass.

2026 scoring universe

A player is eligible for a 2026 LWI only when all of the following are true:

position is QB, RB, WR, or TE;
identity resolves uniquely to the governed Waiverlord player universe and nflverse metadata where required;
governed current 12-team full-PPR redraft ADP is numeric and between 1 and 300;
all required pre-season feature families are either observed or handled by the same preregistered missing-history indicators used in v0.3.

Missing redraft ADP is not imputed from Best Ball, dynasty, standard, half-PPR, superflex, popularity, or an invented market rank. An ineligible player receives no fabricated LWI and is labeled insufficient market evidence.

Frozen production model

Production scoring uses the same target and architecture that passed the v0.3 research gate:

LWS target remains 45% Draft Surplus, 40% Regular-Season Win Value, 15% Availability through Week 14;
historical training cohort remains 2016–2019 and 2021–2025, with 2020 excluded;
the ADP + prior-production ridge baseline is retained;
the residual HistGradientBoostingRegressor retains learning_rate=0.03, max_iter=220, max_depth=2, min_samples_leaf=25, l2_regularization=8.0, and random_state=42;
no production hyperparameter search, coefficient shopping, or post-2026-result tuning is permitted.

After the leave-one-season-out research specification has passed, the production estimator is fit once on all nine governed historical seasons and applied to eligible 2026 players using only information available before the 2026 season.

Public LWI scale

The internal raw model prediction is preserved for audit. The public League Winner Index is the percentile rank, from 0 to 100, of the raw v0.3 prediction among the eligible 2026 scoring cohort. League Winner Rank sorts that index descending. LWI is a relative signal, not a championship probability.

Master Board integration

The four retired Championship Equity columns are repurposed rather than expanded:

AQ → League Winner Index
AR → League Winner Rank
AS → League Winner Archetype
AT → LWI Method / Status

The first production implementation is informational only. Writing these fields must not alter Model Rank, ESPN divergence, Action Rank, projected fantasy points, custom VORP, simulation inputs, or any other core ranking field.

Historical archetype and explanation

Each scored 2026 player may be paired with a closest validated historical league-winner archetype using only normalized pre-season feature similarity within the governed historical population. The public explanation may identify the dominant LWI driver and principal uncertainty/failure mode. Historical outcome information may describe the comparison player but may not leak into the 2026 feature vector.

Production gate

Promotion requires all of the following in the same candidate package:

the saved v0.3 research artifact remains a PASS with no research blockers;
2026 eligibility, identity, ADP-format, missingness, range, uniqueness, and no-leakage checks PASS;
the candidate LWI fields are written to Sandbox Model first;
the complete Validation Suite has zero Production Model failures;
the Production Model Gate reads PASS;
the production fingerprint is recorded before Master Board publication;
only the approved LWI columns are promoted to Master Board and public website feed;
runtime receipts verify the live website is serving the same approved fingerprint.

A stale ALLOW_MODEL_PROMOTION=false flag is non-authoritative. The Production Model Gate is the sole promotion authority. Draft Operations Gate state does not certify or block this auxiliary player-level signal unless a future specification explicitly makes LWI part of simulation logic.

Market coverage integrity

The Market ADP validator must enforce the same governed ingestion contract as the production market synchronizer. The synchronizer accepts only FantasyPros Overall (PPR) consensus rows supported by at least two same-format redraft platform ADPs and agreement with their median within 12 picks; unavailable or quarantined players remain blank. Therefore production validation must test the synchronizer's preregistered hard stops — at least 180 valid same-format rows and at least 90 of the current top 100 — plus uniqueness/range integrity. It must not require 300 fabricated ADPs or silently substitute Best Ball evidence.

Publication state

After Production Model Gate approval, 2026 LWI may be displayed on the League Winner panel and Master Board as a validated production signal, while the League Winner research program remains labeled EXPERIMENTAL. Public output may include player, LWI, Waiverlord rank, governed full-PPR ADP, closest historical archetype, primary driver, and principal uncertainty flag.

Observed production result — 2026-08-21

Status: PROMOTED — PRODUCTION SIGNAL LIVE; RESEARCH PROGRAM REMAINS EXPERIMENTAL.

The frozen v0.3 production estimator was fit on the nine governed historical seasons and applied to the live 2026 Waiverlord board. The candidate was staged in Sandbox Model before publication. The complete Production Model Gate returned PASS with zero production-model failures, and all six LWI-specific hard stops passed: retained research PASS, exact eligible-score coverage, 0–100 index range, continuous unique rank, no missing-ADP fabrication, and exact production headers.

The approved production package contains 176 scored players. Players without governed 12-team full-PPR redraft ADP remain unscored rather than receiving an invented price.

Approved production fingerprint: 31a2e45122f353a4f8b73a80cfd0ed3c39e872c1efbd6a6f05992f2ee7ead47b.

The first live LWI top five are:

Brock Bowers — LWI 100.0, Waiverlord #22, governed ADP 22.0
Ja'Marr Chase — LWI 99.4, Waiverlord #5, governed ADP 3.0
CeeDee Lamb — LWI 98.9, Waiverlord #9, governed ADP 12.0
Puka Nacua — LWI 98.3, Waiverlord #2, governed ADP 4.0
Jahmyr Gibbs — LWI 97.7, Waiverlord #1, governed ADP 1.0

These values are an auxiliary relative draft-value signal. They do not alter Model Rank, Action Rank, projections, custom scoring, expert consensus, or simulator inputs under v0.3.

Iteration Notes

Specification v0.4 — League Winner Edge / Draft Leverage

Status: PREREGISTERED — not yet executed when this specification was published.

Why v0.4 exists

The v0.3 League Winner Index passed its governed research and production gates, but raw LWI is an absolute profile score. Sorting the public panel by raw LWI naturally favors already-expensive elite players. That is scientifically valid but not the most useful draft question for Waiverlord.

v0.4 asks a narrower question: which drafted players look materially more like historical league winners than their current acquisition price and established production already imply?

This specification does not change LWS v0.2 or retrain/tune the validated v0.3 architecture. It tests whether the incremental residual already produced by v0.3 can identify non-obvious league-winning value more effectively than raw LWI.

Frozen historical target

LWS remains unchanged:

45% Draft Surplus;
40% Regular-Season Win Value;
15% Availability;
Weeks 1–14 only;
governed cohort 2016–2019 and 2021–2025;
2020 excluded under the preregistered COVID-season exception.

No fantasy-playoff scoring, championship-roster prevalence, social popularity or post-season information enters the target.

Frozen v0.3 prediction

For every held-out historical season, v0.4 reuses the exact leave-one-season-out v0.3 prediction and the exact ADP + prior-production combo baseline from that fold. No v0.3 hyperparameter, feature definition, training rule or baseline is changed.

League Winner Edge

For player i in held-out season s:

League Winner Edge = LWI_v0.3_prediction - Combo_baseline_prediction

This is the model's incremental estimate of league-winner value beyond what the player's acquisition price and prior production already imply. It is not a championship probability and it does not replace Waiverlord Rank.

For public display only, the raw edge may be converted to a 0–100 percentile among eligible 2026 players after production approval. The raw residual is preserved for audit.

Draft Leverage

Draft Leverage is intended to reward a positive hidden-value signal more when the acquisition cost leaves more roster capital available.

For each held-out season:

positive_edge = max(League Winner Edge, 0)
price_flex = log1p(ADP) / log1p(300)
Draft Leverage raw = positive_edge * price_flex

The transformation is frozen before results are inspected. No optimized coefficient or post-result weighting is permitted.

Public Draft Leverage, if approved, is the 0–100 percentile rank of this raw quantity among the eligible 2026 scoring cohort.

Discovery cohort

Because v0.4 is specifically intended to identify non-obvious draft value, the primary discovery cohort is players with governed historical ADP 37 through 180 in 12-team full-PPR redraft formats. Picks 1–36 remain in secondary diagnostics but do not control the v0.4 discovery gate. Players beyond 180 are excluded from the primary gate because historical market availability and drafted-player coverage become materially less stable.

The 37 cutoff corresponds to Round 4 onward in a 12-team league and was fixed before execution.

Primary outcomes

Within the held-out discovery cohort, evaluate three rankings:

raw LWI v0.3;
League Winner Edge;
Draft Leverage.

Primary target remains continuous LWS. Secondary classification target is League Winner = top 20% LWS within season, using the already-governed LWS percentile definition.

Report:

Spearman correlation with LWS;
MAE only for raw prediction scales where meaningful;
top-20% lift;
top-10% lift;
precision among the top 10 ranked discovery candidates per season;
season-level wins versus raw LWI;
season-block bootstrap confidence interval for top-20 lift advantage versus raw LWI.

v0.4 research gate

League Winner Edge or Draft Leverage may replace raw LWI as the homepage ordering only if one of them independently clears all of the following in the primary discovery cohort:

top-20 lift exceeds raw LWI;
top-10 lift is not worse than raw LWI;
top-10-candidate precision exceeds raw LWI;
season-level top-20 lift beats raw LWI in at least 6 of 9 governed seasons;
the 95% season-block bootstrap lower bound for top-20 lift advantage versus raw LWI is strictly positive;
no identity, leakage, cohort, ADP-format or provenance hard stop fails.

Spearman is reported but does not control promotion because Edge and Draft Leverage are intentionally discovery-oriented rather than absolute player-value estimators.

Production semantics if the gate passes

If a derived metric passes, the League Winner homepage panel will become a mispriced-upside discovery surface, not a list of the best players in fantasy football.

The public panel may show:

Player;
governed ADP;
Waiverlord Rank;
LWI;
League Winner Edge;
Draft Leverage;
closest validated historical archetype;
primary driver;
principal uncertainty/failure mode.

The panel ordering will use the single derived metric that clears the preregistered gate with the strongest governed top-20 lift. Ties are broken by top-10-candidate precision, then by raw LWI. No manual player promotion is allowed.

Production boundary

Even if v0.4 passes, League Winner Edge and Draft Leverage remain auxiliary signals unless a separate future blending experiment authorizes them to move Model Rank, Action Rank, projections, expert consensus, custom scoring or simulator behavior.

The public Method must record the observed v0.4 result before any production presentation change is deployed.

League Winner Experiment v0.4 — Result

Status: FAILED — v0.3 public ordering remains authoritative

Primary discovery cohort: ADP 37–180, 1218 held-out player-seasons.

Raw LWI benchmark

Top-20 lift: 1.641×
Top-10 lift: 2.155×
Top-10-candidate precision: 0.333

League Winner Edge

Spearman vs LWS: 0.177
Top-20 lift: 1.436×
Top-10 lift: 1.303×
Top-10-candidate precision: 0.289
Season top-20 lift wins vs raw LWI: 1/9
Bootstrap 95% CI for top-20 lift advantage: -0.426 to +0.029
Gate: FAIL

Draft Leverage

Spearman vs LWS: 0.142
Top-20 lift: 1.442×
Top-10 lift: 1.185×
Top-10-candidate precision: 0.278
Season top-20 lift wins vs raw LWI: 2/9
Bootstrap 95% CI for top-20 lift advantage: -0.426 to +0.055
Gate: FAIL

Governed conclusion

Approved research ordering metric: NONE.

A research pass does not itself change the live panel. Public presentation may change only after the Method records this result and the Production Model Gate approves the production fingerprint.

Specification v0.5 — Upper-Tail Discovery Model

Status: PREREGISTERED — not yet executed when this specification was published.

Motivation from v0.4

v0.4 tested whether subtracting the ADP + prior-production baseline from the validated v0.3 LWI would reveal hidden league-winner value. It did not. Both League Winner Edge and Draft Leverage underperformed raw LWI across the preregistered discovery tests.

The governed conclusion is that league-winner discovery is not well represented by a simple residual from an absolute-value model. v0.5 therefore changes the forecast question rather than manipulating the v0.3 score.

v0.5 question

Among non-premium draft selections, can a model trained directly on the historical upper tail identify league-winning outcomes better than raw LWI?

The model is designed for discovery, not absolute player ranking.

Frozen historical outcome

LWS v0.2 remains unchanged. No outcome definition is modified.

The binary upper-tail target is:

League Winner Outcome = 1 when historical LWS percentile >= 80 within season; otherwise 0.

This is the already-governed League Winner classification. No new outcome cutoff is fitted from results.

Primary discovery cohort

The primary v0.5 cohort is governed 12-team full-PPR historical ADP 49 through 180, corresponding to Round 5 through Round 15 in a 12-team league.

The purpose of this cohort is explicit: the public discovery surface should find players available after the premium opening rounds rather than announce that established first-round stars are good at fantasy football.

Players outside this cohort remain available for secondary diagnostics but cannot control the v0.5 research gate.

Preseason features

v0.5 uses only the preregistered v0.3 preseason feature set:

governed full-PPR ADP / log ADP;
prior-season production proxy;
prior-season opportunity per game;
prior availability;
prior replacement-week share;
age on September 1;
NFL draft capital;
years since NFL draft;
rookie, second-year and third-year indicators;
two-year PPG trajectory;
two-year opportunity trajectory;
missing-history and missing-metadata indicators;
position.

No same-season regular-season result, playoff result, social attention, championship prevalence or post-draft outcome is permitted.

Frozen model architecture

For each leave-one-season-out fold, training is restricted to the v0.5 discovery cohort from the remaining eight governed seasons.

The upper-tail model is a HistGradientBoostingClassifier with frozen hyperparameters:

learning_rate = 0.03
max_iter = 220
max_depth = 2
min_samples_leaf = 25
l2_regularization = 8.0
random_state = 42

The score used for ranking is the held-out positive-class probability. It is treated as a relative upper-tail score, not a calibrated championship probability.

No hyperparameter search, threshold search, feature selection or post-result tuning is permitted.

Candidate discovery metrics

Two candidate orderings are tested:

Upper-Tail Score — direct positive-class score from the frozen classifier.
Upper-Tail Leverage — Upper-Tail Score * price_flex, where price_flex = log1p(ADP) / log1p(300).

The price-flex transformation is fixed before execution and is not optimized.

Governed benchmarks

Within the same held-out ADP 49–180 cohort, compare against:

raw LWI v0.3;
ADP alone;
ADP + prior-production combo baseline.

Primary metrics

Report for every candidate and benchmark:

top-20% lift against the governed League Winner Outcome;
top-10% lift against Elite League Winner Outcome (LWS percentile >= 90);
precision among the top 10 discovery candidates in each held-out season;
season-level top-20 lift;
Spearman correlation with continuous LWS as a secondary diagnostic;
season-block bootstrap confidence interval for top-20 lift advantage versus raw LWI.

v0.5 research gate

Upper-Tail Score or Upper-Tail Leverage may advance to Production Model Gate review only if one candidate independently satisfies all of the following:

top-20 lift exceeds raw LWI;
top-20 lift exceeds ADP and the combo baseline;
top-10 lift is not worse than raw LWI;
top-10-candidate precision exceeds raw LWI;
season-level top-20 lift beats raw LWI in at least 6 of 9 governed seasons;
the 95% season-block bootstrap lower bound for top-20 lift advantage versus raw LWI is strictly positive;
identity, provenance, leakage, cohort and ADP-format hard stops all pass.

If both candidates pass, the winner is the candidate with the greater top-20 lift; ties are broken by top-10-candidate precision, then top-10 lift.

Production boundary

A v0.5 research pass authorizes only Production Model Gate review for a discovery-oriented public League Winner panel. It does not authorize changes to Model Rank, Action Rank, projections, expert consensus, custom scoring or simulator logic.

The public Method must record the observed v0.5 result before any presentation or feed change is promoted.

League Winner Experiment v0.5 — Result

Status: FAILED — v0.3 public ordering remains authoritative

Primary discovery cohort: ADP 49–180, 1102 held-out player-seasons.

Raw LWI v0.3

Spearman vs LWS: 0.243
Top-20 lift: 1.524×
Top-10 lift: 1.409×
Top-10-candidate precision: 0.244

Upper-Tail Score

Spearman vs LWS: 0.148
Top-20 lift: 1.463×
Top-10 lift: 0.986×
Top-10-candidate precision: 0.178
Season top-20 lift wins vs raw LWI: 3/9
Bootstrap 95% CI for top-20 lift advantage: -0.461 to +0.372
Gate: FAIL

Upper-Tail Leverage

Spearman vs LWS: 0.137
Top-20 lift: 1.432×
Top-10 lift: 0.986×
Top-10-candidate precision: 0.211
Season top-20 lift wins vs raw LWI: 4/9
Bootstrap 95% CI for top-20 lift advantage: -0.476 to +0.351
Gate: FAIL

ADP baseline

Spearman vs LWS: 0.142
Top-20 lift: 1.402×
Top-10 lift: 1.127×
Top-10-candidate precision: 0.222

ADP + prior-production baseline

Spearman vs LWS: 0.157
Top-20 lift: 1.341×
Top-10 lift: 1.550×
Top-10-candidate precision: 0.222

Governed conclusion

Approved research ordering metric: NONE.

A research pass authorizes Production Model Gate review only. No live panel or ranking change is authorized by this research artifact alone.

Specification v0.6 — Opportunity Vacuum / Depth-Chart Mechanism

Status: PREREGISTERED — not yet executed when this specification was published.

Motivation from v0.4 and v0.5

Two governed discovery attempts failed:

v0.4 showed that simply subtracting the market/production baseline from raw LWI made discovery worse;
v0.5 showed that retraining the same player-history feature set directly on the upper tail still did not beat raw LWI.

The common limitation is now explicit: the model knows a player's price, career stage and prior production trajectory, but it does not directly know why a larger role could become available before the season starts.

v0.6 therefore tests a mechanism rather than another transformation: preseason opportunity vacated by teammates, current-team continuity, and earliest available target-season depth-chart position.

Historical target and discovery cohort

LWS v0.2 remains unchanged. The binary discovery target remains LWS percentile >= 80.

Primary discovery cohort remains governed 12-team full-PPR ADP 49 through 180.

New data source

v0.6 uses nflverse historical weekly player statistics plus nflverse historical depth charts. Depth-chart data are available back to 2001. Only the earliest target-season depth-chart snapshot is permitted for each team/player. Later in-season depth-chart records are prohibited.

For seasons through 2024, the legacy depth-chart schema is normalized from fields including team/club, player name, position and depth_team. For 2025+, the ESPN-derived schema is normalized from timestamp, team, player name, position abbreviation and pos_rank.

If an unambiguous earliest depth-chart identity cannot be resolved, the depth-specific fields remain missing and an explicit missing indicator is set. Missingness may not be repaired with same-season outcomes.

Opportunity-vacuum construction

For target season S, all opportunity quantities are calculated from season S-1 regular-season nflverse data.

A prior-season player is treated as returning to the same team only when that player's normalized identity appears on the target-season team's earliest depth-chart snapshot. This definition is fixed before execution.

The following new preseason features are constructed:

team_vacated_target_share — prior-team targets belonging to players not returning to that team, divided by prior-team targets;
team_vacated_carry_share — prior-team carries belonging to players not returning to that team, divided by prior-team carries;
position_vacated_opportunity_share — prior-team targets + carries at the player's fantasy position belonging to players not returning to that team, divided by prior-team position opportunities;
returning_position_competitor_share — prior-team position opportunities belonging to returning same-position teammates other than the player, divided by prior-team position opportunities;
depth_rank — player's normalized rank on the earliest target-season depth chart when available;
depth_competitor_count — count of same-position players on that team in the same earliest depth-chart snapshot;
new_team — whether the player's earliest target-season team differs from his most common prior-season team;
prior_team_opportunity_pg — prior-team targets + carries per team game, using only S-1 data;
opportunity_vacuum_missing — explicit missing indicator when team/depth identity cannot support the calculation;
depth_chart_missing — explicit missing indicator when player depth rank cannot be resolved.

All shares are clipped to [0,1]. No target-season regular-season snap, target, carry, injury outcome, fantasy result or playoff result may enter these features.

Predictor

v0.6 retains the v0.5 upper-tail HistGradientBoostingClassifier architecture and adds only the preregistered opportunity-vacuum/depth features to the v0.3 player-history features.

Frozen hyperparameters remain:

learning_rate = 0.03
max_iter = 220
max_depth = 2
min_samples_leaf = 25
l2_regularization = 8.0
random_state = 42

No hyperparameter search, feature selection, threshold search or post-result tuning is allowed.

Candidate metrics

Opportunity Upside Score — direct positive-class score from the v0.6 classifier.
Opportunity Leverage — Opportunity Upside Score * price_flex, using the unchanged v0.5 price-flex formula.

Governed benchmarks

Within each held-out ADP 49–180 cohort, compare against:

raw LWI v0.3;
v0.5 Upper-Tail Score;
ADP alone;
ADP + prior-production combo baseline.

v0.6 research gate

Opportunity Upside Score or Opportunity Leverage may advance only if one independently satisfies all of the following:

top-20 lift exceeds raw LWI v0.3;
top-20 lift exceeds v0.5 Upper-Tail Score;
top-20 lift exceeds ADP and the combo baseline;
top-10 lift is not worse than raw LWI;
top-10-candidate precision exceeds raw LWI;
season-level top-20 lift beats raw LWI in at least 6 of 9 governed seasons;
the 95% season-block bootstrap lower bound for top-20 lift advantage versus raw LWI is strictly positive;
depth-chart identity coverage is reported and no source/provenance/leakage hard stop fails.

If both pass, select the candidate with greater top-20 lift; ties are broken by top-10-candidate precision, then top-10 lift.

Production boundary

A v0.6 pass would authorize Production Model Gate review for a discovery-oriented public panel only. No core Waiverlord ranking, Action Rank, projection, expert blend, custom score or simulator input may move without a separate preregistered blending experiment.

The observed v0.6 result must be published to the Method before any live panel change.

League Winner Experiment v0.6 — Result

Status: FAILED — v0.3 public ordering remains authoritative

Primary discovery cohort: ADP 49–180, 1102 held-out player-seasons.

Depth target-team coverage: 95.5%. Depth-rank coverage: 95.5%. Complete opportunity-feature coverage: 82.3%.

Raw LWI v0.3

Spearman vs LWS: 0.243
Top-20 lift: 1.524×
Top-10 lift: 1.409×
Top-10-candidate precision: 0.244

v0.5 Upper-Tail Score

Spearman vs LWS: 0.148
Top-20 lift: 1.463×
Top-10 lift: 0.986×
Top-10-candidate precision: 0.178

Opportunity Upside Score

Spearman vs LWS: 0.155
Top-20 lift: 1.158×
Top-10 lift: 1.268×
Top-10-candidate precision: 0.222
Season top-20 lift wins vs raw LWI: 2/9
Bootstrap 95% CI for top-20 lift advantage: -0.773 to +0.000
Gate: FAIL

Opportunity Leverage

Spearman vs LWS: 0.145
Top-20 lift: 1.189×
Top-10 lift: 1.550×
Top-10-candidate precision: 0.222
Season top-20 lift wins vs raw LWI: 2/9
Bootstrap 95% CI for top-20 lift advantage: -0.710 to +0.069
Gate: FAIL

ADP baseline

Spearman vs LWS: 0.142
Top-20 lift: 1.402×
Top-10 lift: 1.127×
Top-10-candidate precision: 0.222

ADP + prior-production baseline

Spearman vs LWS: 0.157
Top-20 lift: 1.341×
Top-10 lift: 1.550×
Top-10-candidate precision: 0.222

Governed conclusion

Approved research ordering metric: NONE.

A research pass authorizes Production Model Gate review only. No live panel or ranking change is authorized by this artifact alone.

League Winner Experiment v0.7 — Championship Roster Probability

Status: PREREGISTERED — results unseen

Research question

Can preseason information identify drafted players who are more likely to appear on championship rosters than their acquisition cost alone implies?

This experiment changes the target from generic season-value resemblance to championship-roster incidence. Frozen v0.3 LWI remains unchanged and may be used only as a benchmark/supporting preseason signal.

Governed seasons

2016–2019 and 2021–2025. The 2020 season remains excluded under the existing COVID-season preregistration.

Historical target

For each player-season with directly attributable championship-roster evidence:

Observed Championship Roster Rate (OCRR): observed championship-roster prevalence from the governed source.
Expected Championship Roster Rate (ECRR): out-of-sample expectation conditional on preseason ADP and position, fit without the held-out season.
Observed Championship Edge: OCRR - ECRR.
Championship Roster Probability (CRP): forward estimate of OCRR using preseason-available information only.
Projected Championship Edge: CRP - ECRR.

Missing championship evidence is censored/unknown, never zero. Absence from a selective leaderboard may not be coded as failure or 0% championship prevalence.

Playoff-spike normalization

Championship incidence remains the endpoint, but predictive player-performance features stop before the target season begins. Weeks 15–17 of the target season are prohibited as predictive inputs and may be used only as a contamination diagnostic. A monster fantasy-playoff game cannot manufacture a preseason archetype.

Frozen preseason feature set

The challenger uses the frozen v0.3 preseason feature set only:

log_adp
adp_pos_strength
prior_ppg
prior_opp_pg
prior_ceiling
prior_volatility
prior_availability
no_prior
cost_x_ppg
cost_x_opp
age_sep1
log_nfl_draft_pick
undrafted_nfl
years_since_draft
rookie
second_year
third_year
ppg_change
opp_change
prior_replacement_week_share
history2_missing
metadata_missing
position as categorical input

No v0.6 depth-chart/opportunity-vacuum feature and no target-season realized production enters the challenger.

Frozen model architecture

The challenger is a direct HistGradientBoostingRegressor for OCRR with the same fixed hyperparameters as v0.3:

learning_rate = 0.03
max_iter = 220
max_depth = 2
min_samples_leaf = 25
l2_regularization = 8.0
random_state = 42

Numeric inputs use median imputation with explicit missing indicators. Position uses one-hot encoding with unknown categories ignored. No hyperparameter search, feature selection, threshold search, post-result calibration or post-result weighting is permitted.

Validation and frozen baselines

Leave-one-season-out validation across all nine governed seasons. Only directly observed championship rows are scored; missing rows remain censored.

Required baselines are:

ADP-only: Ridge regression on log_adp + position, alpha = 10.0.
ADP + prior production: Ridge regression on log_adp, prior_ppg, prior_availability, no_prior + position, alpha = 10.0.
Frozen v0.3 LWI map: out-of-fold v0.3 LWI predictions mapped to OCRR with Ridge regression, alpha = 10.0, fit without the held-out championship season.

ECRR is the ADP-only held-out prediction. Observed Championship Edge is OCRR minus ECRR. Each candidate's Projected Championship Edge is its held-out OCRR prediction minus ECRR.

Primary evaluation

Because championship reporting is selective/censored, evaluation occurs only where observed championship prevalence is directly attributable. Missing rows are excluded rather than treated as zero.

Report:

Spearman correlation with OCRR
MAE against OCRR
Spearman correlation between Projected Championship Edge and Observed Championship Edge
top-decile enrichment: actual top-decile Observed Championship Edge rate among each model's top-decile Projected Championship Edge selections, divided by cohort base rate
held-out season Spearman wins versus the strongest baseline
season-block bootstrap confidence interval for challenger minus strongest-baseline OCRR Spearman
source coverage and censoring diagnostics
position-level OCRR Spearman where n >= 10

Research gate

The challenger passes only if all are true:

Challenger OCRR Spearman is strictly greater than every governed baseline.
Challenger Championship Edge Spearman is strictly greater than every governed baseline's edge Spearman.
Challenger top-decile Championship Edge enrichment is strictly greater than every governed baseline.
Challenger beats the strongest aggregate baseline on held-out-season OCRR Spearman in at least 6 of 9 seasons.
The 95% season-block bootstrap lower bound for challenger minus strongest-baseline OCRR Spearman is strictly positive.
Position-level challenger OCRR Spearman is non-negative for QB, RB, WR and TE wherever n >= 10; any position with fewer than 10 observed rows is reported as insufficient rather than forced through the gate.
No leakage, fabricated zeros, invented championship rates or target-season performance inputs are present, and coverage/censoring diagnostics are populated.

Any FAIL is a hard stop. No Master Board field, production ranking, website ordering or 2026 public score changes solely because of a failed research result.

2026 outputs if and only if research + Production Model Gate pass

CRP: estimated championship-roster prevalence/profile score.
Championship Edge: CRP minus market-expected championship prevalence at current governed ADP.

Championship Leverage is deliberately not authorized here. Any acquisition-cost transformation of Championship Edge requires a separate preregistered validation.

The intended public ordering is Championship Edge rather than raw CRP because the objective is to find championship probability the draft room is underpricing.

Interpretation

A successful result does not mean a player guarantees a title. It means comparable preseason profiles historically appeared on championship rosters more often than comparable acquisition prices implied, and that incremental relationship survived held-out-season validation.

League Winner Experiment v0.7 — Championship Roster Probability Result

Status: FAILED — no production change authorized

Observed drafted championship rows: 70 across nine governed seasons. Missing leaderboard rows were censored, never coded as zero.

Aggregate held-out metrics

|---|---:|---:|---:|---:|

| crp | -0.031 | 4.694 | 0.196 | 0.000× |

| adp | -0.489 | 4.877 | 0.000 | 0.000× |

| combo | -0.355 | 5.003 | 0.008 | 0.000× |

| lwi_map | -0.534 | 4.895 | 0.124 | 1.094× |

Strongest governed baseline: combo.

Season Spearman wins vs strongest baseline: 5/9.

Season-block bootstrap 95% CI for CRP Spearman advantage: +0.042 to +0.497.

Gate

ocrr_spearman_beats_all_baselines: PASS
edge_spearman_beats_all_baselines: PASS
top_decile_edge_enrichment_beats_all_baselines: FAIL
season_wins_at_least_6: FAIL
bootstrap_lower_positive: PASS
position_direction_nonnegative: FAIL
coverage_and_censoring_populated: PASS

promotion_pass = false

Coverage

2016: 5 observed drafted championship rows
2017: 5 observed drafted championship rows
2018: 10 observed drafted championship rows
2019: 4 observed drafted championship rows
2021: 8 observed drafted championship rows
2022: 10 observed drafted championship rows
2023: 8 observed drafted championship rows
2024: 10 observed drafted championship rows
2025: 10 observed drafted championship rows

Playoff-spike diagnostic

Target-season Weeks 15-17 points vs championship prevalence Spearman: 0.078. This is diagnostic only and never enters the predictor.

League Winner Experiment v0.8 — Multi-Source Statistical Archetypes

Status: PREREGISTERED — results unseen

Research question

Can a multi-source historical championship-roster label, combined with nflverse statistical profiles, identify repeatable league-winner archetypes — and do preseason projected statistical profiles similar to those archetypes improve identification of future championship-roster players?

Why v0.8 exists

v0.7 found a positive aggregate championship-edge signal but failed the governed gate because the direct ESPN championship sample was sparse, upper-tail enrichment was weak, and positional stability failed. v0.8 attacks those weaknesses by broadening the historical label across independent public championship-roster sources and moving the mechanism test into position-specific statistical profile space.

Governed seasons

Primary historical championship-label seasons remain 2016–2019 and 2021–2025. The 2020 season remains excluded under the existing COVID-season preregistration.

The preseason-projection backtest uses only seasons for which a directly archived component-stat projection source can be authenticated no later than the calendar date of that season's opening NFL kickoff. If the public archive exposes the kickoff date but not a publication time, the season is tagged KICKOFF_DATE_TIME_UNKNOWN and retained in the primary test, with that provenance limitation reported. A guide dated after opening kickoff is rejected. Missing projection seasons are omitted, never imputed. A minimum of 5 independent non-COVID seasons is required for any projected-stat similarity conclusion.

Historical championship label

The primary label is Multi-Source Championship Consensus (MCC).

Eligible evidence is a directly attributable public platform report of player championship-roster prevalence for that season. Initial governed sources are ESPN and Yahoo Fantasy. Additional sources may be added only before execution if they expose directly attributable championship-roster data with a source URL and season.

Raw platform percentages are not averaged directly because league size, scoring defaults, roster rules, and user populations differ by platform. For each source-season:

Preserve the published championship-roster percentage exactly.
Rank eligible QB/RB/WR/TE rows within that source-season and transform them to a 0–100 percentile score.
MCC for a player-season is the arithmetic mean of available source-season percentile scores.
Each source receives equal source-level influence. A source with more published rows cannot dominate another source merely because it exposes a deeper leaderboard.
source_count and coverage class are tracked separately from MCC quality.

Missing source/player evidence is UNKNOWN/CENSORED, never zero. A player is not penalized because a selective leaderboard omitted him.

Editorial “league winner” articles that do not publish an exact championship-roster percentage may be retained as an external directional diagnostic, but they do not enter MCC v0.8.

nflverse statistical profile

For each historical player-season, nflverse regular-season data through Week 14 only supplies the realized statistical profile. Weeks 15–17 are excluded from every primary profile feature and may be used only for playoff-spike diagnostics.

Profiles are position-specific and use per-game / rate statistics so schedule length does not create artificial similarity.

QB

pass attempts/game
completions/game
passing yards/game
passing TD/game
interceptions/game
carries/game
rushing yards/game
rushing TD/game
PPR points/game

RB

carries/game
rushing yards/game
rushing TD/game
targets/game
receptions/game
receiving yards/game
receiving TD/game
target share when available
PPR points/game

WR / TE

targets/game
receptions/game
receiving yards/game
receiving TD/game
carries/game
rushing yards/game
rushing TD/game
target share when available
air-yards share when available
WOPR when available
PPR points/game

Optional advanced-share fields enter a position model only when at least 70% of the applicable training rows contain a real value. Otherwise that field is omitted for that fold. Eligible missing values are imputed to the training median and the coverage decision is recorded. No missing optional field is converted to a favorable value.

Statistical archetype score

Within each leave-one-season-out fold and position:

Fit robust location/scale parameters on the training seasons only.
Define the historical winner bank as player-seasons with MCC at or above the training-position 70th percentile.
Standardize player statistical vectors using training-only robust scaling.
For each held-out player, use the 7 nearest winner-bank neighbors when at least seven exist. If only 3–6 eligible winner-bank rows exist for that position, use all available rows and record reduced-neighbor coverage. Fewer than 3 is a hard stop for that position-fold.
Statistical Archetype Similarity (SAS) is the inverse-distance-weighted mean MCC of those neighbors, rescaled 0–100 within the held-out position-season.

The winner-bank 70th-percentile threshold, maximum k=7, minimum k=3, Euclidean distance, feature sets, coverage rule, and scaling method are frozen before results are inspected. No post-result tuning is allowed.

Historical realized-stat validation

SAS is first tested using actual Week-1-to-14 nflverse statistics. This is a construct-validation test only: it asks whether championship-roster players occupy repeatable statistical-profile neighborhoods. It does not by itself authorize a preseason 2026 prediction.

Benchmarks:

ADP-only rank
Week-1-to-14 PPR points/game alone
frozen historical v0.3 LWI where available

Primary metrics:

Spearman SAS vs MCC
top-quartile MCC enrichment among top-decile SAS
leave-one-season-out fold wins versus the strongest benchmark
position-specific direction and enrichment
season-block bootstrap CI for incremental Spearman advantage

A position-level direction gate is evaluated only where at least five held-out MCC observations exist across the governed OOF results; lower-coverage positions are reported but cannot create a false PASS.

Preseason projected-stat validation

A separate test uses archived Mike Clay / ESPN component statistical projections as statistical-only input. Mike Clay remains outside the expert-ranking ensemble.

Authenticated guide seasons currently targeted are 2019, 2021, 2022, 2023, 2024 and 2025. Each source must be a directly accessible ESPN projection guide whose displayed update date is no later than that season's opening-kickoff calendar date under the provenance rule above. If fewer than five non-COVID seasons parse and pass provenance/coverage gates, projected-stat validation hard-stops.

For each eligible projection season:

Parse projected games and position-appropriate component statistics from the archived ESPN guide.
Convert projections to the same per-game feature space used by nflverse historical profiles. Features unavailable in the projection guide are excluded from PSAS rather than backfilled from realized data.
Build the winner bank only from governed MCC seasons strictly earlier than the projection season. Future seasons can never define an earlier season’s projected archetype.
Use the same frozen maximum k=7, minimum k=3, robust-scaling and inverse-distance rules. A target position-season with fewer than three eligible prior winner-bank rows is excluded with an explicit coverage receipt rather than silently falling back.
Produce Projected Statistical Archetype Similarity (PSAS).
Evaluate PSAS against that season’s MCC.

Required comparison baselines:

ADP-only
projected PPR points alone
frozen v0.3 LWI when reproducible without leakage

2026 mapping

The 2026 Mike Clay Projection Guide may be mapped into the frozen PSAS feature space only after the historical projected-stat test completes.

A historical projected-stat PASS authorizes a Sandbox-only 2026 candidate list containing:

PSAS
nearest historical statistical comps
MCC of each comp
current governed ADP
v0.3 LWI as a separate supporting field

No 2026 PSAS, Championship Edge, Master Board field, website ordering, Action Rank change, or Model Rank change becomes production merely because the research test passes. Production still requires the complete Validation Suite and Production Model Gate.

Research gates

Gate A — realized statistical archetype

PASS only if all are true:

SAS Spearman with MCC exceeds the strongest governed benchmark.
Top-decile SAS produces greater top-quartile MCC enrichment than the strongest benchmark.
SAS wins at least 6 of 9 governed season folds where MCC coverage permits evaluation.
The 95% season-block bootstrap lower bound for incremental Spearman advantage is strictly positive.
Direction is non-negative at QB, RB, WR and TE where sample size is sufficient.

Gate B — preseason projected statistical archetype

PASS only if all are true:

At least 5 authenticated non-COVID projection seasons pass parsing/provenance coverage.
PSAS beats ADP-only and projected-PPR-only on MCC association.
PSAS improves upper-tail MCC enrichment.
PSAS wins a majority of eligible held-out projection seasons.
Season-block bootstrap lower bound for the primary incremental advantage is strictly positive.
No target-season realized statistics enter projected features and no future-season archetype labels enter an earlier projection season.

Any FAIL is a hard stop for production use of that candidate score.

Playoff contamination diagnostic

For each historical MCC player-season, calculate the share of Weeks 1–17 PPR production occurring in Weeks 15–17. Report the relationship between MCC and playoff-spike share. This is diagnostic only and never enters SAS or PSAS.

Interpretation

If both gates pass, the evidence would support a specific claim: championship-roster players tend to occupy repeatable position-specific statistical neighborhoods, and preseason projections that place current players into those neighborhoods identify championship-roster outcomes better than price or projected points alone.

League Winner Experiment v0.8 — Result

Gate A realized archetype: FAIL

Gate B preseason projected archetype: FAIL

MCC player-seasons: 130; multi-source rows: 23.

Realized-stat test

SAS Spearman vs MCC: 0.128
Strongest benchmark: ppg (0.152)
SAS upper-tail enrichment: 0.000x
Season wins: 5/9
Bootstrap 95% CI advantage: -0.183 to 0.121

Projected-stat test

Authenticated/parsed seasons: 2021, 2022, 2023, 2024, 2025
PSAS Spearman vs MCC: 0.223
ADP: 0.064
Projected PPR: 0.076
PSAS upper-tail enrichment: 0.891x
Season wins: 3/5
Bootstrap 95% CI advantage: -0.153 to 0.491

Governance

No production ranking or Master Board change is authorized by this research artifact. 2026 output, if present, is Sandbox-only.

League Winner Experiment v0.9 — Deep Championship Corpus Replication

Status: PREREGISTERED — results unseen

Research question

Does the projected-statistical-archetype signal observed in v0.8 survive when the historical championship label is materially deepened across public sources, without changing the frozen similarity architecture?

Motivation

v0.8 produced a promising preseason projected-stat signal: PSAS associated with Multi-Source Championship Consensus more strongly than ADP or projected PPR, but failed the research gate because the championship corpus remained small, upper-tail enrichment failed, and the season-block confidence interval crossed zero. v0.9 is a data-depth replication, not a model-tuning exercise.

The v0.8 PSAS architecture, position feature sets, robust scaling, winner-bank threshold, distance metric, neighbor rules, projection provenance rules, 2020 exclusion and no-leakage rules are frozen unchanged.

Governed championship corpus

The corpus is expanded before execution from directly attributable public championship-roster reports:

ESPN: retain governed historical rows for 2016–2021; replace the shallow 2022–2025 extracts with the complete publicly exposed ESPN top-50 championship-roster lists for those seasons, restricted to QB/RB/WR/TE.
Yahoo Fantasy: retain the governed exact-rate snapshot used in v0.8.
FantasyPros: add directly published 2021 championship-roster shares where the article uses the same denominator — share of championship rosters containing the player.
RealTime Fantasy Sports: add directly published 2024 championship-team ownership percentages, again only where the denominator is championship rosters.

No source with a different denominator may enter MCC. Conditional win rates such as “percentage of teams rostering Player X that won” are retained only as external diagnostics and are not mixed with share-of-champions data.

Multi-Source Championship Consensus v0.9

Raw percentages are never averaged across platforms.

For each source-season:

preserve the exact published championship-roster percentage;
rank eligible QB/RB/WR/TE rows within the source-season;
transform the rank to a 0–100 within-source-season percentile;
for each player-season, average available source percentiles with equal source-level influence;
track source count and source names separately from MCC quality.

Missing player/source evidence remains CENSORED / UNKNOWN, NEVER ZERO.

nflverse statistical profiles

Use nflverse regular-season player statistics through Week 14 only. Position-specific realized profiles and the playoff-contamination diagnostic are identical to v0.8.

Frozen SAS / PSAS architecture

No architecture changes are permitted from v0.8:

position-specific feature spaces;
optional advanced fields only when at least 70% observed in the applicable training fold;
robust training-only median / IQR scaling;
training-position winner bank = MCC at or above 70th percentile;
Euclidean distance;
maximum 7 nearest winner-bank neighbors;
use all 3–6 when fewer than 7 exist;
fewer than 3 eligible neighbors = hard-stop/excluded position-fold;
inverse-distance-weighted neighbor MCC;
no target-season realized statistics in PSAS;
historical projected winner banks contain only seasons strictly earlier than the target projection season.

Projection sources

Use the same authenticated Mike Clay / ESPN component-stat projection guides governed in v0.8. At least five non-COVID seasons must pass provenance and parsing. Same-day opening-kickoff-date guides retain the KICKOFF_DATE_TIME_UNKNOWN limitation already frozen in v0.8.

Primary tests

Gate A — realized statistical archetype replication

Compare SAS against:

ADP;
Week-1-to-14 PPR points per game;
frozen historical v0.3 LWI where available.

PASS requires all:

SAS Spearman with MCC exceeds the strongest benchmark.
Top-decile SAS has greater top-quartile MCC enrichment than the strongest benchmark.
SAS wins at least 6 of 9 governed season folds with sufficient coverage.
The 95% season-block bootstrap lower bound for incremental Spearman advantage is strictly positive.
Position-level direction is non-negative where at least five OOF MCC observations exist.

Gate B — preseason projected statistical archetype replication

Compare PSAS against:

ADP;
projected PPR points per game.

Frozen v0.3 LWI is reported as an auxiliary comparator where available but is not required as a baseline because its historical reconstruction has different coverage.

PASS requires all:

At least five authenticated non-COVID projection seasons pass.
PSAS Spearman with MCC exceeds both ADP and projected PPR.
Top-decile PSAS has greater top-quartile MCC enrichment than both ADP and projected PPR.
PSAS beats the strongest required baseline in a majority of eligible held-out projection seasons.
The 95% season-block bootstrap lower bound for the incremental Spearman advantage versus the strongest required baseline is strictly positive.
No target-season realized statistics or future-season championship labels enter projected features or winner banks.

Corpus adequacy receipt

Publish, before interpreting either gate:

source rows by source and season;
unique MCC player-seasons;
multi-source player-seasons;
position counts;
projection-season matching coverage.

A source row without a direct source URL and exact attributable rate is rejected.

2026 boundary

Only a Gate B PASS may authorize a Sandbox-only 2026 PSAS candidate map using the frozen 2026 Mike Clay statistical projections. Each candidate must include current governed ADP, current v0.3 LWI, and nearest historical MCC statistical comps.

No production Master Board, homepage ordering, Waiverlord Model Rank, Action Rank, projection, expert weight, custom scoring or simulator change is authorized by this research experiment alone. Production requires the complete Validation Suite and Production Model Gate.

Interpretation

v0.9 is designed to answer whether v0.8's projected-statistical similarity signal was a small-sample curiosity or a reproducible championship-roster signal after materially increasing the observed historical championship corpus.

League Winner Experiment v0.9 — Result

Gate A realized archetype: FAIL

Gate B preseason projected archetype: FAIL

MCC player-seasons: 236; multi-source rows: 66.

Realized-stat test

SAS Spearman vs MCC: 0.100
Strongest benchmark: ppg (0.220)
SAS upper-tail enrichment: 1.924x
Season wins: 1/9
Bootstrap 95% CI advantage: -0.226 to -0.005

Projected-stat test

Authenticated/parsed seasons: 2021, 2022, 2023, 2024, 2025
PSAS Spearman vs MCC: 0.104
ADP: 0.162
Projected PPR: 0.131
PSAS upper-tail enrichment: 0.998x
Season wins: 2/5
Bootstrap 95% CI advantage: -0.178 to 0.113

Governance

No production ranking or Master Board change is authorized by this research artifact. 2026 output, if present, is Sandbox-only.

League Winner Experiment v1.0 — Winner Mechanics / Predictable Surprise

Status: PREREGISTERED — results unseen

Research question

Which deviations between preseason expectation and Weeks 1–14 reality are associated with disproportionate championship-roster prevalence, and can those deviations themselves be predicted before the season well enough to identify championship upside beyond ADP or projected fantasy points?

This is a two-stage mechanism experiment. It does not tune another nearest-neighbor resemblance score. v0.3 remains the only live League Winner production signal unless this experiment later clears both research and Production Model gates.

Governed seasons

The primary test uses authenticated Mike Clay / ESPN component-stat projection seasons 2021–2025. 2020 remains excluded under the standing COVID-season preregistration. Historical championship evidence uses the governed v0.9 multi-source corpus. nflverse supplies realized regular-season statistics.

Championship label

Use v0.9 Multi-Source Championship Consensus (MCC): within-source-season championship-roster percentiles averaged across available sources. Missing source/player evidence remains CENSORED / UNKNOWN, NEVER ZERO. No incompatible denominator may enter MCC.

Realized-data boundary

All target-season realized statistics stop after Week 14. Weeks 15–17 are excluded from all mechanism construction and prediction targets. They may be used only for an explicitly labeled contamination diagnostic.

Stage A — Championship Surprise Vector

For every projected QB/RB/WR/TE with matching nflverse Weeks 1–14 data, calculate preseason-versus-realized residuals. Residuals are position-season standardized after raw calculation and winsorized to the 2.5th/97.5th percentiles within the applicable training population.

Frozen mechanism families:

Volume Surprise — realized minus projected per-game opportunity. QB: pass attempts plus carries. RB: carries plus targets. WR/TE: targets plus carries.
Role Surprise — change in receiving/rushing mix relative to projection. RB uses target share of projected opportunities versus realized opportunities; WR/TE use carry share as a secondary role channel; QB uses rushing-attempt share of total designed opportunities. This is a role-mix residual, not an invented team target-share projection.
Efficiency Surprise — position-appropriate efficiency residuals: yards per carry, yards per target, touchdown rate per opportunity, and for QBs yards/attempt plus TD rate minus interception-rate change. Undefined denominators remain missing rather than fabricated.
Availability Surprise — Weeks 1–14 games played minus preseason expected games through Week 14, where expected preplayoff games = projected games × 14/17.
Fantasy Output Surprise — realized Weeks 1–14 PPR/game minus projected PPR/game.
Price Surprise — realized Weeks 1–14 PPR/game residual versus the held-out-season-safe historical relationship between governed preseason ADP, position, and realized PPR/game. Missing ADP remains missing and does not imply poor quality.

Stage A uses a fixed ridge linear model (lambda = 10) on standardized mechanism components to estimate MCC. Coefficients are fit only on training seasons for each held-out season. No post-result coefficient selection is permitted.

Stage A is explanatory. A strong Stage A result alone cannot authorize a 2026 score.

Stage B — Predictable Surprise

For every target season, predict each Stage-A mechanism component using preseason-only inputs. Training rows may include projected players without championship labels because the mechanism targets come from projections + nflverse, not MCC.

Frozen preseason feature set:

position one-hot indicators;
governed ADP and log ADP when available;
projected games and projected PPR/game;
projected per-game carries, targets, rushing yards, receiving yards, passing attempts/yards/TDs and interceptions as position applicable;
prior-season nflverse Weeks 1–14 PPR/game, games, carries/game, targets/game, rushing yards/game, receiving yards/game, pass attempts/game and passing yards/game;
projection-minus-prior deltas for PPR/game, carries/game, targets/game and pass attempts/game where applicable;
frozen historical v0.3 LWI only as an auxiliary preseason feature when an out-of-fold historical value exists.

Missing numeric preseason fields are training-median imputed; missingness indicators are included. All scaling and imputation are fit on training seasons only.

Each mechanism component is predicted with a fixed ridge model (lambda = 10). The predicted mechanism vector is then mapped to championship upside using the Stage-A ridge coefficients fit without the held-out season. This yields Championship Upside Score (CUS).

Validation

Use leave-one-season-out validation across the five authenticated projection seasons 2021–2025. No target-season realized statistic may enter the target-season CUS calculation.

Required baselines:

governed ADP score (-ADP);
projected PPR/game;
frozen v0.3 LWI where historical OOF coverage allows, reported as auxiliary comparator.

Primary evaluation is against MCC on directly observed championship player-seasons.

Report:

CUS Spearman with MCC;
ADP and projected-PPR Spearman with MCC;
top-decile CUS enrichment for top-quartile MCC;
held-out season wins versus the strongest required baseline;
season-block bootstrap 95% CI for CUS Spearman advantage over the strongest required baseline;
position-level direction where at least five MCC observations exist;
Stage-A OOF explanatory Spearman and fixed coefficient summaries;
mechanism predictability correlations for Volume, Role, Efficiency, Availability, Output and Price Surprise;
coverage and missingness diagnostics.

Research gate

CUS passes only if all are true:

CUS Spearman with MCC exceeds both ADP and projected PPR.
Top-decile CUS has greater top-quartile MCC enrichment than both required baselines.
CUS beats the strongest required baseline in at least 4 of 5 eligible held-out seasons.
The season-block bootstrap 95% lower bound for incremental Spearman advantage is strictly positive.
Position-level CUS direction is non-negative wherever at least five MCC observations exist.
At least two of the six mechanism families have positive out-of-fold predictability from preseason-only inputs.
No target-season realized leakage, playoff leakage, fabricated zero championship rates, or post-result tuning is detected.

Any FAIL is a hard stop.

2026 boundary

Only a full research-gate PASS may generate a Sandbox-only 2026 CUS map. The 2026 map must use authenticated 2026 Mike Clay / ESPN component projections plus preseason-only features and must include current governed ADP, current v0.3 LWI, predicted primary mechanism, and uncertainty/coverage flags.

No Master Board, homepage ordering, Waiverlord Model Rank, Action Rank, projections, expert weights, custom scoring, simulator behavior, or production package may change from v1.0 alone. Production still requires the complete Validation Suite and Production Model Gate.

Interpretation

The experiment is successful only if it can identify, before the season, players whose circumstances make historically championship-relevant positive surprises more likely. It is not enough to rediscover already-expensive stars, to explain winners after the fact, or to fit random touchdown variance.

League Winner Experiment v1.0 — Result

Winner Mechanics / Predictable Surprise Gate: FAIL

MCC observations tested: 192
CUS Spearman vs MCC: 0.009
ADP: 0.154
Projected PPR: 0.120
CUS upper-tail enrichment: 1.371x
Season wins vs strongest baseline: 3/5
Bootstrap 95% CI advantage: -0.309 to 0.083

Mechanism predictability

Volume: 0.432 mean held-out Spearman
Role: 0.152 mean held-out Spearman
Efficiency: 0.120 mean held-out Spearman
Availability: 0.503 mean held-out Spearman
Output: 0.373 mean held-out Spearman
Price: 0.190 mean held-out Spearman

Governance

No production change is authorized by this research artifact. A 2026 map exists only if every preregistered research gate passed.

League Winner Experiment v1.1 — Championship Breakout State

Status: PREREGISTERED — results unseen

Research question

Can preseason information identify players likely to cross a specific nonlinear combination of regular-season surprise thresholds that is disproportionately associated with top-quartile championship-roster prevalence?

v1.0 showed that several individual surprise mechanisms are predictably related to preseason information, but a linear weighted sum of those predicted surprises did not predict championship-roster prevalence. v1.1 therefore tests a frozen interaction/threshold hypothesis rather than another continuous additive score.

Governed data

Primary seasons are the authenticated Mike Clay / ESPN component-projection seasons 2021–2025. The governed v0.9 multi-source championship corpus supplies Multi-Source Championship Consensus (MCC). nflverse supplies realized regular-season player statistics. The standing 2020 COVID exclusion remains in force.

Missing championship evidence remains CENSORED / UNKNOWN, NEVER ZERO. Realized target-season statistics stop after Week 14. Weeks 15–17 are prohibited from model inputs and state construction.

Frozen realized breakout state

For each leave-one-season-out fold, build the v1.0 raw surprise mechanisms using training-season-only transformations and thresholds. Within each position, using training seasons only, define:

High Output Surprise: Fantasy Output Surprise at or above the training-position 75th percentile.
High Price Surprise: Price Surprise at or above the training-position 75th percentile.
Sustained Availability: Availability Surprise at or above the training-position 50th percentile.

A player-season is in the Championship Breakout State (CBS) only when all three conditions are true.

Thresholds are recalculated from training seasons only for each held-out fold. No threshold may be moved after results are inspected.

Why this state

The hypothesis is explicitly conjunctive: league-winning seasons are created when a player materially beats fantasy-output expectation, materially beats the value normally available at his acquisition cost, and remains available long enough for the advantage to persist. This tests an AND state, not a weighted sum.

Stage A — State validity

Before evaluating preseason prediction, test whether realized CBS is actually associated with championship outcomes. On directly observed MCC rows, report:

top-quartile MCC rate among CBS rows;
pooled enrichment versus the observed MCC base rate;
season-level enrichment;
position-level enrichment where sample size permits.

If the realized state itself does not enrich championship outcomes, the forward classifier cannot pass regardless of predictive accuracy.

Stage B — Preseason prediction

Predict CBS using preseason-only features frozen from v1.0:

position;
governed ADP and log ADP;
projected games and projected PPR/game;
projected carries, targets, rushing yards, receiving yards, passing attempts, passing yards, passing TDs and interceptions as applicable;
prior-season nflverse Weeks 1–14 games, PPR/game and opportunity/yardage features;
projection-minus-prior deltas for PPR/game, carries/game, targets/game and pass attempts/game;
frozen historical v0.3 LWI where historical out-of-fold coverage exists;
missingness indicators.

The classifier is a fixed RandomForestClassifier with:

n_estimators = 500
max_depth = 4
min_samples_leaf = 20
max_features = "sqrt"
class_weight = "balanced_subsample"
random_state = 111
n_jobs = -1

No hyperparameter search is allowed.

The classifier output is Championship Breakout Probability (CBP).

Validation

Use leave-one-season-out validation across 2021–2025. Required baselines are:

governed ADP score (-ADP);
projected PPR/game;
frozen v0.3 LWI as an auxiliary comparator where available.

The primary championship label is top-quartile MCC within season among directly observed MCC rows. The continuous MCC value remains a secondary diagnostic.

Report:

CBP AUROC for top-quartile MCC;
CBP average precision;
top-decile CBP enrichment for top-quartile MCC;
ADP and projected-PPR AUROC and enrichment;
CBP Spearman with continuous MCC as a secondary diagnostic;
held-out season wins versus the strongest required baseline by AUROC;
season-block bootstrap 95% CI for pooled AUROC advantage;
position-level AUROC where both classes and sufficient sample exist;
predictability of each CBS component event from preseason-only inputs;
coverage and class-balance diagnostics.

Research gate

v1.1 passes only if all are true:

Realized CBS has pooled top-quartile-MCC enrichment of at least 1.50x.
CBP pooled AUROC exceeds both ADP and projected PPR.
Top-decile CBP enrichment exceeds both ADP and projected PPR enrichment.
CBP beats the strongest required baseline by AUROC in at least 4 of 5 held-out seasons.
The season-block bootstrap 95% lower bound for incremental AUROC advantage over the strongest required baseline is strictly positive.
Position-level CBP AUROC is at least 0.50 wherever at least 20 directly observed MCC rows and both outcome classes exist.
At least two of the three component events (High Output, High Price, Sustained Availability) achieve mean held-out AUROC above 0.55 when separately predicted from the same preseason feature set.
No target-season realized leakage, playoff leakage, fabricated zero championship rates, or post-result threshold/model tuning is detected.

Any FAIL is a hard stop.

2026 boundary

Only a full research-gate PASS may create a Sandbox-only 2026 CBP map using authenticated 2026 Mike Clay / ESPN component projections, governed current ADP, prior-year nflverse data and frozen v0.3 LWI. A 2026 row must include CBP, current ADP, current v0.3 LWI, predicted breakout-state probability and coverage flags.

No Master Board, homepage ordering, Waiverlord Model Rank, Action Rank, projections, expert weights, custom scoring, simulator behavior or production package may change from v1.1 alone. Production requires the complete Validation Suite and Production Model Gate.

Interpretation

This experiment is not asking whether a player is good. It asks whether preseason information can identify the specific kind of cheap, persistent, expectation-breaking outcome that historically appears disproportionately often on championship rosters.

League Winner Experiment v1.1 — Result

Championship Breakout State Gate: FAIL

MCC observations tested: 192
Realized CBS enrichment: 1.106x
CBP AUROC: 0.453
ADP AUROC: 0.623
Projected PPR AUROC: 0.600
CBP top-decile enrichment: 0.392x
Season wins vs strongest baseline: 1/5
Bootstrap 95% CI AUROC advantage: -0.256 to -0.073

Component-event predictability

High Output: 0.598 mean held-out AUROC
High Price: 0.576 mean held-out AUROC
Sustained Availability: 0.730 mean held-out AUROC

Governance

No production change is authorized by this research artifact. A 2026 CBP map exists only if every preregistered gate passed.

League Winner Experiment v1.2 — Ingredient Recipe

Status: PREREGISTERED — results unseen

Research question

Can preseason forecasts of specific surprise events be combined nonlinearly to predict championship-roster outcomes better than draft price and projection alone?

v1.1 established that three component events are individually forecastable from preseason information — High Output Surprise, High Price Surprise, and Sustained Availability — but a hand-written all-three threshold state was not championship-enriched and its direct classifier failed. v1.2 therefore removes the hand-written AND recipe. It predicts the ingredients first, then lets historical championship incidence learn a fixed, low-complexity nonlinear recipe.

Governed data

Primary seasons are authenticated Mike Clay / ESPN component-projection seasons 2021–2025. The governed v0.9 multi-source championship corpus supplies Multi-Source Championship Consensus (MCC). nflverse supplies realized Weeks 1–14 player statistics. The standing 2020 COVID exclusion remains in force.

Missing championship evidence remains CENSORED / UNKNOWN, NEVER ZERO. Weeks 15–17 are prohibited from predictors, ingredient labels, and championship-mechanism construction.

Ingredient labels

Use the three v1.1 component events exactly as frozen, with thresholds learned inside the applicable training fold and position:

High Output Surprise: output-surprise z-score at or above the training-position 75th percentile.
High Price Surprise: price-surprise z-score at or above the training-position 75th percentile.
Sustained Availability: availability-surprise z-score at or above the training-position 50th percentile.

No v1.1 threshold may be changed after seeing v1.2 results.

Stage 1 — Ingredient forecasting

For each outer held-out season, fit the same frozen v1.1 RandomForest ingredient classifiers using preseason-only features:

n_estimators = 500
max_depth = 4
min_samples_leaf = 20
max_features = "sqrt"
class_weight = "balanced_subsample"
random_state = 111
n_jobs = -1

The preseason feature set is unchanged from v1.1: governed ADP/log ADP, position, Mike Clay projected games and component statistics, prior-season nflverse Weeks 1–14 features, projection-minus-prior deltas, frozen historical v0.3 LWI where available, and missingness indicators.

The three outputs are:

P(Output Surprise)
P(Price Surprise)
P(Sustained Availability)

Mandatory nested cross-fitting

The championship recipe may not train on in-sample ingredient probabilities. Inside each outer training set, ingredient probabilities used to fit the recipe must be generated by leave-one-training-season-out cross-fitting. Only after those OOF ingredient probabilities exist may the recipe be fit. The three ingredient models are then refit on the full outer-training set to score the outer held-out season.

This is a hard anti-leakage requirement.

Stage 2 — Frozen championship recipe

The primary championship label is top-quartile MCC within season among directly observed MCC rows.

Fit a fixed L2-regularized logistic regression to the following recipe features only:

p_output
p_price
p_availability
p_output * p_price
p_output * p_availability
p_price * p_availability
p_output * p_price * p_availability
governed -log1p(ADP)
projected PPR/game

Recipe model is frozen as:

LogisticRegression
penalty = "l2"
C = 1.0
solver = "liblinear"
class_weight = "balanced"
max_iter = 2000
random_state = 112

Numeric recipe features are training-median imputed and training-standardized inside each outer fold. No hyperparameter search, feature selection, threshold search, or post-result coefficient editing is allowed.

The output is Championship Recipe Probability (CRP-v1.2).

Required baselines

Evaluate against all of:

governed ADP score (-ADP);
projected PPR/game;
frozen v0.3 LWI where available as auxiliary comparator;
a Market-Only Logistic Baseline using only governed -log1p(ADP) and projected PPR/game, with the same preprocessing and logistic-regression settings.

The strongest required baseline is the best pooled AUROC among ADP, projected PPR, and Market-Only Logistic.

Validation

Use leave-one-season-out outer validation across 2021–2025.

Report:

CRP-v1.2 pooled AUROC for top-quartile MCC;
average precision;
top-decile enrichment for top-quartile MCC;
continuous-MCC Spearman as a secondary diagnostic;
all baseline metrics;
held-out-season AUROC wins versus the strongest required baseline;
season-block bootstrap 95% CI for pooled AUROC advantage;
position-level AUROC where both classes and sufficient sample exist;
ingredient-event held-out AUROC;
recipe coefficients by held-out fold;
coverage, class balance, and nested-cross-fit receipts.

Research gate

v1.2 passes only if all are true:

CRP-v1.2 pooled AUROC exceeds the strongest required baseline.
CRP-v1.2 average precision exceeds the strongest required baseline.
Top-decile CRP-v1.2 enrichment exceeds the strongest required baseline.
CRP-v1.2 beats the strongest required baseline by AUROC in at least 4 of 5 held-out seasons.
The season-block bootstrap 95% lower bound for incremental AUROC advantage is strictly positive.
Position-level AUROC is at least 0.50 wherever at least 20 directly observed MCC rows and both classes exist.
At least two of three ingredient events retain mean held-out AUROC above 0.55.
Every outer fold has valid nested cross-fitted ingredient probabilities for recipe training; no in-sample ingredient prediction is allowed.
No target-season realized leakage, playoff leakage, fabricated zero championship rates, post-result tuning, or incompatible championship denominator mixing is detected.

Any FAIL is a hard stop.

2026 boundary

Only a full research-gate PASS may generate a Sandbox-only 2026 CRP-v1.2 map. A 2026 row must include CRP-v1.2, the three ingredient probabilities, governed ADP, projected PPR/game, frozen v0.3 LWI, and coverage flags.

No Master Board, homepage ordering, Waiverlord Model Rank, Action Rank, projections, expert weights, custom scoring, simulator behavior, or production package may change from v1.2 alone. Production still requires the complete Validation Suite and Production Model Gate.

Interpretation

This experiment asks whether championship outcomes are better understood as a learned interaction among predictable forms of surprise rather than a hand-written league-winner archetype. A successful result must beat the market out of sample, not merely tell a convincing retrospective story.

League Winner Experiment v1.2 — Result

Ingredient Recipe Gate: FAIL

MCC observations tested: 192
Recipe AUROC: 0.581
Strongest baseline (adp) AUROC: 0.623
Recipe average precision: 0.337
Recipe top-decile enrichment: 1.763x
Season wins: 1/5
Bootstrap 95% CI AUROC advantage: -0.069 to 0.002

Ingredient predictability

High Output: 0.598
High Price: 0.576
Sustained Availability: 0.730

Governance

No production change is authorized by this research artifact. A 2026 map exists only after a full research-gate PASS and remains Sandbox-only until the Production Model Gate approves promotion.

League Winner Experiment v1.3 — Source Uncertainty / Label Robustness Audit

Status: PREREGISTERED — results unseen

Research question

Did v1.2 fail because the predictor is weak, or because the championship-roster label is too sparse / source-sensitive to support a stable conclusion?

v1.3 does not change the v1.2 predictor, ingredient models, thresholds, hyperparameters, preseason features, nflverse realized-stat construction, or out-of-fold predictions. It freezes the v1.2 OOF predictions and perturbs only the championship evidence layer.

Governed championship evidence

Use the governed row-level championship corpus already accepted under v0.9: ESPN, Yahoo, FantasyPros and RTSports direct championship-roster prevalence rows with attributable URLs. Missing player/source evidence remains CENSORED / UNKNOWN, NEVER ZERO.

Additional public sources may be admitted only if they report a directly attributable championship-roster prevalence or an equivalent platform-level championship incidence. The following are explicitly rejected as label evidence: finalist-roster rates, playoff-roster rates, generic win percentage, qualitative 'league winner' lists without a denominator, DFS/contest results, anecdotal league rosters, and single-league outcomes.

No source is admitted merely because it agrees with the model.

Frozen v1.2 predictions

The primary predictor is the exact v1.2 out-of-fold recipe_prob already published. Required comparators remain:

governed ADP score;
projected PPR/game;
v1.2 market-only logistic probability;
frozen v0.3 LWI where available.

No v1.2 model refit is permitted in the primary audit.

Source-resampling audit

For each of 5,000 bootstrap replicates:

Within each season, resample available championship sources with replacement.
For each source-season, convert directly observed championship-roster percentages to within-source-season percentiles exactly as in v0.9.
For a player-season, average only the resampled source percentiles for which that player has directly observed evidence.
Missing source/player rows remain absent; they are never set to zero.
Define top-quartile championship outcome within season from the replicate MCC.
Recalculate AUROC, average precision and top-decile enrichment for v1.2 Recipe, ADP, projected PPR and market-only logistic.

This produces a distribution of label-sensitive performance rather than pretending the observed source mix is exact ground truth.

Leave-one-source-out audit

Rebuild MCC four times, excluding ESPN, Yahoo, FantasyPros and RTSports one at a time where the remaining corpus still supports evaluation. Recalculate the same metrics on frozen v1.2 predictions.

Source-depth audit

Report results separately for:

MULTI_SOURCE player-seasons;
SINGLE_SOURCE player-seasons;
seasons with at least two admitted championship sources;
source-season top-k depth bands where identifiable.

This is diagnostic only; no subgroup may be selected after inspection as the new production population.

Decision rule

Label uncertainty is considered a plausible primary explanation for v1.2 failure only if all are true:

In at least 60% of source-bootstrap replicates, Recipe AUROC exceeds the strongest required market baseline.
The median bootstrap Recipe-minus-baseline AUROC advantage is positive.
Recipe beats the strongest baseline in at least half of eligible leave-one-source-out audits.
Recipe direction is not materially worse in MULTI_SOURCE observations than in SINGLE_SOURCE observations.
No rejected source semantics, fabricated zeros, playoff/finalist substitution, or post-result source selection enters the audit.

If these conditions fail, the experiment concludes that label uncertainty is not sufficient to explain v1.2's underperformance and future work should stop blaming the evidence layer without new data.

Production boundary

v1.3 is a diagnostic audit only. It cannot directly authorize a 2026 score, Master Board field, website ranking, Action Rank change, projection change, or model promotion. A positive result may justify a separately preregistered label-improvement replication. A negative result sends research back to the predictor/target formulation.

Execution note

This non-substantive line exists only to trigger the governed workflow after the workflow file was installed. It changes no hypothesis, threshold, model, source rule, or decision criterion.

League Winner Experiment v1.3 — Result

Label uncertainty as primary explanation: NOT SUFFICIENT

Frozen v1.2 Recipe AUROC on observed label mix: 0.581
Strongest observed market baseline (adp) AUROC: 0.623
Source-bootstrap Recipe win rate: 14.2%
Source-bootstrap median AUROC delta: -0.029
Source-bootstrap 95% delta interval: -0.057 to 0.016
Leave-one-source-out wins: 0/4

Interpretation

This audit freezes the v1.2 predictor and perturbs only the admitted championship evidence. If the decision rule fails, source uncertainty is not sufficient to explain v1.2 underperformance and future work should move back to predictor/target formulation unless genuinely new championship data is acquired.

Governance

Diagnostic only. No production or 2026 map changes are authorized.

League Winner Experiment v1.4 — Championship Contribution Above Replacement

Status: PREREGISTERED — results unseen

Research question

Can preseason information identify players whose championship-roster prevalence materially exceeds what was historically expected from players drafted at comparable prices?

v1.4 changes the target rather than adding another feature stack. The experiment does not ask who appears most often on championship rosters. It asks who contributes unusually high championship prevalence relative to acquisition cost.

Governed seasons and evidence

Primary validation seasons are 2021–2025, using the governed v0.9/v1.3 multi-source championship corpus and the same source-normalization rules. The standing 2020 COVID exclusion remains in force.

Missing player/source evidence remains CENSORED / UNKNOWN, NEVER ZERO. No playoff-roster rate, finalist-roster rate, qualitative league-winner list without a denominator, DFS result, anecdotal roster, or single-league outcome may enter the target.

Historical target: Championship Contribution Above Replacement

For each leave-one-season-out fold:

Use training seasons only to estimate expected Multi-Source Championship Consensus (Expected MCC) as a function of governed preseason ADP and position.
Fit the expected-value curve with a fixed HistGradientBoostingRegressor using only:

log1p(ADP);
position one-hot indicators.

The expected-value model is trained only on directly observed championship player-seasons from training seasons.
For each held-out player-season with directly observed championship evidence, define:

CCAR = Observed MCC - Expected MCC at that acquisition cost and position

Positive CCAR means the player appeared on championship rosters more often than comparable draft-cost peers would have implied. Negative CCAR means less often.

Because MCC is a within-source normalized prevalence construct, CCAR is an index of marginal championship prevalence, not a causal percentage-point increase in title probability.

Primary outcome

The primary continuous outcome is held-out CCAR.

A secondary binary outcome, High CCAR, is defined within each held-out season as CCAR at or above that season's 75th percentile. The percentile is calculated only after the held-out Expected MCC values are produced; it is not tuned to improve results.

Preseason predictor

Use only preseason information already available in governed prior experiments. The challenger is a fixed RandomForestRegressor predicting continuous CCAR with:

n_estimators = 750
max_depth = 4
min_samples_leaf = 15
max_features = "sqrt"
random_state = 114
n_jobs = -1

Frozen features:

governed ADP and log1p(ADP);
projected PPR/game;
frozen v0.3 historical out-of-fold LWI where available;
v1.2 frozen out-of-fold probabilities of High Output Surprise, High Price Surprise and Sustained Availability;
v1.2 frozen interaction terms among those ingredient probabilities;
position one-hot indicators;
missingness indicators.

No target-season realized production, Weeks 15–17 data, championship labels from the held-out season, or post-result feature selection may enter.

The challenger output is Projected CCAR.

Required comparators

Compare Projected CCAR with:

governed ADP score (-ADP);
projected PPR/game;
frozen v0.3 LWI;
frozen v1.2 Recipe probability.

Validation

Use leave-one-season-out validation across 2021–2025.

Report:

Spearman correlation between Projected CCAR and observed CCAR;
MAE for continuous CCAR;
AUROC for High CCAR;
average precision for High CCAR;
top-decile enrichment for High CCAR;
held-out season wins versus the strongest comparator on CCAR Spearman;
season-block bootstrap 95% CI for incremental CCAR Spearman advantage versus the strongest comparator;
position-level CCAR Spearman where sample size is sufficient;
target diagnostics showing the relationship of raw MCC and CCAR with ADP;
source-coverage diagnostics.

Research gate

v1.4 passes only if all are true:

Projected CCAR has higher pooled Spearman with observed CCAR than every required comparator.
Projected CCAR beats the strongest comparator in at least 4 of 5 held-out seasons by CCAR Spearman.
The season-block bootstrap 95% lower bound for incremental CCAR Spearman advantage is strictly positive.
Projected CCAR AUROC for High CCAR exceeds every required comparator.
Projected CCAR top-decile enrichment for High CCAR exceeds every required comparator.
Position-level direction is non-negative wherever at least 20 directly observed rows exist.
Raw MCC remains materially associated with acquisition cost while CCAR is substantially less associated with acquisition cost, demonstrating that the residualization is doing its intended job.
No leakage, fabricated zeros, post-result tuning, or target contamination is detected.

Any FAIL is a hard stop.

2026 boundary

Only a full research-gate PASS may create a Sandbox-only 2026 Projected CCAR map using current governed ADP, authenticated 2026 Mike Clay / ESPN component projections, frozen v0.3 LWI, and the frozen v1.2 ingredient-prediction machinery.

No Master Board, homepage ordering, Waiverlord Model Rank, Action Rank, projections, expert weights, simulator behavior, custom scoring, or production package may change from v1.4 alone. Production requires the complete Validation Suite and Production Model Gate.

Interpretation

A successful v1.4 result would mean Waiverlord can identify preseason players whose historical profiles were associated with more championship-roster prevalence than their draft prices normally purchased. It would not establish causal title probability, but it would directly test the draft-value version of the league-winner question.

League Winner Experiment v1.4 — Result

Championship Contribution Above Replacement Gate: FAIL

Player-seasons tested: 173
Projected CCAR Spearman: -0.197
Strongest comparator (projected_ppr) Spearman: 0.083
Projected CCAR High-CCAR AUROC: 0.387
Projected CCAR top-decile enrichment: 0.214x
Season wins: 1/5
Bootstrap 95% CI Spearman advantage: -0.454 to -0.032
Raw MCC vs price Spearman: 0.154
CCAR vs price Spearman: 0.051

Governance

No production change is authorized by this research artifact. A 2026 map requires a full research-gate PASS and a separately governed Sandbox/Production Model Gate path.

League Winner Experiment v1.5 — Right-Tail Precursor Families

Status: PREREGISTERED — results unseen

Research question

Can preseason features that describe role elasticity and high-value opportunity shape predict Championship Contribution Above Replacement (CCAR) better than the median-projection / market feature set that failed in v1.4?

v1.4 established that CCAR substantially reduced the raw relationship between championship prevalence and acquisition cost, but the existing preseason predictor was directionally wrong. v1.5 therefore keeps the CCAR target frozen and changes only the predictor families.

Governed seasons and target

Primary validation seasons are 2021–2025. The v1.4 CCAR construction is reused without modification:

CCAR = Observed MCC - Expected MCC at comparable preseason ADP and position

Expected MCC is fit from training seasons only. Missing championship evidence remains CENSORED / UNKNOWN, NEVER ZERO. The standing 2020 COVID exclusion remains in force.

No target-season realized data may enter a preseason predictor.

Frozen core feature family

The Core model contains only information already established as available before each season:

governed preseason ADP and log1p(ADP);
authenticated Mike Clay / ESPN projected PPR per game;
frozen historical out-of-fold v0.3 LWI where available;
position indicators;
missingness indicators.

Core is the internal baseline for incremental family tests.

Family A — Role Elasticity Proxies

This family asks whether the preseason projection implies a player is positioned for a materially different workload than his prior regular-season usage.

Features are built from authenticated preseason projections and the immediately prior season's nflverse Weeks 1–14 statistics only:

projected minus prior PPR/game;
projected minus prior carries/game;
projected minus prior targets/game;
projected minus prior pass attempts/game;
relative versions of those deltas, scaled by abs(prior) + 1;
projected opportunity/game (pass attempts + carries for QB; carries + targets for RB/WR/TE);
prior opportunity/game on the same positional definition;
absolute and relative projected opportunity growth.

These are role-elasticity proxies, not claims about an injury or depth-chart event. No post-kickoff role data may enter.

Family B — High-Value Opportunity Shape

Historical component projections do not consistently expose red-zone or two-minute usage, so v1.5 tests the strongest reproducible preseason proxies available across all governed seasons:

projected PPR per projected opportunity;
projected yards per projected opportunity;
RB receiving-opportunity ratio (targets / (carries + targets));
QB rushing-opportunity ratio (carries / (pass attempts + carries));
projected targets/game;
projected carries/game;
projected pass attempts/game;
projected games / availability expectation.

These variables test whether the shape of projected opportunity, not only projected fantasy points, adds CCAR signal.

Families explicitly deferred

Two proposed predictor families are not authorized in v1.5 because the current governed historical store does not yet contain reproducible pre-kickoff snapshots for all validation seasons:

cross-source projection/rank dispersion;
market uncertainty / ADP volatility across time and platforms.

They may not be fabricated from current-season data or reconstructed after seeing v1.5 results. They require a separately sourced and preregistered experiment.

Models

Use the same fixed model class for every family comparison:

RandomForestRegressor(n_estimators=750, max_depth=4, min_samples_leaf=15, max_features="sqrt", random_state=115, n_jobs=-1)

No hyperparameter search is permitted.

Four models are fit inside each leave-one-season-out fold:

Core
Core + Role Elasticity
Core + High-Value Opportunity
Core + Both Families

All imputation is fit on training seasons only.

Validation

Use leave-one-season-out validation across 2021–2025. Report for every model:

pooled Spearman with held-out CCAR;
MAE;
AUROC for High CCAR (held-out season top quartile);
top-decile enrichment for High CCAR;
season-by-season CCAR Spearman;
position-level CCAR Spearman where sample size permits.

For each challenger, report season-block bootstrap 95% CI for incremental Spearman versus Core.

Required external comparators remain governed ADP, projected PPR/game, frozen v0.3 LWI, and frozen v1.2 Recipe probability.

Research gate

The Core + Both Families model passes only if all are true:

pooled CCAR Spearman exceeds Core and every required external comparator;
it beats Core in at least 4 of 5 held-out seasons;
the season-block bootstrap 95% lower bound for Spearman advantage versus Core is strictly positive;
High-CCAR AUROC exceeds Core and every required external comparator;
top-decile High-CCAR enrichment exceeds Core and every required external comparator;
position-level CCAR direction is non-negative wherever at least 20 observed rows exist;
at least one individual family model (Core + Role or Core + High-Value Opportunity) improves pooled CCAR Spearman over Core, so any combined gain is not purely an opaque interaction artifact;
no target-season realized leakage, playoff leakage, fabricated zero championship rates, current-season backfill, or post-result tuning is detected.

Any FAIL is a hard stop.

2026 boundary

Only a full research-gate PASS may authorize a Sandbox-only 2026 right-tail / CCAR map. No Master Board, homepage ordering, Waiverlord Model Rank, Action Rank, projections, expert weights, simulator behavior, custom scoring, or production package may change from v1.5 alone.

Interpretation

A successful result would show that the market misses something measurable about how a player's role can expand and how valuable his projected opportunities are, beyond median projection and draft price. A failure means these reproducible proxies are still insufficient and the next research priority becomes acquiring genuine historical dispersion and market-volatility data rather than adding more variants of the same projection inputs.