Definitions, formulas, validation gates, calibration results, caveats, and null findings are public-facing by design.
Section VI / Methods / Assumptions, validation, change history
Every number has a paper trail.
Risk archetypes, validation examples, formulas, and version notes. Methodology cards summarize what is live, what validated, and where coverage is intentionally limited.
Method audit 2026-04
Calibration harness available
Lane health 18 rows
Scheduler 1 failed / 9 stale
Raw data licenses, credentials, scheduler internals, model artifacts, and exact feature-pipeline code paths stay out of public copy.
If a surface is experimental, missing coverage, descriptive-only, or parked after validation failure, the page should say that directly.
Three Lanes Beta / current market coverage
Post-beta coverage
What each market can honestly show today.
Validated non-HR challengers remain unavailable until their real production inference outputs are wired and logged. Full production status still requires forward evaluation and direct posted-line calibration.
| Market | Palantir | Fangorn | Valinor | Triple Confirmed | Consensus | Receipt path |
|---|---|---|---|---|---|---|
| Home Runs | Production | Production | Production | Full | Full | HR production validation receipts |
| Total Bases | Production | CatBoost validated, not wired | LightGBM + beta validated, not wired | Waiting on live output | Insufficient lanes | outputs/validation/three_lanes_beta/total_bases/ |
| Hits | Production | Random Forest validated, not wired | In active development | Waiting on live output | Insufficient lanes | outputs/validation/three_lanes_beta/hits/ |
| Pitcher Outs | Production | Random Forest validated, not wired | In active development | Waiting on live output | Insufficient lanes | outputs/validation/three_lanes_beta/pitcher_outs/ |
| Strikeouts | Production | In active development | In active development | Waiting on multi-lane output | Insufficient lanes | outputs/validation/three_lanes_beta/strikeouts/ |
| Futures | Production | Intentionally limited | Intentionally limited | Waiting on multi-lane output | Insufficient lanes | docs/codex_context/three_lanes_beta_preregistration_amendment_001_2026_05_24.md |
Mithrandir+ / flagship derived metrics
Stuff+ v1.1 / promoted
Run-value target, full backfill, original feature set.
2020-2023 Statcast pitch-event backfill via pybaseball. Validation 2024; held-out test 2025.
Statcast delta_run_exp. Lower run value is better for pitchers; 100 is league average on the published plus scale.
Velocity, release point, extension, movement, location, zone height, count, pitch type, pitcher hand, and fastball-relative deltas.
STABILIZATION 100 pitches
RMSE VS FIP -0.168
ABLATION features tested
Ablation finding
Spin axis, batter handedness, and platoon context were tested. They added noise for v1.1 and were dropped from production despite a proxy-target variant producing shinier raw Y-Y numbers.
Arsenal Decomposition
Phase 2 initially attempted a Kirby Index. Audit showed public Kirby is a command metric built from release-angle consistency, while Mithrandir had built pitch-type Stuff+ composition. We renamed it Arsenal Decomposition and surface it only as descriptive support under Stuff+, not as a predictive metric.
Two-Strike Adjustment null
Two-Strike Adjustment Index v1.0 followed the Chamberlain chase-delta framework but missed the locked gate: contact-rate r² .147 vs .300 target, and raw two-strike K% remained the better next-year K% predictor (.445 r² vs .080). This stays parked as an honest null, with bat-tracking integration deferred for any v1.1 attempt.
Pitch-Type Vulnerability partial promote
Per-type validation cleared FB, CH, and SI but exposed sample gaps for CB/FC and a slider miss. Family aggregation by hitter-recognition class validated Fastball, Sinker, and Changeup; Breaking horizontal remains caution-only at r .239; Breaking vertical and Cutters/Splitters stay insufficient-sample with explicit tags.
Conditional OAA internal reframe
Conditional OAA was tested as a standalone predictive metric and missed the locked Y-Y gate: r² .174 vs .450 target. The small-sample test was excellent, though: RMSE 2.073 vs raw OAA 6.377. We therefore apply shrinkage internally below 200 defensive outs and label those rows "regression applied" instead of promoting a separate metric page.
Stuff+ v1.1 = 100 + z(predicted pitch run prevention) * 10
production target = delta_run_exp / production features = original pitch-shape + location set
Arsenal Decomposition = pitch-type-relative Stuff+ by pitcher and pitch family
Pitch-Type Vulnerability = regressed same-family wOBA-against, inverted to resistance percentile
SEAGER+ v1.0 / promoted
Plate discipline index, reframed after validation.
SEAGER+ was initially pre-registered against next-year ISO at r ≥ .300. The first formula produced r .023 and failed.
A methodology audit found a denominator error: public SEAGER is rate-based ST - HPT, not a per-pitch run-value average. Correcting it moved ISO r to .274.
Alternative-target diagnostics showed the corrected metric is a plate discipline monster: next-year BB% r .648 and chase-rate r -.692.
PRIMARY TARGET BB% / chase
ISO OBSERVATION r .274
ROADMAP called-strike weighting
SEAGER+ = 100 + z(Selection Tendency - Hittable Pitches Taken) * 10
Selection Tendency = good takes / non-hittable opportunities; HPT = hittable takes / total takes
Transparency note
The ISO threshold miss is documented, not hidden. v1.1 roadmap: called-strike probability weighting plus bat-tracking quality on correct swings if we want to chase the power-prediction framing again.
Lane health / calibration drift moat
Epistemic chrome
Models are allowed to be wrong; they are not allowed to hide it.
Last cycle 2026-07-28T03:50:09.543655+00:00 / 1 failed / 9 stale
Failed: train_stuff_plus
| Market | Lane | Logged | Brier | ECE | 14d trend | Calibration | Flag |
|---|---|---|---|---|---|---|---|
| Home Runs | Palantir | 11499 | 0.122 | 0.011 | flat | recalibrated | healthy |
| Home Runs | Fangorn | 11499 | 0.122 | 0.010 | flat | recalibrated | healthy |
| Home Runs | Valinor | 5904 | 0.116 | 0.006 | flat | recalibrated | healthy |
| Strikeouts | Palantir | 964 | 0.247 | 0.031 | degrading | recalibrated | drift |
| Strikeouts | Fangorn | 53 | 0.257 | 0.119 | insufficient sample | skipped insufficient sample | insufficient sample |
| Strikeouts | Valinor | 460 | 0.247 | 0.038 | degrading | recalibrated | drift |
| Total Bases | Palantir | 5326 | 0.239 | 0.073 | degrading | recalibrated | drift |
| Total Bases | Fangorn | 0 | -- | -- | tracking | skipped insufficient sample | tracking in progress |
| Total Bases | Valinor | 0 | -- | -- | tracking | training time | tracking in progress |
| Hits | Palantir | 344 | 0.177 | 0.049 | degrading | skipped insufficient validation split | drift |
| Hits | Fangorn | 0 | -- | -- | tracking | training time | tracking in progress |
| Hits | Valinor | 0 | -- | -- | tracking | training time | tracking in progress |
| Pitcher Outs | Palantir | 512 | 0.259 | 0.056 | degrading | recalibrated | drift |
| Pitcher Outs | Fangorn | 0 | -- | -- | tracking | training time | tracking in progress |
| Pitcher Outs | Valinor | 0 | -- | -- | tracking | training time | tracking in progress |
| Futures | Palantir | 0 | -- | -- | tracking | skipped insufficient sample | tracking in progress |
| Futures | Fangorn | 0 | -- | -- | tracking | training time | tracking in progress |
| Futures | Valinor | 0 | -- | -- | tracking | training time | tracking in progress |
Healthy
11499props
0.122Brier
0.011ECE
Healthy
11499props
0.122Brier
0.010ECE
Healthy
5904props
0.116Brier
0.006ECE
Rolling projections / Bayesian shrinkage
Daily update layer
The prior stays visible; the season earns weight one game at a time.
The frozen projection from the original season simulator. It remains on-page as the baseline.
The raw extrapolation of observed 2026 performance. Useful, but noisy.
The Bayesian blend: preseason prior plus observed performance, weighted by sample size.
TEAM WINS 60-game regression constant
HITTER RATES 220 PA
PITCHER RATES 240 BF
observed_weight = opportunities / (opportunities + regression_constant)
rolling_projection = prior * (1 - observed_weight) + current_pace * observed_weight
team game lines = rolling team talent + opponent context + starter rolling ERA + bullpen FIP + lineup status
Limitation
Small samples mean high uncertainty; early-season rolling projections intentionally mean-revert toward preseason until the observed sample becomes persuasive.
Per-model write-ups / status
Stable / Projection-Anchored
Palantir Model
Strikeouts / K Valinor
Boosted residual + calibration lane
Total Bases / Fangorn
Selective nonlinear signal lane
Regularized, projection-anchored lane built to stay stable rather than chase the sharpest take.
Stable / Readable mainline probabilities and broad projection context. Nonlinear / Interaction Discovery Fangorn ModelNonlinear interaction-seeking lane that can abstain when it does not see a distinct pocket.
Advanced / Comparing where tree-style interaction logic diverges from the stable anchor. Calibration / Probability Refinement Valinor ModelProbability-refinement lane built to ask whether center, tails, and reliability are shaped correctly.
Experimental / Calibration review and challenger comparison, not declaring the most aggressive pick.Production calibration lane is visible with a caveat banner while forward-logged graded sample accumulates. Rows flagged with the fallback path stay disclosed.
production / caveatedRebadged research lane remains disclosed as non-production independent methodology until a true Total Bases Fangorn engine ships.
rebadged researchValidation
HR Valinor is the worked example.
Non-stacked LightGBM + beta calibration promoted on ECE strength; Brier difference was inside bootstrap noise.
Glossary / formulas
ECE = sum_b (n_b / N) * abs(avg_pred_b - avg_outcome_b)
Brier = mean((p_i - y_i)^2)
RA_FV = FutureValue * P(reaching ceiling)
Versioned changelog