Arm A
[Auto] NFL Spread (ExtraTrees)
- ATS win rate
- 51.30%
- Bets
- 5,961
- Kelly ¼ / $1k
- $0
as of Aug 17, 1:41 PM ET
EDGE · BENCH
Assembly Bench
Snap together a typed model, then test every connection before training.
Build a module on the fly · three ways, one contract
(game, pregame ctx) -> score in [-1, 1], confidence · runs in the sandbox · tested on the holdout before it can touch Tonight
Lean to the home team when they have 3+ more rest days than the visitor, stronger in weeks 10+.
compiles to ->
rest_days_delta >= 3 -> +0.3 - week >= 10 -> x1.5
not graded yetThe ML stage, opened
The default pipeline behind one brick — static defaults, not measured.
| Step | Default | Why |
|---|---|---|
| Target | cover | totals / ML / props / margin are one switch |
| Features | packs + feeds + module outputs | modules vote (weight) or feed (column), never both silently |
| Engineering | rolling 4 - opp-adj - rest D | computed per week with that week's data |
| Learner | LightGBM d4 x 300 | 19 blueprints behind one brick |
| Validation | walk-forward 5 | shuffled k-fold is not a brick |
| Calibration | isotonic | so 0.62 means 62% - heat depends on it |
| Gate to Tonight | beats baseline - n >= 100 - gap <= 2 | otherwise the card says so |
Level 1 · Data stage, opened · typed wires, live
Open one stage without losing the six-socket model spine.
0/4
Data socket
Wires carry their type. A Feed lens joins the Features Frame "as feature"; a Filter lens gates rows through its threshold. A results column wired toward the learner draws red: PostFrame can only feed Eval — leakage is a type error, not a lint.
Wire types
Framerows + columns, pregame PostFrameresults — Eval only, by type Signalscore + claim Probcalibrated % Predicatekeep / drop Modeltrained Metricswith n + baseline · adapters auto-insert where one fix exists; red wires explain themselves
Bench evidence · what builders actually use
Signed-in registry aggregates only; the anonymous demonstration catalogue is never charted.
Sign in to load Bench evidence
The anonymous Builder payload is a local demonstration catalogue, not registry evidence.
Unlocks the moment you're signed in.
Sign inSharp money
Inspector pane · role contract and settled proof.
Sharp money inspector needs complete evidence
lens_registry exposes family, gate, graded_n, hit, lift and drift, but the Bench loader has no remix count or threshold-to-slate result for Sharp money.
A / B — which belief pays?
Two beliefs over the same rows; the edge says which one pays.
Arm B
[Auto] NFL Spread (AdaBoost)
- ATS win rate
- 49.37%
- Bets
- 5,961
- Kelly ¼ / $1k
- $0
as of Aug 17, 1:40 PM ET
Same graded rows — 5,961 bets, 152 pushes — one change.
Build a module on the fly · three ways, one contract
Sentence reads a selected registry module's manifest and joins the current draft through BlueprintSlots.features. Formula and Code are drafting surfaces: they compile or run through the real compound-metric parser and the real Pyodide sandbox — not a stored module field, since brick_registry has no formula/code source columns yet.
Sign in to use registry modules
The anonymous build payload is a local demonstration catalogue; this block only renders signed-in brick_registry rows.
Unlocks the moment you're signed in.
Sign inThe ML stage, opened
Current draft fields are shown exactly as BlueprintSlots represents them; unsupported settings are named rather than guessed.
| Step | Current draft | Why |
|---|---|---|
| Target | Not selected | BlueprintSlots.target owns the market target. |
| Features | Not selected | BlueprintSlots.features holds feed and module references. |
| Engineering | Not represented in this draft payload | BlueprintSlots has no standalone engineering field. |
| Learner | Not selected | BlueprintSlots.architecture owns the learner reference. |
| Validation | Not represented in this draft payload | The build loader exposes no validation setting. |
| Calibration | Not represented in this draft payload | The build loader exposes no calibration setting. |
| Gate to Tonight | Sign in to evaluate | The anonymous demonstration catalogue is not registry evidence — sign in to gate a real model_cards row. |
The ML stage is not configured
Choose a target, features, and learner before training this draft.
Training trace · live
This pane only becomes live when the loader can supply real training-run events.
No live training trace
The reused /build loader exposes no training-run event stream; a scripted live trace would be false.
Time machine · replay the season
A season replay requires real week-by-week settlements from the model ledger.
Season replay needs weekly settlements
The reused /build loader exposes no week-by-week settled ledger for replaying a season.
Level 2 · the whole machine
The full graph is generated from the current draft slots; no training or settlement result is implied.
Populated adjacent layers are connected by the canonical Blueprint graph migration. The Monaco code drawer is not built yet (blocked — brick_registry has no formula/code source column, so no canvas node is ever identifiable as a code brick).
Training trace · live
Fold progress belongs to a real training run, not to the editable Blueprint draft.
run startedbp_et_mlb_ml · mlb · 2026-08-28T03:36:46.751347+00:00
outcometrained
duration80.5s
smoke gatepassed
artifact retention5 prior artifacts retired
walk-forward fold metricsmodel_training_runs stores no per-fold Brier/holdout series for this run
model_training_runs · run 40043 · gate PASS
Open full BuilderTime machine · replay the season
The replay needs the schedule, graded ATS receipts and cumulative weekly units together.
wk 18
Cumulative at $100
-$38,645
4680–4641–35 · 9356 plays2025 house spread record, real graded games only.