e:I[98500,["500","static/chunks/500-2d1136a2b6ea2d76.js?dpl=4b9c4ad4330944e59ba357e78d79721d3b3a54260ea29968a33fe381c2f4e291366139313962633033373632383730303038633731333434","804","static/chunks/app/design-review/page-6035c1a1bc6cd66d.js?dpl=4b9c4ad4330944e59ba357e78d79721d3b3a54260ea29968a33fe381c2f4e291366139313962633033373632383730303038633731333434"],""] f:T123bd, Sunday Lab — design review

Sunday Lab · design review v0 · schema + interface

A forecasting lab
for two people.

Price the Sunday slate, state your number, take a position in practice credits — then get scored against the closing line instead of against luck. Credits are not money: no currency, no payment rails, no sportsbook account. Balances are exact numerics in a Postgres table.

Closing line value: Did you take a better number than the market ended at? Reported per unit, with availability beside every mean.
Calibration: When you say 60%, does it happen 60% of the time? Separates being right from being lucky.
Return on credits: Did the balance grow? The number you feel — kept last, on purpose.

The loop

How a week runs

Five beats, Wednesday to Monday. The lock transition is the one that carries weight: stakes debit and closing lines freeze in the same instant, because if those two drift apart the scoring is meaningless.

Wed

Slate opens. Games, opening lines, injury designations, and model benchmarks ingest. Every row lands with a source and a cutoff.

Wed–Sun

Forecast first. State a quick game belief or a direct proposition probability before market and model comparisons appear. Take a position or pass.

Kickoff

Lock per game. Due forecasts reveal and due entries freeze at their own kickoff. Later games stay editable.

Mon

Settlement. Game forecasts, proposition projections, and positions grade independently; the credit ledger appends exactly once.

Ongoing

Post-mortem. Where were you right for the wrong reason — and where did the market move against you before the ball was snapped?

Screen 1 of 5

The Board

The Sunday slate, laid out as an odds board rather than a card feed. Rotation numbers on the left are the stable handle for a market — you say “I took 452,” not “I took the Ravens thing.” Click any price to stage it.

sunday-lab / board — 2026 · week 1

RotMatchupKick SpreadTotalMoney MoveForecasts

The movement column is the highest-value cell on this screen. A team that opens −2.5 and closes −3.5 is telling you something a single overwritten “current line” would have destroyed — which is why market.odds_snapshot is append-only and there is no current-value column anywhere in the schema.

Screen 2 of 5

Game

The workflow begins with your belief, not the price. Quick mode takes three numbers; direct mode remains available for props. Market and model benchmarks reveal only after the forecast is committed.

1 · Record your game forecast

Kansas City vs Baltimore · cutoff Sat 18:00 ET · benchmarks sealed

KC win

KC margin

Game total

Under a minute per game. Win probability is already a game-result forecast; mean margin and total alone do not invent spread or total probabilities. Choose a versioned dispersion template in Advanced mode or state an exact proposition directly.

Game forecastwhat you believe

→

Projectionone side + line

→

Position or passwhat you chose

Line movement — KC spread

451/452 · opened −2.5 · current −3.5

Consensus line You entered −2.5 Peer activity hidden until kickoff

2 · Same-cutoff comparison

KC to win · all values at Sat 18:00 ET

Comparison sealed
Commit your forecast before seeing market or model numbers.

Efficiency matchup

EPA per play · season to date · model output

Injury designations

nfl.injury_report · latest as_of per player

Player	Tm	Status	As of
I. Likely TE	BAL	Out	Fri 16:20
Z. Orr LB	BAL	Quest.	Sat 11:05
R. Rice WR	KC	Quest.	Sat 11:05
J. Taylor OT	KC	Full	Fri 16:20

Designations are versioned rather than overwritten. A Wednesday Questionable and a Saturday Questionable are different information.

Screen 3 of 5

The Slip

The final decision layer. The return stays hidden until you commit to a probability. Seeing “+240 returns 34.00 cr” first anchors the number, and an anchored number makes the entire calibration dataset worthless. Try it — the panel on the right stays sealed until the slider moves.

sunday-lab / slip — week 1 · draft

Kansas City −2.5

451/452 · spread · −110 · consensus

1 · Direct proposition probability or derived from game forecast

— —

2 · Position size 2.0% of credits

= 20.00 cr

3 · Rationalerequired

draft→ submitted→ locked→ graded→ settled

XState reflects client interaction only. The server locks each due entry, freezes its close, and appends the stake in one transaction.

Screen 4 of 5

Head to Head

The Monday-morning screen, and the one that gets argued about. Two credit curves, directly labeled — the running gap between them is the point.

Jovani · credits

1,184

+18.4% from 1,000

Marcus · credits

927

−7.3% from 1,000

Forecasts graded

positions are not the scoring denominator

Positions taken

86.9% of forecasts · passes retained

Practice credits · season to date

lab.v_credits · a fold over the ledger, never a stored balance

Week 11 ledger

both slips, side by side

Rot	Market	Who	Took	Close	Line CLV	Stake	Result	Net cr

Jovani Marcus

Screen 5 of 5

Post-mortem

Three questions, three scorecards. Forecast quality, decision quality, and market execution never collapse into one headline number.

Forecast quality · 84 games

Brier 0.205

Market 0.199 · model 0.213 · matched cutoffs only

Decision quality

73 / 84

forecasts became positions · 27 direct / 46 derived

Market execution · spreads

+0.7 pt

54 available / 61 positions · 7 unavailable

Spread line CLV

one unit only · 54 available / 61 positions · mean +0.7 pt

Totals and player props have their own panels because points, yards, and receptions are not commensurable. Missing CLV remains in the denominator.

Calibration

proposition forecasts · stated vs. realized · dot area = n

The high-confidence gap is inside the current neutral band, so it is shown without an alert. Thresholds trigger only when deviation and sample size both justify attention.

No alert. 10 forecasts in this bucket; review begins at 25 or a larger sustained gap.

Underneath

Six schemas, because the layers are not the same kind of claim

Lifted straight from the data-layer table at the end of awesome-nfl-data. An EPA figure is a model output, a spread is a market signal, and a completion is an observed fact. Collapsing them into one stats table makes the post-mortem question unanswerable within about three weeks.

nfl

Observed facts

team, player, game
player_week_stat
team_week_stat
injury_report

chart

Human classifications

play_classification
coverage, personnel, pressure
vendor-dependent, optional

model

Model outputs

feature_snapshot, training_run
calibrator, forecast_run
game_forecast → proposition_probability

market

Market signals

odds_snapshot (append-only)
closing_line (frozen once)
fair_probability (devigged)

lab

The product

game_forecast → forecast → entry
outcomes + per-unit scorecards
ledger_entry (credits, append-only)

meta

Provenance

source (license, terms)
ingest_run (watermarks)

One decision does most of the work. nfl.player_week_stat is long-format — (player, week, stat_key, value) — and market.market carries a matching stat_key. Adding “receptions” as a bettable prop is a row, not a migration, and grading it is a join rather than a CASE block.

Build order

Step 2 before step 3 is the ordering that matters. Grading bugs discovered after six weeks of weekly play cost you the dataset, which is the only thing this project produces.

Schema + ingest. Load a completed past season from nflverse. Everything downstream is then verifiable against known outcomes.

Grading and CLV, headless. Replay that season. Confirm settlement math against results you already know. No interface.

The Board and The Slip. The two screens that make it usable at all.

Head to Head. Enough to actually play a week against each other.

Post-mortem. Available from week one with denominators visible; interpretations strengthen as coverage grows.

The core is market-agnostic on purpose: a market is anything with a priced side and a resolvable outcome, and an entry freezes what you took and what you believed. The NFL is this lab’s subject — the same core is meant to be lifted into sibling repos rather than bolted on here. A future NBA lab is a second ingest path against the same five screens, and “will core CPI print above 0.3% month over month” is structurally a total, scored exactly the same way.

4:["$","main",null,{"id":"main-content","className":"min-h-dvh bg-board text-board-foreground","children":[["$","div",null,{"className":"flex flex-wrap items-center justify-between gap-3 border-b border-board-rule px-4 py-3 sm:px-6","children":[["$","div",null,{"children":[["$","p",null,{"className":"font-semibold text-white","children":"Sunday Lab design review"}],["$","p",null,{"className":"mt-0.5 font-mono text-[0.62rem] uppercase tracking-[0.12em] text-board-muted","children":"Illustrative prototype - not production evidence"}]]}],["$","$Le",null,{"href":"/","className":"rounded-sm border border-board-rule px-3 py-2 text-sm font-semibold hover:border-board-accent hover:text-board-accent focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-board-accent","children":"Return to the live workspace"}]]}],["$","iframe",null,{"title":"Original Sunday Lab five-screen design review","srcDoc":"$f","sandbox":"allow-scripts","referrerPolicy":"no-referrer","className":"block h-[calc(100dvh-4.5rem)] w-full border-0 bg-background"}]]}]