How the Sports Lab Files a Pick

ThoroughWiz's sports section covers four leagues — the NFL, the NBA, college football and college basketball. Every league files under one published rule. The rule is printed before the picks, the picks are filed before the games, and the results settle from the official finals into a public ledger. Every pick is labeled UNPROVEN. The section is a paper record: it places no wagers and claims no edge.

The pick is a differential

The spread pick in every league comes from one number: each team's season scoring differential per game — points scored minus points allowed, divided by games played — counting regular-season and postseason games only (preseason and exhibition games never enter the window), then adjusted for the opponents faced: each game's margin is adjusted by the opponent's own average margin over two passes, with the NFL's form double-weighting each team's last 8 games. The team with the better adjusted differential is the pick, taken at the filed spread. When the two adjusted differentials sit within 1.0 point of each other, the pick is the home team. That is the whole selector.

Lines are captured at filing time from the anchor board — DraftKings where it carries the game, otherwise the first US book carrying the markets the pick needs. A game files only when the anchor book carries those markets and both teams have completed games this season; otherwise it is excluded and the reason is recorded.

The four leagues, as they file today

One rule history covers the section, but the leagues do not file the same book. The table is the current state; each league's guide carries its full treatment — its rating, its markets, its data sources and its own history.

The consensus era — and why it ended

The lab did not start with the plain differential. On November 6, 2026 the sports section switched to a points consensus modeled on the racing lab's method: five factors — the market, an opponent-adjusted power rating, key numbers, rest, and venue-split form — each voting 5 or 2 points for a side, with dead-bands where a factor stays neutral. A sixth factor, public opinion, joined the next day (Rule v1.1, below).

Then we found a bug in our own backtest: away against-the-spread picks had been graded with the spread applied backwards. We reran the entire baseline corrected. Thirteen of the sixteen against-the-spread verdicts flipped to NOT SUPPORTED, and no corrected spread cell reached 52.38% — the cover rate a standard −110 bet needs just to break even. The prune that followed (Rule v1.2) is the current shape of the section:

• Totals machines off, every league. Every totals cell in the corrected record graded negative.
• Moneyline column off. The moneyline pick restated the market favorite — it was never a selection, and it graded as one: it won at the market rate and lost the vigorish.
• Consensus overlay off the spreads. The machine added complexity on top of a core the corrected record did not support.
• Spreads reverted to the plain differential — on the corrected record, the least-negative against-the-spread machine the lab has tested. That is a weaker claim than "supported," and it is recorded as one.

Public opinion: the factor that still runs

Rule v1.1 added F6 PUBLIC to the consensus: every morning at 9:05 ET a snapshot tallies positive mentions of each side from forums, social posts and picks pages. Only positive mentions count — backing language such as "best bet," "taking," "rolling with." A mention carrying negative language toward a side is removed, never counted; neutral mentions do not count; one mention per candidate per source. Where a game has no usable chatter, the factor is inactive with the reason recorded — it is never estimated and yesterday's tally is never carried forward.

The v1.2 prune switched off the machines public opinion voted in, so in the sports section it currently selects nothing. The snapshot still runs every morning, the tallies still publish, and the factor still votes in the Numbers Lab's machine. We print that plainly rather than let a retired input look alive.

The two totals that still file

Two totals survive the prune, both NFL-only, both standalone situational filings rather than machines:

• The wind under (Rule v1.2 WIND). When an outdoor NFL game's forecast wind at kickoff is 15 mph or more, the lab files the Under, labeled with the forecast wind, the venue and the forecast source (Open-Meteo, read by stadium coordinates). In the lab's own history study — NFL games 1999–2026 at closing totals — those games went under 56.97% of the time, +10.05% ROI on 684 bets. That is annotation-grade support from a historical study, not a live record; the filing is UNPROVEN like everything here. No forecast, a roofed venue, or wind below 15 mph means no pick — never estimated, never substituted.
• Primetime unders (Rule v1.3). Below.

Rule v1.3 — two spots, adopted with their flags

New angles reach this section one way: published ideas are harvested, frozen word-for-word in a preregistration, and tested on the lab's own full history — the NFL corpus runs back to 1999 — under the lab's gates. An outside finding is a candidate, never a result. On November 8, 2026, two spots from that process were adopted. Both are NFL-only — no other league files either spot. Both file in their own record fields and their own ledgers; neither changes the differential pick; neither can become a parlay leg.

Both records carry the lab's concentration flags, printed here the way they are printed in the rule file. For the revenge spot, the single season 2003 contributes 59.78% of the total profit. For primetime unders, 2022 alone contributes 57.14% of the profit, and the first half of the sample is negative. Under the lab's standards, an edge concentrated like that is presumed a spike until the concentration is explained. So both spots file as UNPROVEN, and the prospective record is the judge from here: each type gets a checkpoint report after 30 settled picks.

Line movement is a display, not an input

Each league page also carries a line-movement block: the first line we captured for a game and the latest, with the move between them. It is a published supplement and a lab ledger — it touches no pick. The ledger's history begins November 5–6, 2026, which the pages state: days of history, not seasons. Any future use of movement as a pick input requires its own pre-registered, gated test first. The ledger existing is not evidence of an edge.

Settling and the record

Every pick settles from the official finals, in the market it was filed in, under the rule version that filed it — picks filed under an older version stand as filed and are never restated. Rule changes ship as new numbered versions with the record that caused them, the way v1.2 carries the corrected-baseline story and v1.3 carries its flags. The ledgers are public, including the losing weeks.