Louisville vs Ole Miss. Our pick: Ole Miss -6.5 (medium). No pick on the total.

Provenance: schedule and finals from the Edge Labs database (2026 season); the line is the latest capture for this game with the book named; our number is the EL rating read (in-season once a team has played, preseason before that) plus 2.5 home field, none at neutral sites, and it is context on this page rather than the call. Signals come from the trend engine, each with its real sample; a tier is how decisively the evidence agreed and is never a win probability. Once the game kicks off this page stops computing and reads our pick lock ledger instead: the pick and the spread we were locked at, written once at kickoff, never updated, graded against that same number. Line movement comes from our permanent capture log, one book across both ends and never a capture taken at or after kickoff, so the second number is the last pregame line and after kickoff it is the close; a game the log has captured only once shows no movement rather than an invented one. Weather is the captured forecast for the venue, and the chip appears only when it is worth saying (wind at 12 mph or more, a 50 percent or better chance of rain, or 35F or colder), never on an indoor venue. Injury counts are our latest daily scan, real report rows only, and a team the scan does not cover is left out rather than shown as zero. The write-up is assembled from those same stored rows, never written around them: the case comes from the lock row, the signals from the trend engine with their own samples, the series from our game database from 2013 forward, and the scoring profiles from completed games only, each with the number of games it averages. A section with no data behind it is left out instead of filled in, it refreshes while the game is pregame, and it can never be edited once the game has kicked off. Trends are descriptive only, computed by the same shared module the write-up uses so the two cannot disagree: straight-up splits come from completed games in our database, situational splits look back three seasons and need at least eight decided games, records against the number and on the total arrive with the closing-line restore, and every line carries the games it counts. Filters like these did not hold up as predictors in our testing, so none of them enter the confidence read. The totals block is a read and not a pick: our number is EL CFB Totals v1 (edgelabs.el_cfb_totals, season 2026, model_version v1, docs/el-cfb-totals-v1.md), a registered candidate that loses to the closing total, so it never posts to the board or the record; the market total beside it is the same capture the spread comes from before kickoff and the frozen consensus the model priced against after it, and the gap is measured against whichever number is printed. The totals trends under it are the trend engine's own rows for this game, deduped on the headline, each carrying its sample. On the games the analyst panel works each week (the games our model prices off the market plus the ranked and Power Four games, at most twenty), the pick and the total call are the panel's synthesis: three independent analyst panels, the same evidence pack, their own research, one side each, tiered by how firmly they agreed, never a win probability. A panel majority on a side is what makes a pick; without one the game is a published no-pick, printed as No pick rather than left blank. Research and context, never a guarantee, 21+.