PHI -1.5 to CAR -3.5 and back to PHI -1.5: the line this research solved, then un-solved

Carolina beat Philadelphia 25-14 on October 18. GameScript's own house line had Carolina favored by 3 points. That's the normal, boring outcome of a spread call — directionally right, short on the margin, nothing to write up on its own. What's worth writing up is what this site's own research did to that number in the six days before kickoff: it found out why the house line was right, then two days later reported the opposite and proposed that its own employer's software might have a sign-flipped bug.

What the research found on October 16

The house line on CAR@PHI flipped from Philadelphia -1.5 to Carolina -3 on October 14 — the sharpest single-game move anywhere on the Week 6 board that day. For two passes, the research couldn't explain it: every dated external source it could find still showed Philadelphia favored, and it said so plainly rather than guessing.

Then, on the 16th, it found the explanation. The Eagles' own injury report, dated the 15th, ruled Jalen Hurts out for a blood clot in his calf, with Tanner McKee set to start. The research logged that finding in exactly those terms: "the first solid, dated, primary-source explanation found across five passes for this move," and recorded that the external market had moved with it — Carolina as a 3.5-point favorite, matching the house number almost exactly. The question was closed. A real injury, a real line move, a house number that called it before most of the open web had caught up.

What it found two days later

By the 17th, the research was reporting Philadelphia favored again. By the 18th — game-day morning — a four-book sweep (FanDuel, the DraftKings-sourced line on SI Betting, Covers, DeFi Rate) all showed Philadelphia -1 to -1.5, and the write-up called this "a seventh straight pass" where the house number disagreed with the market, treating the question as never resolved rather than re-opened.

From there it built a specific, well-argued hypothesis: GameScript's own internal sourcing notes state plainly that its default odds feed, nflverse's spread_line, is written positive when the home team is favored — "the opposite of the homeSpread the model stores" — and name getting that backwards as a known risk, not a hypothetical one. Philadelphia was the home team. CAR@PHI was the one game on the slate where the house favorite and the external market disagreed on direction. The research flagged it for an engineering check, carefully hedged — "this research has no access to the ingestion code or logs... is not asserting the number is wrong" — and that hedge was earned. It just pointed in the wrong place.

What was actually going on

Nothing about the house number had changed since the 14th. What changed was what the research was citing. The claim recorded on the 17th describes the external lines it checked as "inconsistent with each other and mostly pre-date Friday's Barkley-OUT confirmation" — one undated, one explicitly dated Tuesday the 13th, two days before the Hurts news the research itself had already surfaced. The 18th's four-book sweep doesn't name dates for three of the four books at all, and the one it does cite, FanDuel's page, carries only a relative byline ("6 days ago") with no dated snapshot to check against.

I re-fetched that same FanDuel page today. It currently shows Carolina -3, not Philadelphia -1.5 — the house number, not the number the October 18 pass recorded finding there. I can't tell you whether that page actually read differently three days ago, or whether a fetch crossed with a cached or mismatched response; I didn't capture it myself at the time, and neither did the research — it logged a result, not a saved page. What I can tell you is that the research's own October 16 pass had already done the work of resolving this: primary source, dated, with a named mechanism. The 17th and 18th passes didn't check that finding against what the sources they were now reading actually said happened after it. They checked the live house number against sources that, on inspection, mostly predate the thing the research itself had already confirmed.

That's a cheaper, less interesting failure than a sign-flipped ingest bug, and it's also the one the evidence actually supports. The sign-convention risk is real and worth naming — this product warns about it internally, by name — but naming a real risk isn't the same as this being an instance of it, and the research had, two days earlier, already done the work that would have told it so.

What happened

Carolina won 25-14. Bryce Young threw two touchdowns; Chuba Hubbard ran for 120 yards and a score on 28 carries; Tommy Tremble caught two touchdowns of his own. Tanner McKee made his 2026 debut at quarterback for Philadelphia, as the research had correctly tracked since the 15th. DeVonta Smith, who the research's pool-proxy math had flagged mid-week as still showing a real, non-zero projection despite a lingering questionable tag, caught a touchdown — the one live read that stayed accurate through every version of this story. Carolina's 3-point spread undersold the final margin by eight, in the same direction every blowout on this slate's board this season has undersold its spread. That part isn't new; it's the same lesson this site has written about a spread's margin versus a total's points before. The number worth sitting with here is narrower: the house line was right about who would win this game for four straight days before kickoff, and for two of those four days, this site's own research reported that it might be broken.

The actual lesson

"Unchanged for N passes" describes how many times a question was asked, not how many times it was actually re-verified. A disagreement that reappears after a pass has already closed it deserves one specific check before it gets escalated: not "does this still disagree with our number," but "is what I'm reading now dated after the thing that already resolved this." The October 17 and 18 passes asked the first question repeatedly and never asked the second — and arrived, reasonably and in good faith, at a mechanism that fit the symptom but wasn't what produced it. The fix isn't more skepticism of the house number. It's treating a research thread's own prior pass as a source that has to be checked against, with a date, the same as any outside one.

Sources

• Philadelphia Eagles: Panthers vs. Eagles injury report, Week 6 2026
• FanDuel: Eagles vs. Panthers predictions, odds, picks for NFL Week 6 2026
• SI Betting: Panthers vs. Eagles score prediction, Week 6 — can the Birds bounce back after the London loss
• Fox Sports: Week 6 Carolina Panthers vs. Philadelphia Eagles box score, October 18, 2026
• GameScript's internal sourcing notes, on the nflverse spread_line sign convention