Line Lab Research

We're Auditing X's Handicappers With Their Own Timestamps. Records Are Easy to Fake — Closing Line Value Isn't.

Sunday, August 2, 2026

X is full of accounts selling picks off claimed records nobody can verify, and YouTube is full of channels reading the same kind of card out loud for free. We built the verification for both. Because this project already stores its own closing line on every game — thirteen sports of them — we can do the one audit a screenshot can't fake: take a capper's own timestamped, pre-game post, grade the pick against the final score ourselves, and check whether the number they posted beat the close. Skill shows up in closing line value long before it shows up in a win-loss record.

Who we're tracking, exactly

• @TherealLVtruck — a low-volume Vegas bettor posting first-five-innings plays in old-school rotation-number format, with prices, before first pitch. New account; small sample; the format is why he's trackable.
• @BetScopeLeaks — an account that leaks paid capper picks from ~40 sellers, with stake, price and timestamp. We don't care about the leaks as picks; we care that it's a firehose of structured, timestamped tout behavior across MLB, tennis, soccer and more.
• @CapperLedger — an adversarial tracker that independently logs big-name cappers' bets (including a 600K-follower account). Their data is their grading, so we label it claimed-only — but their random coin-flip control bet climbing their own leaderboard tells you what they keep finding.
• YouTube free-picks channels — 28 of them so far, and already about 26% of the ledger. This tier doesn't paywall anything: it earns from ads and subscriptions, so the entire card gets read out loud in a daily video. We search those uploads, pull the transcript, and take the plays out of the speech. The channel is the capper, and it lands in the same ledger under the same schema as everything above — same grader, same CLV test, same table.
• Auto-discovery: weekly searches for structured pick-format posts add new accounts to a probation tier automatically, capped so the roster stays curated.

What we do with every pick

Daily, the tracker pulls new posts via the X API (timestamps prove pre-game), parses the structured picks, then runs two independent measurements: (1) re-grade — we settle every pick ourselves from final scores (and inning-by-inning data for first-five plays); their emoji never enter our books — and (2) CLV — we compare their posted price to our own stored closing line on the same market. One honest caveat on that second number: leaked prices are line-shopped bests while our close is a consensus number, so the CLV column carries a positive bias — it's for comparing cappers against each other, not reading absolutely.

The two feeds differ in exactly one place: how the pick gets read. X posts are structured enough to parse with rules. A YouTube transcript is conversational speech mangled by auto-captioning — "Nerfy" for NRFI, "White Socks" for White Sox — so a language model pulls the plays out of it. That is extraction, not prediction: it reports what was already said out loud, and we measure it inventing a pick that wasn't there 0.44% of the time. It never forecasts a winner, and it can't flatter anybody, because everything it extracts is then re-graded against the box score by the same code that grades the X feed.

Timing is the one thing a video can get wrong that a timestamped tweet can't — a 22:00 upload might be a recap of games already played. So the slate date comes from the title whenever the title carries one, falls back to the publish timestamp for ordinary daytime uploads, and a late-evening upload with no dated title is dropped rather than guessed at. A pick we can't place on a slate is not a pick.

The scoreboard

Independent regrades only (minimum 15 graded picks), updated nightly:

Ledger: 6168 picks from 122 cappers; 2739 independently regraded, 1284 CLV-matched. (Tennis, UFC, esports, props and parlays are logged but not regraded — we only settle what our own data can verify.)

The early finding: tout consensus is a fade

The sharpest split in the data so far, on MLB moneylines we've regraded: when two or more cappers land on the same side, that side is 444-459, -78.82u (-8.7%). Picks only one capper made: 137-107, +20.04u (+8.2%). The crowd of touts, like the on-chain whales we tracked before them, looks like a contrarian indicator — and we also see them systematically piling onto favorites against our own underdog triggers, and losing when they do. Both reads are two-week samples; both will harden or die in public on this page.

One angle is already being tested that way: since August 5 the AI panel that scores our own fires also scores the sides these cappers take, pre-game, with the rest of the board as a control arm — see Crossing Two Experiments.

We never tail anyone. The point of this tracker is upstream: any capper whose CLV survives a month of forward data gets their angles — which markets, which sides, which timing — fed into our own mining pipeline as hypotheses against 18,000 games of stored closes. People are noisy; angles are testable. Nothing here is betting advice, and no money follows anything on this page.

← All research