Line Lab Research

Crossing Two Experiments: We Pointed the AI Panel at the Touts. So Far It Says Every One of Their Sides Is Overpriced.

Thursday, August 6, 2026

Two of the experiments on this site were built to answer different questions. The capper tracker asks whether anyone selling picks on X has real closing line value. The Consensus asks whether a three-model AI panel can rank our own mechanical fires by profitability. Since August 5 they're wired together: the touts generate candidates, and the panel scores them before first pitch. Neither one is allowed anywhere near a dollar.

Why cross them at all

Our standing position on handicappers is that they are hypothesis generators, never tail targets. Across the cappers we track, none survives multiplicity — the best ROI on the board carries a family-wise p-value around 0.25 once you account for how many accounts we're testing, and its closing line value is indistinguishable from the pool average. So "who should I follow" is the wrong question. The right one is narrower: the touts surface a dozen or more live sides a day that our own rules never look at. Is there a filter that separates the good ones from the noise?

The panel is a candidate filter. On each side it runs a research pass over the real-world context (probable starters, form, injuries, park), then an adversarial debate — one frontier model argues the backed side, another argues the opposite — and a third model reads both blind and returns two numbers: an independent win probability and a 1-10 confidence score. Crucially, the panel does not know a capper is involved. It scores the side.

Three pools, never pooled

All three arms write to one ledger under a source tag, and we do not merge them, because they are different populations answering different questions:

• fire — the bets our Rule A/B/C triggers actually fire. Pre-registered, gated at 100 plays, published on its own page. Currently 5 graded. Nothing on this page touches it: pouring extra volume into a pre-registered analysis destroys the question it was registered to answer.
• tout — MLB moneyline sides tracked X cappers took, scored pre-game. 98 scored, 98 graded. This is the only pool where the backed team can be the favorite, which is exactly why the ledger settles on the backed team's own scoreline rather than our database's underdog column.
• slate — every other game on the board, dog side, scored the same way. 66 scored, 66 graded. This is the control arm, and it's the most important of the three: a panel that ranks equally well everywhere is a general baseball ranker, not evidence about touts or about our rules.

The early disagreement is unanimous

Every scored side carries an edge number: the panel's blind win probability minus the price Polymarket was actually asking for that team at score time. Positive means the panel thinks the market is too cheap. So far the panel has priced 86 of 98 capper sides below the market's own ask, at a mean edge of -0.120.

Read that carefully, because it is not a result yet — it's a description. The touts overwhelmingly play favorites, favorites are the expensive side, and a value-oriented ranker will call expensive sides overpriced almost by construction. The interesting question isn't whether the panel dislikes the touts. It's whether the ordering within those sides — which ones it dislikes least — has any relationship to what they earn.

The ledger

Every capper side the panel has scored, its Polymarket ask at score time, which accounts were on it, the panel's blind probability, its score, and the outcome. Written pre-game, never edited, rebuilt at every nightly grade:

Where the three arms stand

And the three secondary reads, all of which need far more data than they have:

• Does the score rank profit? Rank correlation between the 1-10 score and realized P&L per $1: +0.095 inside the tout pool, +0.177 inside the control. The gap between those two numbers — not either one alone — is the whole point of running a control arm.
• Does the panel beat the price? Brier score, paired, panel probability against the market's own implied probability: panel 0.2588 vs market 0.2446 (n=98) — the market is ahead by 0.0142. Beating a liquid market on public information is rare, and this project has killed every previous claim to have done it.
• Does the consensus fade hold up here? Our capper database says sides two or more cappers agree on are a fade (444-459, -78.82u (-8.7%)) while solo picks are not (137-107, +20.04u (+8.2%)). Inside the panel-scored subset: solo 25-25, -3.92u (-7.8%) n=50, consensus 23-25, -5.37u (-11.2%) n=48.

What would make this real, and what would kill it

A result means the score orders P&L inside the tout pool and does it better there than in the control — that would say the panel is finding something specific in capper-surfaced sides. If the ordering is flat, the experiment dies and the touts go back to being nothing more than a source of angles to test against our own 18,000-game database. If the ordering is good everywhere, we've built a general baseball ranker, which is a different and much harder thing to trust than it looks.

There is also a cheap kill switch built into the design. The control arm accrues roughly a dozen plays a day against the fire pool's two. If the panel cannot rank several hundred ordinary games, it will not rank fifty fires, and we can retire the whole idea months before the pre-registered gate would have told us.

Honesty notes: the samples on this page are tiny and mean nothing yet — they are published from day one precisely so the record can't be curated later. Scores are computed before first pitch by a leak guard that refuses to touch a game that has started. Prices are the real Polymarket ask at score time, not a line-shopped best. No money follows any of this: there is no code path from the panel or the capper tracker to a bet, and the pre-registered fire experiment is reported separately so this page can never inflate it. Nothing here is betting advice.

← All research