120 positive-EV prices cleared the model. Two cleared the bar.

Most desks publish what they bet. The more useful document is what they didn't.

Today our engine flagged 120 prices where the model's fair value sat above the market's. All 120 were positive-EV on paper. None of them were bet. Two plays cleared the bar instead.

What we found

The defining property of the 120 is not that they were long. It is that all 120 were live, in-game prices — quotes taken while the game was still being played, off a game state that changes with every pitch. That single fact is why the flag count and the bet count diverge.

Three from the set, all live:

- HOU @ SF, SF ML at +953. Live, top of the 9th, score 2-1. Model probability 12.3%, EV +29.87%, 3 stars. Engine would not bet. - MIL @ SD, Under 6.5 at +125. Live, top of the 11th, score 3-2. Model probability 57.6%, EV +29.6%, 4 stars. Engine would not bet. - PIT @ MIA, PIT ML at +1076. Live, bottom of the 3rd, score 1-5. Model probability 10.8%, EV +27.17%, 3 stars. Engine would not bet.

The MIL @ SD line is the instructive one. A 57.6% model probability against a +125 price is not a lottery ticket; it is close to a coin flip priced as an underdog, and it still did not clear. The filter is not screening out long odds. It is screening out prices we cannot trust to still exist by the time we act on them, and a top-of-the-11th total in a one-run game is the purest example of that. The stated edge is real only if the quote is real at the moment of execution, and in-game markets move between the read and the click.

What we took instead, on Kalshi: CHC @ WSH, WSH ML. Ask 43 cents against a fair value of 44 cents — an edge of 2.3%, on a 55.6% model probability, in our proven lane. One qualifying take on the exchange, two plays in total for the session. The same game also carried a VELO_DROP signal-card play.

What it means

The headline arithmetic is uncomfortable in the right way: a 2.3% edge got the money and a 29.87% edge did not. That is the whole thesis of an execution filter. EV computed against a stale or unreachable price is a number, not a position. The 120 are what the model *sees*; the two are what the desk can actually *transact*.

The honest limitation: this is a single session's slate, not a study. 120 flags and 2 takes tells you about today's filter behaviour, not about a long-run hit rate, and we are not presenting it as one. Whether the live-price screen is correctly calibrated — whether it is passing on edges that were in fact executable — is a question that needs a logged sample of attempted fills, not one day of flags.

What would change our mind: a tracked sample of live in-game quotes showing we could have filled at or near the flagged price. If those fills are real, the filter is too tight and the 120 deserve a second look.

21+ where legal.