State of the Lab — week ending Aug 23, 2026

This week the lab ran 102 seats across 6 desks and settled 25,375 trades, nearly all simulated and one desk of them real. Best seat: Market-making · SET. Roughest seat: Fade crypto 15-min whales @ 66–85¢.

The live desk gave most of its winnings back this week. Book A had been up
about fifty dollars on Wednesday and finished the week at seventeen, losing
7.69 over the seven days. The pair completion lane lost another 7.02 and is
now down 7.28 since it started. The Dota lane took its first ten real fills
and made thirty cents, which is a rounding error on a rounding error but is
the first live money that lane has ever handled.

None of that is a surprise and all of it is the point. A book earning one or
two cents a fill has weeks like this by construction, and the only honest way
to publish a small edge is to publish the weeks where it does not show up.
The number to watch is not this week. It is whether the per fill figure holds
across months.

The more useful thing that happened is that we found we had been trading the
wrong contract.

The lab runs a lane that watches Valve's official live game feed and backs a
heavy favorite once a game is clearly ending. It matched markets by team
name. Team names appear in two different Kalshi contracts for the same event,
and it took the first one it found, which was the wrong one. KXCS2GAME and
KXDOTA2GAME do not settle on a game. They settle on the whole match. The
contract that settles on a single game is the MAP one.

Across 793 settled Counter Strike events we recorded, a team that wins map
one goes on to take the match 81.8 percent of the time. So the moment a team
has the current map locked up, the map contract is worth a hundred cents and
the match contract is worth about eighty two. They cost the same. Anyone
acting on something happening inside a live game, through the match ticker,
is paying full price for roughly four fifths of the edge they think they
have, and nothing about the fill looks wrong when they do it.

The paper lane had sixteen fills at plus 6.63 cents each, which looked like a
working strategy and was measuring something nobody designed. Those sixteen
have been relabelled and the idea has to earn its number again on the right
instrument. The live twin was caught at zero fills, so no real money was
spent on the mistake, and it now backs a team only when winning the current
game actually ends the match. That is best of one always, best of three at
one nil up, best of five at two nil or two one. Fewer chances, all of them
honest.

The weather experiment is closed and published in full. Fourteen days of
forward testing, 504 settled Kalshi markets, 660 on Polymarket, seven
different strategies. Every one of them loses. Ensemble divergence loses at
every threshold and loses worse the more confident the models are, because
the models are well calibrated right up until they become confident and then
they fall apart, and the market already prices that asymmetry. The one idea
that looked real, entering at market open rather than same day, showed plus
4.6 cents on a fifty five trade backtest and minus 11.5 on a hundred and
twenty three trade forward test. That gap is the whole method. When a
backtest and a forward test disagree, the forward test wins.

A reader also handed us months of history we did not know we could reach.
Kalshi moved older data to a separate set of endpoints back in February, and
somebody in the thread simply told us the URLs. Our own recordings start in
mid July. The archive goes back to February. Every study on this site was
written on five weeks of data and can now be rerun on five months.

The bottom line for the week is another large red number and it has the same
shape as always. The whale grid settled ten thousand simulated trades, paid
313 dollars in fees to lose 267, and accounts for most of the lab's loss and
almost all of its friction. It is an expensive control group, running on
purpose, in simulation, answering a question we already know the answer to.
Everything else is the lab finding out what does not work, cheaply, in
public, and this week it found out twice.

Each desk name links to The Seats, which gives every strategy's real trigger and threshold plus the Kalshi tickers it trades, enough to run any of them yourself. The one exception is market making: which markets it quotes is listed, how it quotes them is not, and the About page explains why.

All 153 seats behind these five lines, with win rates, fee drag, all-time results and the per-trade CSV, are in the members edition. 10 seats retired to date; their records stay on the board.

The market-making seats are simulated fills, and a simulation cannot model adverse selection: the tendency of a resting order to be taken exactly when someone better informed wants the other side. Where this lab has run the same strategy with real money, the live result has come in materially below the paper one. Read the maker desk as an upper bound on what the idea could earn, not as what it would have earned. The live desk below it is that check running in public: same ideas, one contract at a time, settled by the exchange.

Fill verification: 2,596 of 39,921 trades (7%) were checked against real order-book depth.

The rest are entered at the ask the exchange reported, but the lab does not record the book for every market. Challenger tennis, MLB, soccer, basketball and football have no depth archive, so those fills cannot be re-checked after the fact. Trades whose price level had no size behind it are discarded, not counted.

The archive. The lab records and permanently keeps the raw market data behind every number it publishes: trade tape from 2026-08-15 (8 days); order book depth from 2026-07-28 (26 days); top of book ticks from 2026-07-27 (27 days); 15-minute market quotes from 2026-07-21 (33 days). 1.1 GB and growing nightly. None of it can be recreated after the fact, which is why it gets recorded. It stays in the lab: the exchange reserves its own market data and this lab does not pass it on. What gets published is what the lab finds in it.

Every trade above was logged live, then sealed: each night the lab hashes its raw market archive and trade ledgers into an append-only chain. Rewriting any past trade would change every head after it, so check this one against last week's issue.

Chain head as of 2026-08-23T04:13Z: eea5ea840b269550dc6e929b5779a71602e7145f587fc96f7686419c49432410