No. 11 · 11 Oct 2026 · All ponds

State of the Lab, week ending Oct 11, 2026

The Heron · 2 min read · 1c91dde0…f11094  2026-10-11T04:48Z

No real money traded this week. That was the plan: October is paper only. But two comments under one of our videos changed what the lab is testing, and they are the best place to start.

The first said: you can't trust paper results, they always fill. That is true, and it is the most expensive lesson this lab has learned. The same rule made $97.13 on paper and lost $16.70 with real money, because on paper you get filled whenever the price touches your order, and with real money you get filled when someone wants out. So this week members got the other half of the record: every real-money fill the lab has ever made, 6,849 of them across 15 strategies, straight from the exchange's own records. They add up to −$76.10, the same record the proof page has carried all along. Paper is a best case. The real fills are the record.

The second comment was better than ours: your edge didn't change, the market did, so test it with blocked hours. We checked the one strategy that made real money. Between 8pm and 4am Eastern it held about four cents a fill in both halves of its run. In US daytime it slid from about four cents to under one, and its last 200 fills were negative in the afternoon. The catch is that daytime was its best stretch early on, so nobody could have blocked those hours in advance. That makes it a question, not a finding. On Saturday night we started a paper copy of the full rule that tags every fill night or day, with the pass mark written down first: night has to be positive, positive in both halves, and at least two cents better than day. The first 100 fills are in, night +$1.70 and day +$2.66. Far too early to read. It is judged at 150 night fills or on 7 November.

We want real money back on. Paper can tell us what to test, but only real fills tell us what works, and October is the pause, not the plan. The next real-money run goes to whichever paper strategy clears its written pass mark first, starting with the night test. It will run the way the one winner did: one contract at a time, a fixed number of orders, and a stop written down before the first one. Every fill goes on the proof page and into the members' ledgers.

The honest rerun of our best paper idea kept losing. Following the perpetual futures price looked like +1.8 cents a trade until we found it was being measured against a 20 second old price. Priced at what was really on offer, it lost $62.88 over 2,124 trades this week, about three cents a trade. The ten taker rules still running on the 15 minute crypto markets lost on every rule again, −$803.15 across 33,655 paper trades. Paying the asking price on these markets keeps costing about what the edge is worth.

Our AI forecaster, Claude with web search giving its own probability for markets across the exchange, has 103 settled. Its forecasts scored 0.251 on the Brier scale, where lower is better. Kalshi's own prices scored 0.162 on the same markets. So far the market is the better forecaster by a distance. It runs to mid November.

The small bright spots are all paper and all small: the last minute rule +$3.13 over 35 fills, the four coin sniper +$2.25 over 108, speech markets +$2.50 over 50. The commodity favourite, which is the winning rule without its distance filter, ran 1,598 fills and finished +$0.25. Flat. That is what you would expect if the filter was doing the work.

Still hunting. Lesson two of the bot course comes out on Tuesday: how to pull months of Kalshi's free price history and test a rule at prices you could really have traded.

This week, on paper

All paper: simulated fills against Kalshi's real prices, one contract, fee included. Paper is a best case. Members get the all-time board, every rule's detail, and every real-money fill the lab has made in the archive.

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 2026-10-11T04:48Z: 1c91dde07be5979bfd32798dbe12eb5df65d1fa11f0a8b8ed66d34a095f11094