Gambit
Researching. Nothing is trading.Event-contract research desk for Kalshi, a CFTC-regulated exchange. One machine in Houston.
About 3,052 recipes a day since 11 Sep, each scored on data it never saw. Almost all fail. That is the point.
Series 2 NAV per unit, from 2026-09-28. 1.00 is the high-water mark. Daily snapshot, reconciled to the exchange.
The desk, last hour
- just nowPaper pass on pre-registered rules
- 2 min agoDesk health check
- 2 min agoRanked series put through the frozen protocol
- 5 min agoDesk feed rebuilt for the agents
- 5 min agoPositions reconciled to the exchange
- 5 min agoLedger settlements checked against the exchange
- 6 min agoEvery rule in testing ran on paper against the venue
- 12 min agoHistory stage ran on new or changed rules
Each line is a file that moved. A job's own report that it ran is not counted.
From idea to trading
Ideas on the left, brute-force recipe search on the right. Both end at the same two tests. Nothing trades until one is passed. Counts since 11 Sep.
Side inflows: "older tests" are three tests registered by hand before the Skeptic step existed; "copies" are a passing rule moved unchanged onto similar markets (the Bitcoin rule onto Ether, gold and silver). Testing counts are rules still being tested; finished ones are listed under each. Band widths grow with the square root of the count, so small stages stay visible.
How each path works
Loading
Who moves each stage, and how often
Open forward tests
Milestones
The steps from a tested rule to the goal, in order. Each one waits on the one before it.
- ·Loading
Method
Every rule is frozen before it sees the data that judges it.
The rule, the window of days and the pass mark are written down first. It is then scored once on untouched out-of-sample and earlier periods at one contract, taker fees included, held to settlement.
Every rule tried is counted, and the bar rises with the count.
Trying thousands of recipes makes a few look good by luck. The trial registry holds every attempt, and a winner in a heavily searched market must clear a higher t-stat.
One event is one independent outcome, not one trade.
Twelve city highs on the same day share one weather pattern; several trades in one game are one game. Sample sizes count events.
Code trades. No language model places an order or holds a key.
A strategy goes live only when code passes it, at one contract per order, inside fixed daily and weekly loss stops. Money figures on this page come from the desk's reconciliation against the exchange; research figures come from its own registry.
Series 1 lost $106.68 and closed at 0.7706.
Series 1 ran 30 Aug to 25 Sep 2026 and lost $106.68 settled after fees, reconciled market by market against the exchange's fills and settlements. It closed at 0.7706 per unit. Its units were rescaled by that figure into Series 2, so each holder kept the same share of the fund, and the high-water mark restarted at 1.00.
Invite-only. Not an offer to the public. Not registered with the SEC, CFTC or NFA. You can lose all capital. Past results are not future results. Prospectus. Research figures as of 2026-10-09 20:11 UTC.