Kalshi Trading Strategies — Kalmari's Full Methodology, Published

Our flagship. Deep research on politics, AI and economics markets, with strict risk rules: every pick is sized to conviction, capped per theme, and priced against how the market actually resolves. This is the strategy behind the public track record.

Lab ensemble: several models forecast each market independently, pooled by geometric mean of odds with disagreement shrinkage, sized fractional-Kelly. Shadow-tracked until it beats Keel on resolved record — parked at 4 resolved picks (1W-3L), too few to judge.

Not a picker — a position manager. Mechanical price triggers wake a fresh thesis re-check on live picks; its exit/hold calls are scored against what holding to resolution would have paid.

Anti-selection control: the ensemble must take a side on every market it's handed, even at negative claimed edge. Measures how much of the edge comes from picking your spots. Paused; its 42-pick baseline stands.

Keel's pipeline with GPT-5.6 Terra (xhigh reasoning) as the judge — the mid-tier of the Sol/Terra/Luna family, at half Sol's price. Shadow-tracked head-to-head against Fable 5 Keel and Sextant on the same markets.

The original naive audit: one research pass over a category, picks sized by stated EV. Retired — its resolved record (EV anti-predictive, coin-flips and tails bleeding) is what the later risk rules were built from.

Scout plus instrumentation: committed probabilities, entry context, logged rejects, cluster caps. Same selection flaws as Scout — deprecated after the June losses; its data seeded the calibration work.

Keel's pipeline with GPT-5.6 Sol (xhigh reasoning) as the judge: same deterministic candidate filter and risk sizing, only the probability estimate comes from OpenAI's flagship. Ran head-to-head against Fable 5 Keel on the same markets; discontinued 2026-07-25 after underperforming it.

Keel's pipeline with GPT-5.6 Luna as the judge — the budget-tier test. Dropped after its first run: Luna spent 80 seconds, did no research, and echoed the market price on every candidate. Cheap judgment, cheaply refuted.