What this is.
A working catalogue of unusual poker games — the kind that get invented around kitchen tables and only haltingly written down — documented twice over: once as rules anyone could sit down and play, and once as a research artefact, with a simulator and a findings page.
This site is maintained by Jesse. The prose is drafted by an AI tool from Jesse's onboarding notes, table memory, and simulator output, then edited by Jesse where he has time and attention. So far only the 7s, 8s, and 9s pages have been reviewed by Jesse. Every other variant page is labelled as a draft — the AI has done its best to render what's in the 2024 notes, but the notes themselves are incomplete, and inferences from incomplete notes are how misinformation gets started.
Jesse started playing 7s, 8s, and 9s in 2024, with a group that has been playing for considerably longer. Where the site expresses opinions about the game — archetype names, strategy claims, what a trained agent ought to learn — those are Jesse's, distilled through the AI's prose. They do not necessarily reflect how the other regulars at the table think about anything.
The night vs. the research
A real Tuesday night is dealer's-choice across many games. The same 6–7 players sit down with $5 buy-ins and play the whole evening together; the dealer rotates and chooses whichever variant they want to deal, except in the Mold-‘Em window (typically a 45-minute block in the late evening) when the rotation locks into Mold-‘Em with a 10¢ ante and check-raises allowed.
The research wing of this site treats each variant as if it were the only game being played, repeated many thousands of times. This is not how the table actually operates. We model it this way for the same reason any researcher does — to isolate the strategic surface of one variant from the others, so that simulated archetypes, EV measurements, and (eventually) learned policies are comparable hand-to-hand. The 7s, 8s, and 9s findings page is the first deliverable of this framing; other variants will get analogous treatment once their rules pages are confirmed.
There are perhaps a few thousand poker variants that have been given a name. A smaller number have been formally analysed. A smaller number still — a few dozen at most — have ever been studied with a simulator. The variants in this catalogue tend to share three properties: their rules are not obvious, their strategic surface is wider than the canonical games, and the people who play them have strong opinions about how they should be played that the formal study can confirm, refute, or quantify.
Methodology
Every variant in this catalogue is documented along the same axes, in roughly the same order.
- Rules, as canonicalised from the version we have seen actually played, with rule ambiguities resolved in favour of whichever resolution preserves more of the strategic surface.
- A reference implementation of the game engine in Python, with each game phase represented as a discrete state transition. The engine is the source of truth for any edge case the prose rules leave open.
- A small zoo of heuristic agents that play the game in characteristic styles — aggressive, defensive, direction-locked, hate-drafting. These are not optimal agents; they are vertices of the meta-game.
- A trained agent (when feasible) that plays against a fixed mix of the heuristics. The training is usually PPO over a flat-vector observation encoding; details vary by variant.
- A descriptive sweep: a million or so simulated hands with all heuristic agents in fixed proportions, used to derive frequency tables of winning hands, declaration patterns, and per-archetype EV.
- A findings page with whatever holds up under statistical scrutiny. We are conservative about what we promote to a stated result.
What counts as a finding
We promote an observation to a stated finding only when:
- It holds under at least two random seeds spaced by an order of magnitude in sample size;
- It is robust to plausible perturbations of the heuristic agent thresholds (no result that depends sensitively on a single hand-tuned constant);
- It is reproducible from the published seed, configuration, and code at HEAD;
- And the effect size is large enough to matter at the table — nothing is published purely because a p-value crossed a threshold.
What we are looking for
For each variant, the specific questions are listed on its findings page. Across variants, the recurring questions are:
- Which mechanics in the game generate strategic surface, and which are just chrome?
- What does the rock-paper-scissors topology of player archetypes look like, and how stable is it under perturbation?
- Are there mechanics that are net-EV-negative for every player — rule modifications that the table tolerates because they feel exciting but actually just add variance?
- Where does folk wisdom about the game match the empirical results, and where does it diverge?
Who runs this
Unsubscribe LLC, a small New York studio interested in research projects that sit at the intersection of folk transmission and formal study. The catalogue is one of several research lines we maintain in parallel. If you have a variant we should document, write to us.