What does it take to play cricket well?

We spent five months finding out — three different ways. Hand-coded strategies refined across hundreds of automated experiments, a neural policy trained by self-play, and an exact solve of billions of endgame positions. This site is the write-up: what was asked, what was measured, and what a league player can actually use.

Research frozen — results as of July 2, 2026

Three numbers to start with

Elo 1222 As of July 2, 2026, the strongest player we built is The Closer (it3), a self-play neural policy — top of the unified 27-strategy ladder at Elo 1222 over 24,000 rated games for the policy , beating the best hand-coded champion 60.0% head-to-head (1,500 games) .

4.16B Endgame positions solved exactly . Splicing perfect endgame play onto the old champion beats the original 54.40% of the time — +4.4pp over 50,000 games .

+9.4pp The single largest mechanic discovered: faucet denial — shut the opponent's scoring lanes before claiming empty ones — lifted a champion's mean win rate from 55.0% to 64.4% in one step .

The unified Elo ladder: 27 strategy artifacts across three lineages on one Bradley–Terry scale (anchor S1 = 1000), dots at each rating with ±standard-error whiskers. The Closer (it3) tops it at Elo 1222, as of July 2, 2026.

One question, three instruments

Darts cricket looks simple: close seven numbers, keep the score in your favor. But under a realistic accuracy model, every turn hides a genuine decision — score, close, or deny — and folk doctrine disagrees about all three. Over five months and 36 research sessions (February 5 to July 2, 2026) , this project attacked the question with three independent instruments: an automated discovery loop that invents, benchmarks, and prunes hand-coded strategies; a reinforcement-learning arc that trains a neural policy from imitation through league self-play; and an exact dynamic-programming solve that computes true win probabilities for the whole endgame.

The three methods check one another. The discovery loop's champions were re-run in a clean-room replication, rated against the neural policies on a single Elo ladder, and audited move-by-move against the exact solve. Where all three point the same way — ties go to the closer, take the minimum-miss single, shut the scoring faucet before mopping up — the advice is as solid as simulation can make it. Where they disagree, that disagreement is itself a finding, and the chapters say so. The synthesis is a single verdict: all three lineages, placed on one Bradley–Terry scale, form a near-total order — a 27-strategy field with a clear top, headed by The Closer (it3) at Elo 1222 . All results on this site are frozen as of July 2, 2026 .

The Story — seven chapters, in order

The story wing is a narrative in research order: each chapter states its question, its method, and its numbers, and each ends where the next begins. Read straight through, it is the account of a champion being crowned, doubted, replicated against, out-learned, and finally graded against ground truth.

  • The Field — the arena: the classic cricket playbook of Frongello strategies and experimental variants, and the simulator that scores them.
  • The False Summit — the first invented champion collapsed under two simulator bugs. The retraction became the replication gate every later result had to survive.
  • The Shape Reader — an autonomous discovery loop ran 89 iterations in two days and produced X188: #1 at all 11 tested skill levels, within its 31-strategy field .
  • The Rerun — the same loop, restarted with a clean context, found a decisive endgame law at iteration 3 that the original never found in 115 entries — and its champion, The Shutout (X181), outrates the original's .
  • The Learner — imitation, PPO, then league self-play: from a 49.3% historical ceiling to the strongest artifact in the project — including one instructive failure by reward hacking .
  • Ground Truth — the exact solve of 4.16 billion endgame states grades every champion's endgame. No one plays it perfectly; the neural policy comes closest, agreeing on 74.0% of 71,961 audited decisions .
  • One Ladder — all three lineages rated on one Bradley–Terry scale, where the 27-strategy field turns out to be, to measurement precision, a total order .

Play Better — what to do with it at the oche

The advice wing distills the research into rules a human can hold in their head mid-game. No interactive advisors, no votes — each rule cites the measured number that earned it a place.

  • The Rule Card — seven rules from the champions and the exact solve, led by the one that is arithmetic rather than advice: ties go to the closer, so on a level score, close .
  • Underdog Play — what changes when you are the weaker thrower: The Grinder (X214_UD) denies at any deficit and shrinks the farming cap, worth 17.63% vs a 15.31% baseline underdog win rate — a 15% relative improvement .

Reference — look anything up, verify everything

The Reference page is the lookup wing: the full strategy catalog on one Elo scale, the benchmark and validation methodology, the raw data files (ladder.json, endgame_audit.json), and a register of every citable fact on this site by stable ID. Every load-bearing number in every chapter links there. If you read one wing with a skeptical eye, read that one.