Reference

Every number on this site traces to one fact sheet

This site makes quantitative claims — rankings, win rates, Elo ratings, head-to-head results — and every one of them was transcribed from a single facts sheet, itself read from committed research artifacts on July 2, 2026, the project's freeze date. Each fact carries a stable ID (F-ELO-IT3, F-TB-TIEBLUNDER, …). When a chapter states a load-bearing number, it links that ID to this page; the anchors land either on the row of a table below or on the entry in the fact register at the bottom. If a claim has no fact ID, it is context, not evidence.

Two conventions apply everywhere. First, results are frozen: "champion" means champion as of July 2, 2026, and no page updates that date. Second, every win rate carries its sample size, because a win rate without an n is a rumor.

The catalog: 27 strategies, three lineages, one scale

The table below is the project's roster, ordered by the unified Elo ladder that put all three lineages — the classic catalog, the hand-coded discovery champions, and the neural policies — on one scale (One Ladder tells the story). Ratings are Bradley–Terry maximum-likelihood fits over 253 C-engine pairings (4,000 games each) plus 58 Python-engine pairings (1,500 games each), anchored at S1 = 1000 . As of July 2, 2026, the top of the ladder belongs to the reinforcement-learning policies: The Closer (it3) rates 1222 , about 47 Elo clear of the best hand-coded strategy, The Shutout (X181), which it beats 60.0% head-to-head (1,500 games) .

A scope note on coverage: the classic catalog is larger than its ladder representation. The project's pre-discovery field held 30 strategies — the 17 Frongello S-bots plus experimental E-variants — but only the representatives that later work trained against or audited were rated. Per the naming convention, five artifacts carry style-of-play names ; everything else goes by its research ID.

The fit is honest about its own error: the mean gap between observed head-to-head results and the Bradley–Terry prediction is 5.08pp, with a worst residual of 16.53pp (S10 vs E1: observed 39.1% where the model predicts 55.6%) . And the field is unusually well-behaved — an intransitivity scan over every ordered triple found zero rock-paper-scissors cycles in which every edge clears 52% (≈2.5σ at 4,000 games): to measurement precision, the 27 strategies form a total order .

Selected head-to-heads

A sample of observed pairwise results from the ladder's match data; row beats column at the stated rate. Raw win/loss counts for all 311 pairings are in ladder.json.

Shading: head-to-head win rate 51% → 90%.

Off the ladder: rated by other instruments

Five artifacts matter to the story but do not appear on the unified ladder, either because they compete on a different grid (skill-mismatched play), because no Elo fit was ever committed, or because they were superseded before the ladder existed. Their headline numbers come from their own benchmark protocols and are not directly comparable to ladder Elo.

Methodology: how the numbers were made

Benchmarks and error bars

Two simulators produced every result: a C engine (fast_sim.c) for high-volume tournaments and the original Python engine. Games alternate the first thrower and use the pro/pro skill profile unless a fact says otherwise. The standard cell sizes and their 95% Monte-Carlo intervals :

Key results were replicated under a second seed before being trusted: the ladder ran on seed 42 ; The Shutout's home-pool sweep reproduced under a seed-7 re-bench (64.40% / 50.41% vs 64.40% / 50.44%) , as did The Grinder's underdog result (17.65 / 13.22 vs 17.63 / 13.31) , and the PPO pool mean replicated on an independent seed at 76.01% vs 75.79% .

The ladder fit

The unified ladder is a Bradley–Terry maximum-likelihood fit with anchor S1 = 1000 and scale 400/ln 10, over 253 C-engine pairings (a full round robin at 4,000 games each) plus 58 Python-engine pairings (1,500 games each, 0 stalls). Standard errors come from inverse Fisher information with the anchor fixed. Cross-engine trust rests on port validation: clean-slate champions ported into the C engine were verified bit-identical to their home engine, and Python ports agreed within 1.5pp of C against a common opponent . Independent corroboration: the RL arc's own keep-rule protocol pegged it3 at Elo 1215.9 from different match data — a 6-Elo (≈0.9pp) gap from the ladder's 1222.0 .

Stall accounting

Cricket has a degenerate corner: a player can refuse to close its last targets and farm points forever. The RL arc's first PPO run found and abused it — 13–34% of greedy-eval games deadlocked, and counting those stalls as losses dropped the run's mean to 46.29% . Every frozen result since uses the fix and reports the count: a 900-dart episode cap with stall scored as a loss. The ladder's Python pairings ran 0 stalls ; it2_det ran 0 stalls in 26,000 eval games (0 in 52,000 across two seeds) , and it3 ran 0 stalls ; the exploitability probes ran 0 stalls in all cells . A "0 stalls" note next to a result is that accounting, not a boast.

The tablebase solve and its gates

The exact endgame solve (Ground Truth) covers the slice of cricket with at most 6 of 14 lanes still open and score difference within ±250, both players on the pro profile: 2,767,444 mark patterns × 3 darts × 501 score diffs = 4.16 billion states, 16.6 GB of exact win probabilities, solved shell-by-shell so each openness shell is a usable checkpoint on its own . Four validation gates all passed: mechanical equivalence against the game engine, analytic spot-checks, a boundary-bracket re-solve of the inner shells, and a Monte-Carlo gate of 20 solved states × 100,000 games each (worst |z| = 2.63, 0 fails) . Validation earned its keep: it caught one real bug — biased first-dart RNG seeding under small sequential seeds, fixed by splitmix64 hardening (commit 8b49948) .

The endgame audit

With the solve in hand, 13 strategies were audited: 6,000 games each (2,000 vs each of X188, CS_X181, and S2), with every in-slice decision — 52,000 to 75,000 per participant — scored against the exact win probabilities. "Agreement" is tie-aware (a choice within 10⁻⁶ of the best move counts as agreeing); a "blunder" forfeits more than 0.05 win probability in one decision . Together with the unified ladder, this audit is the evidence base for the neural policies — there is no policy tournament grid.

Shading: tablebase agreement 39% → 74%.

Read the ranking by EV lost, not by agreement: H1_Hoarder agrees with the solve only 39% of the time yet leaks less per game than The Shutout at 67.7% agreement, because most of its disagreements are near-ties . A second instrument from the same solve: hybrid matches, where a champion plays itself except that in-slice decisions are replaced by tablebase-greedy play (50,000 games each, ±0.44pp; plain-vs-plain is 50% by symmetry):

Exploitability probes

How much can a dedicated adversary punish each strategy? For five frozen targets, a PPO best-response was trained against that target alone (warm-started from the champion) and evaluated at 2,000 games per mode, stalls counted as losses, 0 stalls observed . Lower is harder to exploit: the attack barely beats a coin flip against The Closer but wins 82% against E12.

Shading: best-response win rate 53% → 82%.

Data downloads

Two data files ship with this site, both copied verbatim from committed research artifacts on the freeze date. Everything a chapter claims about the ladder or the endgame audit can be recomputed from them.

The previous site's data trees are retained at their old paths for link stability; the chapters on this site transcribe only from the frozen facts sheet, not from legacy JSON.

The fact register

Every citable fact on this site, by stable ID. Chapter links land here (or on the catalog and audit rows above, which carry their own IDs). Statements are condensed; the frozen facts sheet in the repository holds the full wording and the source path for each. All results as of July 2, 2026.

Named artifacts

The unified ladder

Endgame audit

The tablebase solve

Exploitability

The RL arc

The clean-slate arc

The Opus control arm

The AlphaZero cell

The main lineage and the classics

Timeline