The Shape Reader

A loop with a journal for a memory

The previous chapter ended with a retraction and a rule: after the Phase Switch's championship collapsed under an independent check, no result would be kept without a pre-registered bar, and no champion without a fresh-seed replication. The obvious next move was to keep inventing strategies by hand and testing them under that rule. Instead, on April 20, 2026, we handed the search — rule and all — to a loop.

The loop worked like this. A fresh agent — no memory of any previous run — read a shared journal end-to-end: the objective, the baseline to beat, and every experiment already tried, kept and discarded alike. It then proposed exactly one new hypothesis, wrote it as a strategy in the simulator's C engine, and benchmarked it against a fixed opponent pool across three skill profiles. Pre-registered keep rules decided the outcome: a candidate had to beat the reigning champion's mean win rate by a margin beyond the bench's noise floor, or make real progress on the worst matchup, or it was discarded. Either way, the journal entry was written first — before any code was reverted — so failed ideas survived as annotated dead ends rather than being silently rediscovered three sessions later.

The journal was the loop's only memory, and that constraint did most of the work. One hypothesis per iteration meant every result was attributable to one change. Logging discards meant the search space shrank monotonically. And because each agent inherited the full reasoning of its predecessors, the lineage compounded: mechanics found in early iterations became the raw material later iterations composed.

It compounded quickly. The loop was born on April 20; by April 21 it had written 89 journal entries, and entry 089 was kept as champion — a strategy the journal called X188_X186DetectorChaseDenial. We call it The Shape Reader (X188).

What the loop built: a strategy that read the board

Every strategy in the classic field was a short recipe — check a priority, pick a dart. The Shape Reader was a different kind of object: four layers stacked on the same board state, each inherited from an earlier iteration of the lineage, each with a narrow job. What united the layers was that all of them read the shape of the board — how many targets remained unclosed, and whose — before deciding anything.

The base layer was a priority list, run on every dart: score on a number you've closed that the opponent hasn't; failing that, finish any opponent-closed number you're one mark from killing, with a single; failing that, and only if the opponent has closed more numbers than you, chase their highest; otherwise open a new number. The logic behind the ordering is economic. A one-mark finish on an opponent-closed number is the cheapest durable move on the board — two of its three marks are already paid for, and one reliable single permanently kills a lane they could score on. Opening a fresh number is a three-mark investment from scratch. When both are on offer, the half-finished thing is the better buy.

Above the list sat a lead envelope: the strategy kept scoring as long as its points lead stayed within a multiple of the highest number the opponent hadn't yet closed. Early in the game, with 20 still open, that envelope is so wide it effectively always holds — the strategy banks points freely. As numbers close, the envelope shrinks with the board. This is the shape-reading mechanic in its purest form: the same lead that reads as "comfortable, keep scoring" on an open board reads as "enough, go close" on a nearly finished one.

The third layer was a phase gate. When the board entered its closing stretch — only a few numbers left to close, only a handful of marks between the strategy and a full board — the envelope collapsed to zero, and any lead at all flipped the strategy into cover-and-finish mode. The fourth layer was a pattern detector, inherited from the previous iteration: a cheap test for an opponent racing through closes without scoring. When the detector tripped and the strategy was ahead, priorities flipped to denial — finish their almost-dead lanes, chase their highest close, shut the income off. The increment that finally got entry 089 kept was a refinement inside that branch: chase denial specific to the detector's trip.

None of the four layers was novel on its own. Priority lists, lead thresholds, phase gates, and chases all existed in the classic field — the Frongello research had even warned against chasing, and the Shape Reader's chase survived only because three conditions demoted it to a narrow slice of the opening game. What the loop found was the composition. Only the stack won.

Champion at every skill level — within the field it knew

The headline evidence was a full round-robin sweep: 31 strategies , 20,000 games per matchup , both players at equal skill, repeated at 11 skill levels from beginner (0.8 marks per round) to elite (5.6). As of July 2, 2026, The Shape Reader ranks #1 at all 11 skill levels — within that 31-strategy field. Its average win rate against the field rose with skill, from 59.0% at MPR 0.8 to 68.0% at MPR 5.6 : the more accurately both players threw, the more the shape-reading layers mattered.

Shading: win rate 59% → 68%.

Second place at every one of the 11 levels was E12, at 57.8–59.5% — the best of the classic bots, and the strategy the champion's own lineage was seeded from ; the best classic S-bot was S2, at 57.4–59.2% . On its home benchmark — the expanded 14-bot opponent pool the loop evolved against — the Shape Reader's kept baseline was a 60.65% mean and 51.8% minimum win rate, beating all 14 opponents .

Two qualifications, both of which the first version of this site under-weighted. The sweep is a matched-skill result: in asymmetric matchups the stronger thrower wins, and no strategy in this field changes that. And the sweep's field is the 31 strategies that existed in April 2026 — it predates every artifact the rest of this story is about. The old site said the Shape Reader "never loses a head-to-head." Whatever that claim was worth inside its own field, it is falsified on the unified field , and this page does not repeat it.

The sweep saturated, and we called it done

The Shape Reader was kept at entry 089 on April 21, 2026, and no later entry ever replaced it. The project then slept for roughly ten weeks — zero sessions in May and June — before a final loop session on July 1 pushed the journal from entry 092 to entry 115. That session produced zero keeps.

What it produced instead was anatomy. Entry 107 proved that one of the Shape Reader's own endgame-gate clauses was dead code — removing it changed nothing, bit for bit. Entry 115 proved its tie-chase branch was unreachable. The champion had grown by composition, and some of what it carried was vestigial: clauses kept because a bench once improved when they were added alongside something else. Parameter sweeps around its thresholds came back flat. Every structural variation lost or washed. By the loop's own three-phase gate, the search had saturated.

We read that saturation as evidence about cricket. If 115 documented experiments, 89 of them in the champion's direct lineage, could no longer find half a percentage point of improvement, the natural conclusion was that the Shape Reader sat near the ceiling of what a hand-coded strategy could be — maybe near the ceiling of the game itself. That conclusion had a hidden premise: that the loop's search space, its opponent pool, and its inherited doctrine were the right ones. All three were about to be tested from outside.

A confession: the bull tap was wrong the whole time

One concrete flaw deserves to be named here rather than discovered politely in a later chapter, because it shows what "saturated" did and did not mean. When the Shape Reader taps at the bull — finishing a lane it is one mark from closing — it aims HIT_DOUBLE. That doctrine was inherited from the win-finish written in entry 003, the third experiment the loop ever ran, and it survived unexamined through entry 115.

Two independent lines of evidence later proved it wrong. A context-isolated rerun of the whole discovery loop (the subject of the next chapter) derived the opposite rule — take the minimum-miss single at the bull — and measured it as a small but real gain, +0.3pp with no losing matchup . Then the exact endgame solve (see Ground Truth) put a number on the worst case: in a tied endgame one single away from winning, the Shape Reader throws at the double and forfeits 0.542 of a win probability — because a tie goes to the player who closes, the safe single wins on the spot. Eighty-nine iterations of composition never once re-litigated a line written at iteration three. The loop's memory was its strength; it was also how a mistake became doctrine.

What we believed at the freeze — and what came next

Here is where this chapter's history ends, in the first days of July 2026: a champion discovered in two days, unbeaten across the 26 later entries (090–115), first at every skill level against every strategy we had ever written. We believed the Shape Reader was near-optimal cricket. We were wrong in three separate ways, and the rest of this story is the accounting.

When every lineage in the project was finally rated on one Elo ladder, the Shape Reader placed ninth of 27, at Elo 1118 — below every rerun champion from X129 up and far below the neural policies. When an adversary was trained against it and nothing else, the best response beat it 72.1% of the time (2,000 games) — against the strongest policy in the project, the same attack managed barely more than a coin flip. And when its endgame decisions were scored against the exact solve, it ranked 10th of 13 audited artifacts, forfeiting 5.2 win-probability points per game inside the solved slice (52,415 decisions) . Not near-optimal. Not even the best rule-based strategy in its own repository, once the field grew.

The Shape Reader is still, as of July 2, 2026, the best strategy its own lineage ever produced, and the sweep result stands within its stated field. What fell was the inference we drew from it. Saturation of a search is evidence about the search, not about the game — a lesson we could only learn by starting the whole loop again from nothing, which is exactly what happened next.