# Decoding records > The most literal way to learn from a strong board: rebuild it exactly, piece for piece, until your search can reproduce it. What the reconstruction forces you to add is the ingredient your search was missing, and the reproduction is the proof. - Canonical page (with interactive figures/demos): https://eternity2.dev/research/build/learning/decoding-records/ - Updated: 2026-07-16 - Topics: backtracking, learning - Source: Community record boards (this project's records timeline): the witnesses a replay reconstructs — https://github.com/raphael-anjou/eternity2/blob/main/web/content/research/records.mdx --- Corpus priors and anti-pattern mining learn a *statistic* from many boards. This technique learns from a *single* board, and it learns the hardest thing to see from a distance: the one move an ordinary search cannot make. The method is brutally literal. Take a known record board, and try to make your own search rebuild it exactly, piece for piece. Whatever you have to add to your search before it can walk that path is precisely the ingredient it was missing, and the exact reconstruction is the proof that the ingredient is real. ## Reproduction as proof The setup is a depth-first search run in a deliberately constrained mode, aimed at a specific witness board rather than searching freely: - **Prior-over-cost ordering.** An ordinary break-tolerant backtracker ranks candidate placements by immediate mismatch cost, cheapest first. A replay flips the priority so that the placements the witness board actually used outrank cheaper-looking ones. The search is pulled down the known good path instead of wandering off it. - **An exact tail.** The last handful of cells are solved exactly rather than heuristically, so the endgame an ordinary run fumbles is closed deterministically. - **The relaxation under test.** With the ordering and the tail fixed, you vary the one rule you suspect is the barrier, and watch for the moment the witness rebuilds. This is the [REPLAY](/research/lab/experiments/raphael-anjou/learning/replay) experiment, and what it found on the community's strict-460 boards is a clean example of the payoff. The project's own break-tolerant search had saturated at 457 or 458 no matter what, and the reason turned out to be a single rule nobody had questioned: it allowed at most *one* mismatched edge per cell. The record boards each contain four or five cells that pay *two* mismatches at once. A search capped at one break per cell literally cannot represent those boards, so it was fenced off from the very targets it was chasing. > **[Figure]** Interactive: the double-break cells — interactive: DoubleBreakDiagram. Rendered on the canonical page (link above); not shown in this markdown export. > **[Figure]** Interactive: replay the break schedule — interactive: DoubleBreakLab. Rendered on the canonical page (link above); not shown in this markdown export. Raise the per-cell budget from one to two, and both community strict-460 witnesses rebuild exactly, every piece in place, the score checking out edge by edge. The reconstruction *is* the finding: a search-completeness result dressed as a record attempt. The wall was in the move set, not the compute. ## What decoding can and cannot buy Decoding a record is the sharpest diagnostic in this family, because it turns a vague suspicion ("our search is missing something") into a specific, checkable claim ("it cannot pay two breaks at one cell"). It is also the most candid about its limits, and in the same breath. Recovering the double break lifts the strict ladder to 460, but it does not, by itself, say whether two breaks per cell open a path to 461 and beyond or merely to the known 460s. Decoding tells you the ingredient a witness used; it does not promise that ingredient carries any further than the witness did. That gap, a learned move that reaches the plateau without proving anything past it, is the theme of the [collapse page](/research/build/learning/when-learning-collapses). ## Related - [Anti-pattern mining](https://eternity2.dev/research/build/learning/anti-pattern-mining) — The subtle idea in learning from strong boards: not every agreement between them is good. Some shared placements are real structure; some are a shared trap that caps every search just short of the top. Separate the two, and attack the trap. - [When learning collapses](https://eternity2.dev/research/build/learning/when-learning-collapses) — The sobering half of learning from strong boards. A learned signal reaches the top of a search's own range reliably, and then stops. Over-trust it and the search collapses; even used perfectly it does not raise the ceiling, because the ceiling is not a thing the corpus knows. - [REPLAY](https://eternity2.dev/research/lab/experiments/raphael-anjou/learning/replay) — Rebuild the community's strict 460 boards exactly, and in doing so discover the move ordinary solvers can't make: paying two mismatches at a single cell. - [Records & solvers](https://eternity2.dev/research/records) — Eternity II has never been solved, but fifteen years of community effort have pushed the best board to 470/480. Who holds what, how they did it, and why some headline "480" boards are not actually the real puzzle. - [No-good learning: remembering why you failed](https://eternity2.dev/research/build/reduce/nogood-learning) — A failed subtree is a theorem: this partial state can never extend. Store it and never re-enter. The community tried both flavours: chess-style transposition tables over the search frontier, and mined constraints about the puzzle itself. The full ledger of what memory buys at E2 scale, and the small boards where it genuinely pays.