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.
Reproduce this result: prose — no computation behind this page
The other four pages in this section are about what learning from strong boards buys. This one is about where it stops, and it is the page that keeps the family grounded, because every technique here has the same two failure modes and it is better to state them plainly than to let a reader mistake a bag of learned tricks for a way past the wall.
Every learned signal in this section is safe only as a tiebreak. It decides between moves that are otherwise equal on the one thing that actually counts, the matched edge in front of you, and it must never outweigh that thing. The moment a learned prior is promoted from tiebreak to part of the objective, the search starts trading a real matched edge against a statistical hunch, and it collapses.
LODESTONE is the cleanest measurement of this. At a tiny weight its scarce-piece-early prior lifts the median from-scratch score from 449 to 451 and tightens the spread. Turn the weight up even slightly and the score falls to 422, then 380, because chasing the corpus's scarce demands then trades directly against matching the edge in front of you. The same shape appears in anti-pattern mining from the other direction: using the trap list as a gentle steer reached 463, but hard-banning trapped placements made boards worse. A learned signal is a good nudge and a bad law. The collapse is not a tuning accident to be engineered away; it is the signal telling you it was only ever a hunch.
The deeper limit is quieter, because it looks like success. Used correctly, as a tiebreak, a learned signal reliably carries a search to the top of its own range and no further. LODESTONE is explicit that it "does not touch the basin ceiling that stops every method near the top." PRIOR and KEYRING reach 460 from scratch, the top of the from-scratch range, but not past it. PALIMPSEST, the strongest of them, reaches 463 and then finds that rebuilding the trapped regions lands back on the same known top board rather than a new one.
The reason is structural, and it is the thesis of the whole section turned around. Every board in the corpus is itself stuck below 480. A signal mined from that corpus encodes where the strong boards are, which is the plateau, so a learned signal that merely re-encodes the plateau cannot lift a search above it. The corpus knows the shape of the basin its boards are caught in; it does not know the way out, because none of its boards found one. What stops every method near the top is the rigidity wall: near a strong board the remaining mismatches lock into a single interlocking σ-cycle that no local move can open. That wall is a property of the puzzle, not of any search's ignorance, and no amount of corpus mining moves it, because the corpus is made of boards that hit the very same wall.
None of this makes learning from strong boards worthless. It makes it precise. A learned signal is the fastest, most reliable way to get a search to the plateau: to build a competitive board from nothing (PRIOR, KEYRING), to reach it in a fresh basin (KEYRING's new corner family), to land at the top of the range every time rather than sometimes stumbling (LODESTONE's tightened spread), or to diagnose the exact move a search was missing to reach a known height (REPLAY's double break).
What it is not for is getting past the plateau, and a section on learning from strong boards that pretended otherwise would be selling the reader the one thing the corpus provably cannot contain. The construct-then-refine pipelines that reach this project's best boards divide the labour cleanly: a learned constructor to reach the top of the range fast, and then a separate reckoning with the rigidity wall, which learning from strong boards can describe but cannot dissolve.