The open frontier of Eternity II in one place: every angle still worth a run, the wall it attacks, what has been tried and where it stopped, and whether it is a newcomer-tractable target or a hard, well-mapped one.
This is the board of open angles: one row per direction still worth a run. Each entry states the question in a line, names the wall it attacks (from why it's hard), points to what has already been tried and where it stopped (in the wall-and-method map and the dead ends), and tags how approachable it is. The tags mean what they say: newcomer-friendly is a target you can pick up without a decade of context; hard, well-mapped is a target that many strong attempts have charted and none has crossed; unattended is an angle nobody is actively working, distinct from one that has been attacked and held.
For the state of the art the score lives on the records page; this page is only the part that is still open.
The canonical track. The best board respecting only the mandatory starter clue has stood at 470/480 since the community reached it, and a designed 480 is a given: the publisher built the puzzle from one solution, so a perfect board exists and only its location is open.
Wall. Rigidity: the record boards are locally frozen islands, and the ten missing edges are on the far side of every structural wall at once.
Tried, and where it stopped. Every method in the wall-and-method map ends against rigidity; the from-scratch and corpus builds top out at 458 to 463, and none crosses to the ceiling. Raw compute alone does not close it either, which is recorded under just throw more compute at it.
Entry tag. Hard, well-mapped. This is the summit of the whole puzzle: a great deal is known about why it resists, which is exactly why a fresh crossing would be a headline.
The strict five-clue line is tracked separately from the open 470 line, because most progress on the open record was made without the clues. The best board respecting all five placements is 464.
Wall. Rigidity again, plus the clue placements narrow the reachable region: fewer boards satisfy all five, so the haystack is smaller but so is the needle count.
Tried, and where it stopped. The 464 is the standing strict-line mark on the clue puzzles page; the project's own strict builds and re-squeezes reach the 460 class but not past it, as charted in the wall-and-method map.
Entry tag. Hard, well-mapped, but a smaller step than the 470-to-480 gap: 464 to 465 is a single edge, on a line where the reachable region is tighter and therefore easier to reason about than the open record.
Brendan Owen published a pair of generated hint-free 10x10 puzzles with their statistics. One (set_1) fell; the other (set_2) has never been solved and remains the community's agreed next rung below the full 16x16.
Wall. The same edge-matching hardness as the full puzzle, at a size a single machine can actually finish. Set_1 took a large but bounded search, so set_2 is a search-engineering target, not a structural one.
Tried, and where it stopped. The full story is on the benchmarks page: set_1 was solved, set_2's statistics are known, and it is still standing.
Entry tag. Newcomer-friendly. This is the one open target sized for a personal machine and a well-tuned backtracker rather than a compute cluster, so it is the recommended first real attempt.
Every record board sits in a basin that MIP proofs show is locally optimal out to a multi-cell halo. The open question is whether any sequence of small, legal moves reaches a different basin at all.
Wall. Rigidity, stated in its sharpest form, and why basin-hopping looks impossible: the step from one great board to a better one is one giant, indivisible swap with no gradient to follow.
Tried, and where it stopped. The rigidity theorem records MIP proofs of halo-optimality across several basins, and the search for a short escape chain between basins failed for a structural reason set out on why basin-hopping looks impossible. The recombine two good boards entry is the same wall from the crossover side.
Entry tag. Hard, well-mapped. A single crossing move, or a proof that none exists, would reshape how every local-search method is judged.
Meeting two partial searches in the middle is complete in principle, but the number of partial boards to store grows about twentyfold per extra allowed mismatch, so the halves stop meeting well before the full 16x16.
Wall. The exactness cost measured directly: the meet-in-the-middle approach is exact but its memory grows past what any machine holds.
Tried, and where it stopped. BANDSAW solved an endgame band to proven optimality and in doing so measured the roughly twentyfold-per-mismatch growth on both sides, so meeting in the middle stops paying at full size; that result is the BANDSAW row of the map. A better representation, or a lossy meeting that stays sound, is the open lever.
Entry tag. Hard, well-mapped. The wall here is quantified, so progress is measurable: any encoding that lowers the per-mismatch growth factor is a real gain even short of a full solve.
Almost every method keeps rediscovering the same handful of basins. In a full method sweep, only parallel tempering reliably produced genuinely new ones, and even it plateaus quickly.
Wall. Rigidity seen as a diversity bottleneck: the constraint is not the score a method reaches but the number of distinct basins it can find, and the standard arsenal finds too few.
Tried, and where it stopped. The diversity survey is written up on the wall-and-method map: adaptive local search, snake placement, and the top of a trained generator's distribution all keep landing in the known basins, and only parallel tempering reliably escapes. Generate boards with a transformer is one of the diversity attempts that did not hold.
Entry tag. Unattended, and open-ended. Unlike the mapped targets above, this one has no charted attempt list waiting to be beaten: a genuinely new diversity mechanism is an idea nobody has planted yet, which makes it the most speculative entry here and the one least constrained by prior work.
The smallest moves that separate the best known boards and the scale at which genuinely distinct partial boards collapse are both a patch of low hundreds of cells. The open question is whether that is a causal link, rigidity following from the area law, or two independent facts that happen to share a length scale.
Wall. Rigidity and the area law. Rigidity is established independently, by MIP halo-optimality proofs and a committed SAT halo; the area law is a distinctness estimate. A reduction of one to the other would tie the two together, but nothing on-site derives it.
Tried, and where it stopped. The area-law page observes the shared scale and links onward to why basin-hopping looks impossible; it does not claim rigidity is caused by the collapse. The rigidity result is exact and region-by-region, and stands on its own without invoking the area law. No page asserts the causal bridge, which is exactly why it belongs here as a question.
Entry tag. Hard, well-mapped. Both endpoints are among the most studied walls in the section; the open part is the edge between them, and closing it either way would reframe how the two walls are read.
Every interior cell keeps 73 to 137 legal pieces, and the puzzle sits at the phase transition of about one expected solution. Both are grouped as "nothing local to prune," but no derivation ties the branching count to sitting at roughly 17 interior colours.
Wall. No forced moves and the hardness peak. The branching count is a per-cell candidate count on the piece set; the peak is a statement about the number of solutions at 17 colours. A puzzle could in principle have one without the other.
Tried, and where it stopped. Prune versus speed and the wall-and-method map group both walls under the same theme, that there is nothing local to prune, which is a thematic union rather than a causal claim. No page derives the branching profile from the colour count.
Entry tag. Hard, well-mapped. The two walls are each measured exactly; the missing piece is the derivation connecting them, which would turn a shared theme into a mechanism.
The move between two record boards is one indivisible cycle of many cells, and every proper prefix scores worse; that shows why local repair cannot cross between basins. The open question is whether the same geometry is what caps constructive DFS and beam producers, which saturate at 458 to 463 for reasons computed only between finished boards, not inside the search tree.
Wall. Sigma-cycles and rigidity, seen on the diversity axis. The sigma-cycle result is an exact statement over every bundled board pair, but it is computed between finished record boards, not about the constructive search tree that produces them.
Tried, and where it stopped. The wall-and-method map documents that from-scratch and beam producers saturate like the others, and the diversity survey found the standard arsenal keeps rediscovering the same basins. Extending the sigma-cycle mechanism to explain that constructive ceiling, rather than only the local-repair one, is unproven and asserted nowhere on-site.
Entry tag. Hard, well-mapped, and close to the diversity question above: a proof that the same cycle geometry governs constructive saturation would unify two independently observed ceilings.
Every claim above is grounded in a page already on this wiki: the walls in why it's hard, the stopped attempts in the wall-and-method map and the dead ends, the benchmark targets on the benchmarks page. If you take one of these on and it moves, the notebook is open and there is a place for the write-up.