The lab's named search experiments, one section per researcher. Each is a real run against Eternity II with its idea, its best board, and the questions it left open. Raphaël Anjou's notebook is here in full; the notebook is open to anyone else's.
Experiments are named search runs against Eternity II, each written up with its idea, its best result, and the questions it left open. They are credited page by page to the researcher who ran them, and gathered into a section per researcher. This is one corner of the community's wider research; the methods, records and history live across the rest of the research section.
The full ledger: every named experiment that reached a board, across all researchers, with the score it reached and how that run ended. Rows are discovered from the pages themselves, so this stays in step with them; click a column head to reorder. A score is marked with the convention it is under (matched edges, or the stricter five-clue track), because a raw NNN/480 number means different things under each.
| Rigor | Reproduces | |||||
|---|---|---|---|---|---|---|
| Raphaël Anjoubest 463/480 · 13 experiments | ||||||
| PALIMPSEST | 463/480matched edges | plateaued | Jul 2026 | 4.0 | measured | stochastic |
| PRIOR | 460/480matched edges | plateaued | Jul 2026 | — | measured | stochastic |
| KEYRING | 460/480matched edges | new-basin | Jul 2026 | — | measured | stochastic |
| REPLAY | 460/480matched edges | plateaued | Jul 2026 | — | measured | seeded |
| GAUNTLET | 458/480matched edges | new-basin | Jul 2026 | 0.67 | measured | stochastic |
| CLOISTER | 453/480matched edges | plateaued | Jul 2026 | 0.067 | measured | seeded |
| MIDDEN | 452/480matched edges | plateaued | Jul 2026 | 0.67 | measured | seeded |
| LODESTONE | 451/480matched edges | plateaued | Jul 2026 | — | measured | stochastic |
| LADDER | 451/480matched edges | new-basin | Jul 2026 | 0.67 | measured | seeded |
| MOSAIC | 448/480matched edges | plateaued | Jul 2026 | 0.067 | measured | exact |
| Verhaard reimplementation | 438/480strict 5-clue | Jul 2026 | 0.033 | measured | seeded | |
| BANDSAW | 437/480matched edges | plateaued | Jul 2026 | — | proven | exact |
| STAGED | 436/480matched edges | plateaued | Jul 2026 | — | measured | stochastic |
| Joshua Blackwoodbest 454/480 · 1 experiment | ||||||
| Blackwood's solver, decoded and run here | 454/480matched edges | Jul 2026 | 0.033 | measured | heavy | |
Not every result is a solver, and the table above is only the scored runs. To browse the lab the other way, by what kind of result each page is (the analyses, the theory, the measurements, and the dead ends kept as first-class negatives), see the by-contribution index.
Some are original ideas; some faithfully reimplement a known community technique to measure exactly what it buys. Each write-up says which it is, what it attacked, where it stopped, and what it left open, so anyone can pick up where it ended. And every result is reproducible: deterministic runs ship a script and its exact output; searches that depend on randomness ship the same script plus the board they found, loadable in the viewer and checkable edge by edge.
Every experiment also states the hardware it ran on, in a fixed spec card: the cores, the machine, the time budget, and one derived number, core-hours, that says what the run actually cost. It is mandatory, because a score means nothing without the compute behind it. Where a run is directly comparable it wears a standardized bench badge (one core, a fixed minute budget, restarted from randomized corners); a bigger run is a native run that records its real kit and the best score it reached. That way a from-scratch minute on one core and a datacenter sweep on four hundred cores sit in the same table, each judged against what it actually spent.