Most of this wiki is history: searches that ran, walls that held, records set
years ago. This page is not. There is a community search running right now on
the strict five-clue board, it is short of
machines, and joining it takes about two minutes.
It is run by Benjamin Riotte, who holds the
strict five-clue record
and who opened the effort to everyone in August 2026
(the announcement thread).
The invitation is plain: "the next machine could be the one that finds 466, or
goes much further."
Join in two minutes
Download the worker for Windows or
Linux,
choose how many CPU threads you want to give, and leave it running. It
updates itself, and you can stop it whenever you want. Windows gets a desktop
app, unsigned, so it warns you before it runs; Linux is one command and needs
Python 3, with a prebuilt solver on x86-64 and a source build on ARM. A
Raspberry Pi 5 is enough to contribute. The
dashboard
shows the fleet, the discoveries and your own contribution, live.
The interesting part is not the client, it is the orchestrator behind it. A
depth-first search over the five-clue board is one enormous tree, and the only
way a crowd helps is if the tree is cut into pieces that nobody searches twice.
That is the job the control plane does, and the thread behind this page has
the project describing each piece of it:
- Distributed root allocation. The orchestrator hands out roots, typically
a fixed 3x3 pattern, so each machine starts somewhere of its own instead of
racing the others down the same branch.
- Bounded tickets. A root alone is far too coarse to distribute, so the
unit of work is a ticket: a root, a deterministic offset into that root's
search stream, and a budget, currently capped at two billion DFS nodes. The
cap is deliberate. Without it a machine can sit for months on a branch that
returns nothing, which is the worst outcome for a volunteer fleet.
- Completed-ticket accounting. Finished ticket ids go into a ledger that
every worker downloads before it starts, and a completed ticket is not handed
out again. An audit of 3,458,636 completed tickets found 3,458,636 distinct
ids: no unit of work had been computed twice.
- Canonical board deduplication. A finished board is fingerprinted with a
canonical SHA-256 and checked against the central database. A new board is
registered and credited; a board already in there is recorded as a re-find,
and the first discoverer keeps the public credit.
- Independent validation. Every submitted board is re-scored on arrival, so
a claimed score is a checked score, not a trusted one.
That is the same lesson the community learned the hard way across fifteen years
of distributed attempts: the
compute was never the scarce resource, the bookkeeping was.
This page is not a live tracker. The
dashboard
and the board library carry the
current figures; what follows is a dated snapshot, the totals the project
published when it opened the effort on 30 August 2026, on the strict five-clue
board (all five official clue pieces at their official cells):
| Score | Boards found, as of 30 August 2026 |
|---|
| 463/480 | 7,645 |
| 464/480 | 327 |
| 465/480 | 11 distinct boards |
By that date the fleet had processed more than 8.1 quadrillion search nodes,
was running at roughly 5 to 6 billion nodes per second, and had peaked at 28
machines at once above 12 billion. Every one of those numbers has moved since,
and the 465 count moved within days: more 465s were credited to other
contributors' machines in the first week of September, after a change to the
seed set. Read the table as the starting line of the public effort, not as its
score.
The 465 line is the reason the fleet exists, and it is now the strict five-clue
record on this wiki's records page: the fleet took the line
from Riotte's own 464 of July 2026, which had itself ended Bruno Gauthier's
three-year 460. The counts above are reported as the project reports them. The
boards live in the public board library,
this wiki bundles none of them, and no recount was performed for this page.
For a sense of scale, compare the whole of eternity2.net, the BOINC project
that ran in 2007: something over 10¹⁹ operations in five months, stopping in
the mid-460s under the far looser one-clue convention.
The first thing a new contributor notices is that their own counter and the
dashboard disagree. The worker prints every high-scoring board it observes;
the site credits only boards that are new, and only to whoever got there
first. One contributor running a Ryzen workstation and a Raspberry Pi 5
reported 128 local finds against 40 credits in a day, and reasonably concluded
that two thirds of the machine's time had been thrown away.
It had not. The tickets behind those boards were all distinct: two different
searches, from different roots and different offsets, had converged on the same
finished board, which the fleet can only detect once the board exists.
Preventing that in advance would mean shipping and comparing an enormous number
of intermediate search states between machines, which costs far more than the
compute it saves. Over the measured window, re-finds were about 0.15% of
completed tickets. The board is a side effect; the ticket is the work.
There is a structural reason it happens at all. The solver is steered toward
fertile regions rather than sampling finished boards uniformly, and strong
boards cluster into families. Carlos Fernandez, the
community's board surgeon, took the first two 465s and showed they belong to
one such family: exchanging two groups of border pieces, rotating a third
group and swapping a pair generates 32 boards at 465 out of them. That is 32
boards, not 32 discoveries, and it is the
rigidity wall seen from the other side: pieces
can move without a single mismatch closing.
Not a brute-force screensaver. The project describes a native search engine
combining backtracking DFS with exact pruning, tuned
heuristics, an optimized endgame
stage, and the distributed root allocation above. It is still changing as new
search ideas are tested. The project publishes its own write-up of the
techniques and experiments it runs, which is worth reading next to
this wiki's algorithm pages: two independent
accounts of the same problem, from people who have each spent a long time on
it.
Nothing, which is the point. There is no prize left to win: the $2 million
expired in 2010, and this is the five-clue board, not a Tomy submission. What
you get is a share in a search that has resisted everyone since 2007, a line on
a dashboard, and the chance that a board found on your machine is the one that
moves the record.
If you would rather contribute a finding than a few threads, the wiki takes
those too: see contribute your research.