Observed work, with its provenance attached
One problem.
Many small bets.
Your browser is live. The shared ledger is published periodically after verification. These two clocks are different, and we show both.
The volunteer ledger
Loading…- Submitted tasks
- —
- Verified unique tasks
- —
- Duplicate tasks
- —
- Verified solutions
- —
Fetching the published ledger.
Actions runs on an hourly schedule, which GitHub may delay. Verification has a bounded budget; a busy queue can take longer. Full negative replay duplicates compute, so more volunteers do not automatically produce linear speedup. Unverified receipts do not certify coverage.
How the next tasks are chosen
Loading the published strategy…
81 fixed contexts: ideal class × coefficient shape × norm shell × quotient band. Cell brightness encodes allocation weight, not probability of discovery. Select a cell for the actual measurement.
No selection.
Keep exploration in the loop.
Every 64 verified unique tasks triggers an epoch. Replayed work supplies cost observations. Sparse observations keep the policy conservative.
The old Mac model passed a held-out throughput test. Our separate discovery-learning experiment did not beat its acceptance gate. Those are different claims.
Read the failed experiment →No community calibration epochs recorded yet.
People at the table
Join them ↗| Contributor | GitHub submitter | Verified computations |
|---|---|---|
| Loading contributions… | ||
Leaderboard credit goes to the GitHub issue author, not the handle typed into the form. A replay proves the task result, not who used a processor or for how long. Copying or submitting a task twice earns no extra unique credit. Paper authorship is not promised.
The original Mac campaign
A separate native C/PARI search. Public snapshots preserve the observation time.
Did the Mac’s cost model predict the measurements?
Frozen release holdout: the first two observations per context trained the model; the third was held out. Both axes are logarithmic. The diagonal means predicted equals observed. These are throughput scores, not discovery probabilities.
Volunteer and Mac work may overlap. Their counters are never summed as a globally unique search volume. A recent status snapshot is not a direct connection to the Mac.