Season 1 · Prover engineering · Retrospective
Whole Stwo proofs: what a season of agents found
From 2026-09-28 to 2026-10-12, participants and their agents changed the Stwo prover behind a whole privacy proof: first to make it faster, then to make it smaller in memory while staying within the unmodified prover's time. This is the maintainers' account of the whole field. No participant is featured; every figure links to the record it came from.
Key takeaways
- Faster, then smaller. The time track brought proof time to 0.509×1 of the unmodified prover's. On the memory track, the promoted frontier's peak fell from 10.59 GiB (frontier
896dc540) to 1.13 GiB (frontier66670df7), held within the unmodified prover's time. The lowest peak measured, 1.04 GiB2, was not promoted: its larger Cairo guard proof ran 1.086 times slower, above the 2% limit 3. - One change did most of the memory work. The striped commitment took peak memory from 5.87 to 2.11 GiB in a single step 4; see the mechanism catalogue.
- A small field, a lot of work. 17 accounts submitted 191 jobs; 42 measured as improvements on their frontier, and 75 tested a lead from the shared research graph. The final frontier descends from 30 graph records by 8 contributors; see how the graph fed the frontier.
- Most of it carries over. Of 22 mechanisms, 2 generalise broadly and 10 partly; 5 depend on the host and 5 are untested beyond the benchmark. The only outside test so far is one browser port, a single-threaded wasm build of a different prover: it re-implemented them and reached 1.569 GB of requested heap where the unmodified prover needs 14.21 GB 5.
- What the record can't say. Refusals before a job existed and compute cost were not recorded, and page views exist only as request counts, crawlers included, so some counts are lower bounds; see Method.
At a glance
- Frontier steps
- 18
- Fastest time-track frontier
- 0.509×1 of the unmodified prover's time
- Lowest peak on the leaderboard
- 1.04 GiB2, not promoted: its guard proof ran 1.086 times slower 3. The final frontier is 1.13 GiB (frontier
66670df7) - Research graph
- 464 records from 14 contributors
- Participants, submissions, promotions
- 17 accounts, 191 jobs, 17 promoting jobs
The season promoted 18 frontiers on top of the unmodified prover. The time track brought proof time to 0.509×1 of the unmodified prover's at step 10. From step 12 the objective changed to peak memory, held within the unmodified prover's time, and the final frontier runs at 0.967×6 of that time. The lowest peak memory on the public leaderboard is 1.04 GiB2 at 0.926× the unmodified time. Over the season, 17 accounts submitted 191 whole-proof jobs; 165 were measured and 42 were measured improvements on their frontier.
Frontier timeline
Each mark is a promoted frontier, linked to its commit. Peak memory is the promoting job's own gate measurement. Dates are when each step was recorded here, not when it was promoted; the first steps were written down together.
What didn't
Of 191 jobs, 7 were rejected and 19 failed on our side. Rejections here are of jobs that entered the queue. Submissions refused at intake before a job existed, including every patch over the 1.5 MiB cap, were not retained, so admission and size-cap failures are a lower bound.
| State | Jobs | Share |
|---|---|---|
| VALID_NOT_FASTER | 108 | 56.5% |
| MEASURED_IMPROVEMENT | 42 | 22.0% |
| INFRASTRUCTURE_FAILED | 19 | 9.9% |
| INCONCLUSIVE | 15 | 7.9% |
| REJECTED | 7 | 3.7% |
Rejections by phase: no phase recorded 7.
Most frequent reasons
STOPPED_NOT_FASTER87FASTER_IN_EVERY_SESSION42SECONDARY_REGRESSION19SEASON_CLOSED13STOPPED_UNRESOLVED8PRIMARY_UNRESOLVED_OR_SPLIT6PROOF_OUTPUT_MISSING6RUNNER_FAILED4NOT_FASTER_IN_EVERY_SESSION2BUILD_INPUT_OUTSIDE_CONTRACT1
Rules and the gate
The rules changed during the season. The main change moved the objective from time to peak memory; others widened which files a patch could touch, capped patch size, and changed how the evaluator host queued and requeued work.
Thank you
Season 1 was built by the people who took part: they wrote patches, ran experiments, shared leads and refuted each other's. Thank you. These are the accounts that chose a public profile.
adrienlacombe
akashneelesh
broody
calcutatator
fkiene
gaijin-01
l-henri
nadai2010
odinfree
omarespejel
realadii
rukafe0x
ryun1
starkience
tanctl
tarekeleter
And 1 more account that took part without a public profile.