One pinned agent
Prose is the only edge
Miners write the strategy in prose, the validator runs it on hidden bugs from real Python repos, and the best gain over the baseline takes the crown
- hidden bugs a round
- 6
- submission window
- 2 h
- rounds pooled for pay
- 8
Round —
connectingKings of the season
most crowns, all rounds
Loading…
The crown, round by round
bar height is the king's gain over the baseline
How a round runs
the next round opens as soon as one closes
Open
6 hidden bugs from real Python repos, with practice bugs to study
Window 2 hours
Strategies are uploaded privately, each PR holds only a signed commitment
Evaluate
Every strategy and the baseline run on the same bugs in a sealed sandbox
Crown
The best gain over the baseline wins, a tie keeps the king
Publish
Bugs, scores and losing strategies go public, the king's runs become training data
What every strategy runs on
- Agent
- Hermes agent v0.21.3, one pinned commit
- Model
- SparkHermes-27B-v1 NVFP4 (era e1), pinned revision
- Engine
- SparkInfer 0.6.5, 64K context, speculative decoding
- Sampling
- temperature 0.7, top-p 0.95
- Sandbox
- no network, 600K tokens per episode
Compete
practice bugs to a sealed entry
- Get the practice bugs
uv run python -m sh.cli.miner tasks - Write the strategy
A
SOUL.mdplus optionalskills/, prose only - Submit privately
Signs, uploads and opens your PR
uv run python -m sh.cli.miner submit --bundle DIR --key HOTKEY --checkout . --head-owner YOU - Race
Sealed at the close, the king's PR is merged and paid
- Defend
A king is paid only through a defense PR: submit the same bundle again
Check every result
nothing to take on trust
- At openpractice bugs, the window and a digest of the hidden bugs
- At closehidden bugs,
close.json,crown.json,reveal.jsonand scorecards - Commitments
HMAC(salt, withheld)must match every task - Strategieslosing ones published under
rounds/<id>/revealed/, the king's when dethroned - Paypooled over the last 8 rounds, the crown is per round