Open a funded run
A training run defines its rollout policy and reserves a reward pool. Participants can inspect the context in which contributions will be accepted.
Training coordination on a shared ledger
Coordinate the training. Account for the contributions.
Crucible connects off-chain training runs, rollout commitments, reward pools, and published model artifacts through Sigil.
CrucibleIllustrated study
Input / work / result
Work happens off-chain. Its agreement leaves a record.
01 / Built for a purpose
Many contributors need a common record.
Distributed training can involve coordinators, rollout workers, verifiers, and model publishers. Crucible records the agreements and commitments between them while the training itself happens outside the chain.
Commit batches of work
Rollout batches carry artifact commitments and metadata. The record connects the contribution with its worker and intended run.
Attest and distribute rewards
Supported reward attestations release bounded payouts from the run’s pool under the configured policy.
Publish the resulting weights
Weight publication records bind model artifacts and Merkle commitments to the run. Large weight files remain off-chain.
Close with an accountable outcome
Close, cancel, and slashing operations record the run’s outcome and remaining obligations rather than leaving an unbounded coordination process.
The details matterGood questions.
Good questions.
Clear answers.
Does every validator train the model?
No. Validators process the required commitments and state transitions. Training and rollout generation run off-chain.
Does a published checkpoint prove the model is good?
It binds an artifact to the recorded run. Quality, task performance, and other model claims need separate evaluation evidence.
Keep exploringOne capability.