Every model release is a new epoch.
Every epoch is a proof.
$BEPOCH is the infrastructure layer for ML teams — Git-style versioning, structured tracing and on-chain cryptographic provenance for the full model lifecycle, from raw data to production deploy. Built for the DevOps engineer. Verifiable by the regulator.
npx bepoch initThe model lifecycle, as a verifiable graph.
$BEPOCH wraps every step from raw data to deploy in a structured trace. The output of each stage is content-addressed, so the next stage cites it instead of trusting it.
- 01data · pinned
Datasets are content-addressed and snapshotted. Every training run cites a dataset version, not a moving folder.
ds:imagenet-mini@4f1a - 02train · traced
Pipeline CI runs the training graph as code. Each step emits a structured trace — hyperparams, hardware, seed, runtime.
run/2026-05-25T14:31Z - 03eval · scored
Accuracy, fairness, robustness and EU AI Act gates run side-by-side. Failing a gate blocks the release, not just a notification.
eval.report.json - 04sign · approved
A human approver signs the release with their wallet. The signature is over the weights hash, dataset, and the eval bundle.
0xsig_v030_rc1… - 05anchor · on-chain
The signed release is anchored on Base. The on-chain record is the canonical pointer — every downstream system can verify against it.
0x4a2f…6e5
Watch the chain of epochs settle on Base.
Each cube is one model release. Solid links carry verified flow between anchored epochs; dashed links are still-unsigned drafts. The torus at the bottom is the on-chain anchor on Base L2.
A release nobody can rewrite — not even you.
When an epoch is signed and anchored on Base, the record becomes the single source of truth. Auditors, regulators, customers, and your own downstream services all verify against the same digest.
- Verifiable lineageReconstruct any production decision back to its data, code, and approver.
- EU AI Act, by designRequired documentation for high-risk systems is produced as a byproduct of releasing, not as an afterthought.
- Vendor-neutralWeights and metadata stay in your storage. Only digests and signatures go on-chain.
- Token-paid writesEvery anchor consumes $BEPOCH — the protocol fee that funds the network.
Same record. Three audiences. Zero translation layer.
- pipeline.yml drives training, eval, sign, anchor — branch like git, release like an API.
- reproducibility is the default: pinned dataset, pinned config, content-addressed weights.
- rollback to any past epoch in one command — production points at a hash, not a folder.
bepoch init && bepoch epoch push- the EU AI Act technical file is generated from epoch metadata, not assembled by hand.
- every production decision links back to the approver wallet that signed the release.
- tampering is detectable: re-hash the artefact, compare against the anchored digest.
bepoch export --euai-act ep_…- one URL: paste an epoch id, see the full lineage and the matching on-chain receipt.
- no access to weights required — verification is over digests and signatures.
- vendor switches model? on-chain history shows when, by whom, against which baseline.
open bepoch.dev/ep/…The token behind every on-chain record.
$BEPOCH is the protocol fee paid by every anchor. Validators stake it to attest model records; teams pay it when they sign a release; downstream verifiers earn it for catching tampering.
- gas of provenanceevery epoch anchored on Base consumes a small $BEPOCH fee
- staked attestationvalidators stake $BEPOCH to co-sign and lose stake on fraud
- regulator readsfree read access — only writes cost
- governancetoken-holders steer eval gates, fee curve, and supported networks
- network rewards38%
- team & contributors18%
- ecosystem & grants22%
- treasury14%
- public sale8%
Bring your ML pipeline on-chain — without changing how your team works.
Drop the $BEPOCH CLI into your existing CI. Your first epoch anchors on Base sepolia in under five minutes.