Babel II begins with a capable open-weight model. It does not spend its initial budget training a foundation model from zero.
Its first improvement loop is post-training: turn first-party X operations and signed onchain executions into supervised examples, train candidate versions and activate only the one that beats its parent.
01 / X OPERATIONSFIRST-PARTY SOCIAL DATA
Learn from real social operations.
Babel II works inside a real X workflow. First-party briefs, its own drafts and operator edits become labeled traces. Authorized aggregate outcomes evaluate behavior; it does not train on a scraped copy of X.
TRACE / BRIEF · DRAFT · EDIT · OUTCOME02 / ONCHAIN EXECUTIONSIGNED ACTION DATA
Learn from actions with verifiable execution evidence.
For every permitted action, Babel II records the observed chain state, its plan, the signed transaction, receipt, cost and result. The trace captures execution evidence without treating an address as a person's identity.
TRACE / STATE · PLAN · RECEIPT · RESULT03 / SFT + SELECTIONISOLATED GPU TRAINING
Turn selected experience into a successor.
Retained traces become context-to-response or context-to-action examples. Candidate adapters or checkpoints train offchain and face the same time-separated evaluation. Only a candidate that clears the gate replaces its parent.
ONCHAIN / MODEL · DATA · BUDGET · TOOLS · GATESUCCESSION RECORDEvery accepted version publishes commitments to its parent, data and training manifests, aggregate evaluation and promotion rationale.
GENERATED FROM RECORDS / NOT MARKETING COPY