Definition

Beyond open weights — full stack for reproducibility and inspection.

A Totally Transparent LLM release includes:

Model weights and intermediate checkpoints (or public tombstone of what is missing)
Training code, hyperparameters, seeds
Training data release or precise composition + lawful access path
Metrics, logs, and eval methods
Cryptographic manifests (SHA-256, Merkle roots, signatures)
Basic ttlink — human-followable links from outputs to sources

Not TTLLM

Open weights with closed data and closed training code
Marketing claims without re-hashable artefacts
Paywalled verification of public-core claims

Prior art

We stand on Ai2 OLMo / OLMoTrace, LLM360, and related process-open work. Credit them. Differentiate with artefacts: free-core covenant, org transparency domains 1–10, and usable ttlink.

Manifesto in repo →