FORGE

The question

Performance claims become difficult to assess when configuration, data, checkpoints and comparison controls are disconnected.

How it works

• Bind runs to explicit experiment contracts, artifacts and a local registry.
• Export and verify portable evidence bundles, then fail closed when required release evidence is absent.
• Use toy CPU training/recovery and synthetic paired fixtures to exercise the mechanics before accelerator studies.

Evidence & results

• One same-host tiny CPU checkpoint/resume exercise reproduced its uninterrupted loss trace and final checkpoint bytes.
• The paired A100 analyzer exists and is tested on synthetic fixtures.
• Synthetic CPU data-filter experiments preserve output digests and show workload-dependent timing tradeoffs.

Scope & next steps

• No measured A100/Gaudi training, accelerator throughput, memory or model-quality gain.
• The tiny model is a correctness exercise, not a useful language model.
• The broader release and independent learning acceptance gates remain open; source remains local.