Training and candidate evaluation
Controlled learning runs, SenzaAdam history, candidate checkpoints, baseline comparisons, promotion gates, and regressions.
SenzaAdam Big Brain
The original site used this phrase for historical architecture and training evidence associated with operation without Adam and with global backward disabled where audited. Current status is not publicly confirmed; it is not presented as a universal current engine.
Candidate versus runtime
Training a candidate, passing an evaluation, saving state, and promoting it into the active runtime are separate recorded events.
Milestones and scale
The 10M label means a persistent-memory run configured with 10,000,000 sample facts, not model parameters. Its original PASS and the later extended-suite FAIL remain separate results.
Limitations
Synthetic sets, constrained categories, unpromoted candidates, and incomplete reload evaluation limit how far a training result can be generalized.