Native training
Independent model lineage, controlled training stages, targeted curricula, and iterative improvement.
Native AI research
An independently developed AI research project focused on model training, evaluation, instruction following, runtime intelligence, and the systems required to move from experiments to usable capability.
Research areas
The public page describes the direction without exposing private training data, internal evidence, checkpoints, or development-only implementation details.
Independent model lineage, controlled training stages, targeted curricula, and iterative improvement.
Candidate evaluation and promotion decisions are based on measured behavior rather than version numbers alone.
Research into how trained capability is exposed through a dependable runtime and application layer.
Experiments in instruction following, semantic discrimination, context use, identity, and broader capability.
Native model research progress · August 31, 2026
Orayth continues to use its native model lineage. Stage 4 has not been promoted into the live/default runtime. The public status now reflects the completed FIX9 curriculum and later repair/checkpoint work without presenting an unevaluated checkpoint as a promoted model.
Working runtime baseline
Stage 3.5 remains the active/default Orayth model while Stage 4 research continues behind promotion gates.
Strict score: 0.6875
FIX8 remains the last completed strict Stage 4 evaluation retained for public reporting and did not satisfy promotion requirements.
64,000-record true-holdout curriculum built
The grounding-retention and anti-attractor curriculum uses separated train, validation and test template families and remains the current Stage 4 curriculum foundation.
Repair line continuing
Research-best checkpoints have been observed through later FIX10, FIX11 and FIX12 repair work, but no Stage 4 candidate is being presented as promoted until strict evaluation and promotion gates are satisfied.
Promotion status: Stage 4 remains research-only. Stage 3.5 remains live/default.