On this page
Current configuration
The initial Phase-I context target is 8,240 tokens across domains.
Earlier 12K/54K domain-conditioned curricula are historical. Longer-context scaling such as 16K, 32K, or 54K is postponed until smaller-scale SGCA has been evaluated.
Why record the decision?
Context affects memory, batching, throughput, and the comparison. The matched baseline should follow the same current configuration. An overview should not silently merge every previous target.
Last updated 01 Oct 2026Discuss this work