The POC sticks to a frozen base model, a two-level gist tree, and synchronous updates so contributors can prove end-to-end behavior quickly.
- Frozen base: no fine-tuning during initial loop; LoRA is follow-up work.
- Hierarchy: two-level MegaContext Tree suffices for moderate contexts.
- Update cadence: synchronous ingest/refocus between autoregressive steps.
- Linkages: informed the legacy POC deliverables; retain for historical context while MegaContext PRD Index defines the active requirements alongside POC Architecture assumptions.
Details
Guardrails that keep the proof-of-concept focused and verifiable while we build out the full MegaContext Tree loop.
- Frozen base LLM: start with an off-the-shelf checkpoint; no fine-tuning during the initial loop. Lightweight LoRA adapters are a follow-up once the pipeline is stable.
- Two-level MegaContext gist tree: limit lifetime memory to moderate sizes so two hierarchical levels (LOD0 tokens, LOD1 gists, LOD2 roots) suffice for the demo.
- Synchronous updates: keep the MegaContext Tree resident in RAM/GPU for the POC; ingest → focus → decode happens between autoregressive steps with no background streaming yet.
These constraints tied directly to the original POC milestones; treat them as historical guardrails while the active requirements come from MegaContext PRD Index and inform the performance envelope summarized in Performance Sketch.