Popular repositories Loading

1. PPBExt-Knowledge PPBExt-Knowledge Public

   Modular knowledge injection into frozen LLM backbones. Measured on Qwen3-4B / Qwen2.5-3B / SmolLM2-1.7B: a 64 to 128 sample phase transition, and a 2,560-parameter module form with no measurable dr…

   Python

2. PPBExt-Cache PPBExt-Cache Public

   Application-side prefix-cache orchestration for LLM sessions. Three-zone layout (immutable prefix / append-only log / volatile tail) and three measured caching laws; 99.81% steady-state hit rate on…

   Python

3. PPBExt-Sample PPBExt-Sample Public

   Equal-compute measurement of reasoning-effort knobs. Across 18 task points on three models the highest effort setting was never the best; zero-thinking with 4 samples reached the same accuracy at a…

   Python

4. PPBDec-Core PPBDec-Core Public

   A shared decision sub-layer for LLM applications: decision points typed by kind, solvers along a determinism spectrum, an immutable JSONL decision record, and evidence that merging candidates (F1 0…

   Python

5. PPBExt-Memory PPBExt-Memory Public

   Cache-friendly paged memory for LLM sessions: page lifecycle under prefix-stability discipline, an append-only fold mechanism for staleness, and a hash-chained audit of user deletions that keeps ha…

   Python

6. PPBExt-Repair PPBExt-Repair Public

   Multi-layer repair chain for LLM applications - tool calls, structured output, retrieval queries, memory state and plan progress. Five deterministic layers sharing one three-state discipline (ok / …

   Python