1<!-- Author: Simon-Pierre Boucher — simon-pierre.boucher@uqo.ca -->2# 🧠 Beyond Text: Cache Compression and Knowledge Transfer in Multi-Agent Systems34**UQO Working Paper No. 30** — *A Reproducible Study of Fidelity, Communication Cost, and Repeated Handoffs*56> **TL;DR** — Under which task, model, cache, network and reuse conditions does each communication representation (raw text,7> summaries, structured evidence, lossless codecs, learned prompt compression, native/quantized/evicted KV caches, learned cache8> mappings) preserve useful information at the lowest *measured* end-to-end cost? Real experiments on Apple Silicon (MLX) with9> Qwen2.5-3B / Llama-3.2-3B, genuine two-node cache transfer over a 1 GbE LAN, repeated-handoff chains with an evaluator-only10> fact ledger, fan-in with conflicting evidence, a same-family learned KV mapper, hosted-API receivers, and a development-fitted router.1112## 📌 Status13See `STATUS.md` (phase gates), `DECISIONS.md` (dated decisions), `FINAL_REPORT.md` (completion status and limitations).1415## 🗂 Repository layout16```17protocol/ hypotheses.md (registered), analysis_plan.md, frozen.json (protocol hash), run matrices18literature/ sources.csv (34 verified sources), audit_group{1,2,3}.md, method_notes.md19configs/ models, data, resources (budgets/timeouts), pilot/core experiment configs20src/wp30/ backends (MLX), cache adapters (sparse/quantized/serialize/selection), methods (A0–A8), runner, chains, fanin,21 transport (two-node worker), hosted (API track), mapper (Track C), router (E8), analysis, CLI22tests/ cache correctness (12), leakage (3), scoring/data (5)23data/ manifests (splits, licences), generated/docs (agent-visible), generated/keys (evaluator-only), raw/longbench24runs/raw per-attempt JSONL records (schema wp30-attempt-1), chains/ hop messages ; runs/logs ; runs/checkpoints (mapper, router)25results/ tables/ figures/ provenance.json claim_ledger.csv26paper/ LaTeX source + compiled PDF27reproduction/ requirements-lock.txt, minimal.yaml, report.json28```2930## 🚀 Quick start (analysis-only reproduction, no inference)31```bash32uv venv --python 3.12 .venv && source .venv/bin/activate && uv pip install -r reproduction/requirements-lock.txt && uv pip install -e .33make reproduce-analysis # rebuilds results/tables and results/figures from runs/raw34make paper # compiles paper/main.pdf35```3637## 🔁 Inference reproduction (Apple Silicon, ~15 GB of checkpoints from Hugging Face `mlx-community`)38```bash39python -m wp30 doctor --output evidence/environment.json40python -m wp30 data prepare --config configs/data.yaml && python -m wp30 data validate --manifest data/manifests/splits.json41python -m wp30 models probe --config configs/models.yaml42python -m pytest -q # cache correctness suite runs a real 1.5B model43python -m wp30 run --config configs/core.yaml --resume44python -m wp30 reproduce --profile reproduction/minimal.yaml45```46Two-node experiments need a second Mac with the same checkpoint revision: `python -m wp30 worker --port 9430 --model <repo>` on the47receiver node, then `python scripts/e6_network.py --host <ip> ...` on the sender node. Hosted-API experiments need `ANTHROPIC_API_KEY`48/ `OPENAI_API_KEY` and are capped by `configs/resources.yaml`.4950## 📄 Citation51Boucher, S.-P. (2026). *Beyond Text: Cache Compression and Knowledge Transfer in Multi-Agent Systems*. UQO Working Paper No. 30.52