Home

kompress Research Wiki

Welcome to the wiki for the longrun-eval-kompress project — the ICLR 2027 manuscript on the Voting Ensemble Paradox and the 3.0× asymmetric loss modulation fix.

Pages

• Home — this page
• Voting-Ensemble-Paradox — the theorem, the proof, the intuition
• Three-Mechanisms — A (3× penalty), B (regex override), C (self-labeling)
• Version-Lineage — all 17 models (v2–v17), what worked, what didn't
• Baseline-Comparison — kompress-v8 vs TextRank, LLMLingua-2, random eviction
• Pareto-Frontier — the λ-ablation (3×→5×→10×)
• Loop-Engineering — Osmani, Greyling, LangChain, LoopKit
• Compute-Costs — $37.19 + $1.76 = $38.95
• Genesis-Seal — the honesty loop
• Interactive-Sites — marimo WASM site + notebook
• Taskfile — 16 task commands for paper, site, CI
• Nix-Flake — reproducible dev shell with treefmt + pre-commit hooks
• Template-Guide — using this repo as a loop-experiment-researcher template

Quick links

• Interactive site: https://peterlodri-sec.github.io/longrun-eval-kompress/
• Repo: https://github.com/peterlodri-sec/longrun-eval-kompress
• Model: https://huggingface.co/PeetPedro/kompress-v8
• Blog: https://pocoo.vaked.dev
• Engine: https://github.com/peterlodri-sec/loopkit
• Dataset: https://huggingface.co/datasets/PeetPedro/ultrawhale-dogfood