site

Interactive paper companion

A single-page interactive essay for Models Take Notes at Prefill: KV Cache Can Be Editable and Composable — transformer-circuits-style visualizations of every study in the paper, driven entirely by the released result records.

Build & view

# 1. (optional) regenerate the curated site data from the result records
python3 data/build_data.py        # reads ../results, ../mem/results, ../e1, ../e2
                                  # writes src/data/*.json and prints a 22-row
                                  # assertion table checking extracted values
                                  # against the paper's stated numbers

# 2. build / preview
npm install
npm run build                     # tsc + vite -> dist/ (static, host anywhere)
npm run preview                   # serve dist/ locally
npm run dev                       # dev server with hot reload

The generated src/data/*.json are checked in, so npm run build works without running the Python step.

Data fidelity

• Every chart reads from src/data/*.json, extracted 1:1 from results/*.json, mem/results/, and the run logs (comp_div_*.log, parsed).
• Prompts in the explorer are regenerated verbatim by the deterministic harness builders (e1/contexts.py, e2/scenarios.py).
• The handful of values that exist only in the paper text (no released record) live in src/data/constants.json with an explicit source field, and render with a "⊙ from paper text" badge.
• Recorded model outputs are shown exactly as stored (tool call, thinking-token count, truncated answer head) — never reconstructed.
• build_data.py exits non-zero if any extracted headline number drifts from the paper's claimed value.

Attribution

Author: Bojie Li (Pine AI). Code and result records: https://github.com/19PINE-AI/programmable-kv. Deployed at https://01.me/research/programmable-kv/.