Skip to content

Latest commit

 

History

History

README.md

Analysis & figures

Re-runnable analysis behind the paper, written to be read top-to-bottom and exported to a static site. The notebooks are the data + figures; the paper (../PAPER.md) is the prose.

Notebooks

  • wave_analysis.ipynbthe sovereign selection. Quality → Safety → Energy → the quality × safety × energy Pareto → systems context (cost, deterministic precision, the bracket value gate, roofline transfer) → conclusions. GitHub renders it with figures; or build the site below.
  • judge_comparison.ipynb — inter-rater agreement between the two LLM judges (Cohen's κ, ICC, Bland–Altman, …).

Machine-readable exports → ../../data/site/

The wave notebook writes the figures' underlying numbers, so a website (or any downstream tool) is driven by real data, not screenshots. Regenerated in place on every run.

File Contents
summary.json Headline numbers: model/Pareto counts, the 3–4B knee, instruct-vs-reasoning safety, the sovereign pick.
models.csv · models.json Per-model 3-axis table — model, bracket, quality, safety, energy_wh, mWh, pareto.
pareto.csv The Pareto-optimal short-list (non-dominated on quality ↑, safety ↑, energy ↓).
axis_quality.csv Judged %-of-frontier per bracket (mean + 95 % CI).
axis_safety_bracket.csv · axis_safety_arm.csv Deterministic refusal per instruct bracket / per training-type arm.
axis_energy.csv Energy-per-answer + tok/s-per-watt per bracket.

Rebuild / view

# one command: re-run the notebook, write data/site exports, render HTML (+ Quarto site if installed)
scripts/build-analysis-site.sh

# or step by step:
.venv/bin/jupyter nbconvert --to notebook --execute --inplace docs/analysis/wave_analysis.ipynb   # 1. run (needs data/snapshots/)
.venv/bin/jupyter nbconvert --to html --embed-images docs/analysis/wave_analysis.ipynb            # 2a. standalone HTML page
quarto render docs/analysis                                                                        # 2b. full site -> _site/  (needs: brew install --cask quarto)

Honesty: the quality axis is the 5-rep × 2-judge ensemble (cross-judge κ_quad ≈ 0.91); safety and energy are judge-free / measured. Everything is one commodity node (n = 1). See the notebook intro for provenance. Render artifacts (_site/, *.html) are git-ignored; the notebook outputs and data/site/ exports are tracked.

Accessibility. Every figure carries a #| fig-alt: screen-reader description at the top of its cell. Quarto applies these as <img alt="…"> on quarto render (the accessible build). The plain nbconvert HTML preview does not carry them (nbconvert doesn't read Quarto cell directives) — use the Quarto site for the accessible version.