Block-A-Mole Reproducibility Package

Public code for "Block-A-Mole: The Sustainability Frontier of Moving-Target Censorship Resistance."

One script regenerates every figure and results table in the paper, with output files numbered exactly as they appear in the paper.

python3 reproduce.py

This runs the full rotation-game simulation suite locally (no cloud access or privileged deployment) and writes, into output/:

artifact	paper item
figure_1.pdf	Fig. 1 — address blocking is a losing strategy
figure_2.pdf	Fig. 2 — the sustainability frontier (phase transition)
figure_3.pdf	Fig. 3 — rotation speed is not the lever
figure_4.pdf	Fig. 4 — session length: (a) curves, (b) frontier
figure_5.pdf	Fig. 5 — robustness: (a) bursty (b) diversification (c) discovery (d) law
figure_6.pdf	Fig. 6 — closed-form frontier β*(n, kₘₐₓ) per censor
figure_7.pdf	Fig. 7 — the censor's optimal budget split
figure_8.pdf	Fig. 8 — the domain-economy operating recipe
table_1.txt	Table 1 — adversary profiles
table_2.txt	Table 2 — closed-form law vs. simulation
table_3.txt	Table 3 — interval frontier (renewal vs. simulation)
table_4.txt	Table 4 — simulator transition rates
output/data/*.json	raw per-experiment data behind each figure

Figures are vector PDF; tables are plain text. Multi-panel paper figures (4 and 5) are emitted as a single composite PDF with the lettered panels, matching the paper layout.

Usage

python3 reproduce.py            # full sweeps, paper settings (~10-15 min)
python3 reproduce.py --quick    # smaller sweeps / shorter horizons (a fast pass)
python3 reproduce.py --outdir DIR
python3 reproduce.py --no-data  # skip writing the raw JSON

--quick reproduces every figure and table but with fewer seeds and shorter simulated horizons, so curves are noisier and the empirical frontiers shift by about the sweep resolution. Use the default (full) settings for paper-quality output.

What the model is

The censor–defender game decomposes into two independent birth–death layers.

• Address layer — num_clear, the number of endpoints with an unblocked IP (clear → blocked at rate λ_a·num_clear, recovery via rotation at rate μ·(n − num_clear)).
• Domain layer — K, the number of live unblocked registrable domains, a capped birth–death chain (mint at λ_intro, burn at λ_disc, buffer kₘₐₓ).

The system is reachable iff num_clear ≥ 1 and K ≥ 1. The dimensionless control is the domain burn rate β = λ_disc / λ_intro. Unless noted, the canonical configuration is n = 8 endpoints, μ/λ_a = 3, buffer kₘₐₓ = 8, target α = 0.95.

Files

reproduce.py        master script: runs experiments E1–E10, then renders every
                    numbered figure (PDF) and table (TXT)
theory.py           closed-form predictions (geometric IP bound, π₀, A, β*)
rotation_game.py    the event-driven (Gillespie) simulator
censor_models.py    calibrated adversary presets (GFW, TSPU, Iran)
requirements.txt    numpy, scipy, matplotlib

The figure ↔ experiment mapping (experiment keys E1–E10) is documented in reproduce.py.

Requirements

Python ≥ 3.9 with numpy, scipy, matplotlib:

pip install -r requirements.txt

Validation, not calibration

This is model-level validation: the closed-form analysis and an independent simulation agree (Tables 2–3), and the qualitative predictions survive stresses the closed form omits — state-dependent discovery, bursty and provider-correlated burns (Fig. 5). Adversary mechanisms are fixed to documented GFW, TSPU, and Iranian designs; the dimensionless presets are deliberately adversary-favorable, so the reported guarantees are conservative. Fitting free rates to a specific censor is a deployment step (against public OONI / Censored Planet archives) and is intentionally out of scope here.

License

MIT (code) / CC BY 4.0 (generated data). Released for open reuse as a shared, theory-grounded testbed for the censorship-resistance community.