Methods and verification¶

How the numbers on this site are produced, checked, and labeled. Read this before quoting anything.

The three gates¶

A quantitative claim ships only when it clears three independent gates:

1. Proof. An analytical statement with a written proof, technique named, peer-legible.
2. Numeric cross-check. A verification script that checks the closed form against brute-force computation. Every proposition in the theory and game-theory papers has one; the theory paper's core theorems are additionally verified on hundreds of random structures.
3. Monte-Carlo robustness. The qualitative claim must survive random parameter draws, reported as the share of the sampled space on which it holds, with the failure region characterized.

One gate alone is a conjecture, a coincidence, or a stylized fact; a claim needs all three to be stated as a result.

Epistemic labels¶

Results carry their status explicitly, and the labels mean what they say:

• Measured: computed from sourced, per-edge documented data (registries, price lists, regulations, benchmark suites), with a confidence flag on every edge and, for load-bearing claims, an independent adversarial re-verification against primary sources.
• Pipeline: computed end-to-end on stylized intensities under an ordinal discipline; reported only as signs, ordinal strengths, and rankings robust across weight realizations, never as point estimates.
• MC-supported: a conjecture that survived Monte-Carlo search without a counterexample, awaiting proof.
• Programme: stated, motivated, and scheduled, with no result claimed.

Policy instruments get the same treatment: a provision is treated as an upper bound on the protection it names, its delivery as a question of enforcement, and its compliance costs as part of the ledger. No instrument is credited with an outcome without a dated receipt.

Reproducibility¶

Each project carries a verify/ directory whose scripts regenerate every in-text number and figure from the raw data, with fixed seeds and self-checks that abort on any mismatch. The two interactive explorers are generated from the same data and models: the graph explorer is built from the measured CSVs, and the model explorer embeds test vectors from the Python verification suite and re-runs them in your browser at every load, refusing to display if its JavaScript disagrees with the verified models beyond 10⁻⁹.

Status¶

Everything here is a working draft; none of it is peer-reviewed, and version numbers move. Comments are welcome: sf@fermigier.com.