The researcher-blind publication-boundary measurement algorithm

The argument

A public repository's current state X passes through normalization, a detector ensemble, sensitive-value suppression, feature extraction and adjudication to a research-safe output: X → N → D → S → F → A → O. Seven detectors cover paths, secret-like material, networks, authentication, semantics, metadata and recoverability, and each declares its scope.

Everything that touches the data sits inside the threat model — detectors, sanitizers, logs, feature generators, debug output and the disclosure module. The algorithm forbids repository identity in the analysis layer; raw paths and endpoints in the dataset; credential validation; quotable source text; raw exception logs; group labels before classification is locked; and turning UNKNOWN, or a single detector's PASS, into a privacy verdict.

It is a bounded, auditable, machine-first measurement architecture — not a perfect anonymizer and not a formal proof of semantic privacy. Its implementation is the RBPM runtime whose versions and results are on the evidence page.