A source-stratified opacity profile of six public negative-results record families

Public records of negative outcomes in AI research and AI governance exist in six systems, each with its own documentation rules. The Nekropolis v0.2 corpus assembles 17,115 records from these systems into one schema. This article uses that corpus to profile how much each system actually explains when it removes, withdraws, revokes, or archives something. Five per-family measurements — cause-presence, explanation-length distribution, category coverage, inferred-cause coverage gap, and stated-reason uniqueness — produce a source-stratified opacity profile. Three regimes separate the six families. Retraction Watch and arXiv withdrawals are reason-rich: the stated-reason field carries variable per-record text with high string diversity. EU LOTL/EUTL trusted-service records and GitHub archived repositories are templated-status: the field is fully populated, but with a single pipeline-generated sentence that repeats across nearly all records. OpenAlex retracted works are empty in this snapshot of a fallback discovery query. Cause-presence alone collapses all three into two regimes and misreads a templated field as an explained one. The profile is stable between the v0.1 frozen base (14,416 records) and the v0.2 working corpus. All measurements reproduce from seven open scripts against a SHA-256-verified input. Target venue: Quantitative Science Studies (MIT Press for ISSI). Working-paper preprint; under preparation for submission. Dataset of record: Nekropolis v0.2 corpus, SHA-256 08231b5058f91533003296f51b76b7e38285f21e4efc391ad579ddf7c3ad1ae2. Analysis pipeline: sapsan14/tyche-research-vault, branch nekropolis-opacity-analysis-v0.1.