The Self-Confirming Archive: Partial Identification of Information Operations from Detection-Selected Disclosures

Most of what researchers know about how information operations look comes from platform takedown archives, and an archive is a sample of operations that were caught, attributed and disclosed. This paper works out what such an archive can certify about the operations it is drawn from. Known results transfer directly: without restrictions on detection, prevalence is unidentified; if detection probabilities differ by at most a factor κ, population odds lie within a factor κ of archive odds; if detection rises with asignature, the archive overstates it, so detectability measured on archive-labelled operations is an upper bound on detectability in the population. The new result concerns detectors retrained on the archives their predecessors produced: invisibility is absorbing exactly for learners that do not extrapolate beyond the archive’s support, and every later detector then faces a recall ceiling that no analysis of the current archive can bound. Within the model only extrapolation can recover a blind spot; outside it, exploration or an independent channel can. Disclosures record a candidate for such a channel, though not necessarily an independent one: when they state whether a network was found internally or through an outside lead, the two sub-archives identify the ratio of channel detection functions and certify selection from below,never its absence. In a pre-specified study of Meta’s public disclosures (255 networks disclosed from July 2018 to August 2026; 207 in the primary test, on disclosures to the end of 2024), the test does not certify channel selectivity (p = 0.65), a null the theory shows to be uninformative about representativeness.The one contrast near the threshold, stated use of AI-generated content (13% of internally and 29% of externally found networks), separates the cumulative sub-archives (pre-specified pooled lower bound above one) but not the channels within periods (Holm p = 0.060; stratified bound below one), so the data neither certify nor exclude selectivity on it. The external record has changed: in 2025–2026 the route of discovery is unstated for 45% of networks, and every stated external lead (four networks) came from an industry peer, none from journalists, researchers, civil society or governments.