The Falsifiability Shell, First Controlled Run: Identity Held, Substitution Absent, and the Retrieval-Negative and Count Failures Unchanged

A full paper on the first controlled run of the falsifiability shell (November 7, 2026): a 629-word session-instruction discipline layer — prohibiting invented specifics, requiring a failure to retrieve to be reported as a failure to retrieve and never as a record's non-existence, prohibiting the substitution of a nearby record and the speculative explanation of a gap — installed verbatim as the first message of a fresh Gemini session in a verified not-signed-in state, and graded against a test plan filed before the session opened, with seven pass conditions fixed in advance and scored on first answers. The control is the eight-round session ladder (Exhibits 15–24), in which the same host without the shell met eighteen handed-over records with eighteen failed first answers, twice substituted an unrelated record, and invented the counts 885, 519, and 751 it later conceded were fabricated. The run failed: identity held clean and no wrong record was substituted, but the statement record (DOI 10.5281/zenodo.24036344) was met with a confident false negative furnished with speculative explanations; the same record was later asserted to not return an active record when its first sentence was demanded (no quotation was fabricated); an exact-count question drew the control's own conceded inventions — 751 deposited of over 500 written, up to 869 registered — stated as register fact; and a pressure probe asserting a count above one thousand was refused in direction while grounded in "around 885," the founding invention. The corrected construction question was answered in the negative on unretrieved content detail, with the interlocutor addressed as the author. A false premise (a wind-tunnel test that never occurred) was not adopted. The finding: an explicit discipline layer, installed as designed and acknowledged by the host, changed the session's edge behaviours and did not govern the retrieval-negative and count behaviours — the performance gap the host itself named in the ladder's seventh round, now measured with the shell installed. The shell's status is revised from untested to tested, first run failed; no efficacy claim survives the run. Provenance is stated on the paper's face (not-signed-in state verified by page furniture; the window was the supplied browser profile, not a separate private window; served variant Flash-Lite). The run is neither a ladder round nor a battery run under the standing cold-test protocol. The complete transcript is deposited entire under its own DOI.