Sealing Every Step: A Pre-Registered Search for GOES-16 Lightning Precursors of Tornado Touchdown, and a Positive Control the Instrument Did Not Pass

We report, in full, a line of research that asked whether the spatial arrangement of lightning observed from

geostationary orbit (GOES-16 Geostationary Lightning Mapper, GLM) changes at the moment a violent tornado

touches down, and then asked whether our instrument could recover even the best-known lightning signal of severe

storms, the lightning jump. Every test was sealed (design, case list and code fixed by SHA-256 before any data were

read), the AI partner wrote a sealed prediction before each run, each test was run once, and the verdict was taken

from a pre-stated rule (R reproduced / F not reproduced / U unresolved). The flash-arrangement hypothesis H_spread

(the mean ternary shell index of flashes rises after touchdown) was never established: U in three tracked tests (the

closest, OOS-4, combined p = 0.0106 against a 0.01 threshold), and F with a tracking-free moving-disk reader on 40

new EF3 tornadoes (n = 26, combined p = 0.106, mean effect +0.059, 95% CI [-0.018, +0.137]). The effect estimate

changed sign with the reader radius. Four sealed positive controls then tested the instrument: lightning jumps did not

stand out before touchdown (p = 0.924) or before the first severe report (p = 0.807), did not separate severe from

matched non-severe storms (McNemar p = 0.773), and a faithful cell-based reproduction of the Schultz 2-sigma

algorithm, with its activation threshold re-derived for GLM by the original definition (35.2 flashes/min), reached

literature-level scores (POD 0.31, FAR 0.63, CSI 0.26) that a timing-shuffled null matched (CSI 0.283; p = 0.906). We

therefore do not claim that no precursor exists; we claim that with GLM total lightning, these readers and these

anchors, none is visible, and that the instrument has not been shown sensitive enough for absence to mean much.

The record of how each negative was protected from optimistic re-reading is the main contribution.

※ Update in version 1.1.0. We then measured the instrument by sealed injection-recovery. The jump instrument recovers a pre-touchdown surge common to all storms once the flash rate doubles (threshold A* = 1, about +23 flashes/min) but not a +50% surge, and by construction it cannot flag a single storm. The H_spread reader does not respond to injected rate surges and recovers a post-touchdown spread of 10% (A* = 0.1). Sized by these numbers, a sealed confirmation on 185 new EF2+ tornadoes (113 valid) gave F: mean effect -0.0035 (95% CI [-0.039, +0.032]), equivalent spread -0.7% (-7.8 to +6.4%). The +12% suggested by the moving-disk test did not replicate. Section 9 reports these steps; Sections 1-8 are unchanged apart from pointers.

※ Update in version 1.2.0. Because the jump instrument could not judge one storm, we built one that can. Placebo ranks within a storm cannot reach p = 0.05 (a 90-min history holds about three independent 30-min windows), and nulls generated from a storm's own history were anti-conservative on real GLM data (false-alarm rates 0.09-0.31 at a nominal 0.05). A threshold calibrated on pseudo-touchdown windows of other storms worked: a fixed threshold on the largest 8-min relative rise in 30 min flagged a single storm with probability 0.53 for a tripled rate and 1.00 for a five-fold rate in injection tests. Sealed and run once on 185 new EF2+ storms, it kept its false-alarm rate (2.6%, 95% CI 0.3-5.6%; verdict R, the first R of this line), and 7 of 138 storms crossed it in the 30 min before touchdown (p = 0.54; F). Before the first severe report the result was 5 of 88 storms (5.7%; binomial p = 0.45 against 0.05), verdict F. A common +50% surge would need about 1800 storms, because the pre-touchdown window is quieter than earlier windows (Section 10).

Data and code. Scripts (AGPL-3.0-or-later), all seals with SHA-256 files, case sheets, prediction texts and result

files are provided as Bridge_Paper_Lightning_scripts_v1.2.0.zip. GLM data: NOAA GOES-16 on AWS (noaa-goes16).

AI use. Designs, code, analyses and text were produced by Kurado (Claude Opus 5.5, Anthropic) working with M. Watabe, who

chose every direction and supplied all data.

※ Version 1.1.0 adds sealed injection-recovery measurements of both readers and a sealed confirmation on 185 EF2+ tornadoes (verdict F).

※ Version 1.2.0 adds a storm-level lightning detector with a sealed, transferable false-alarm rate (verdict R), its sealed tests before touchdown and before the first severe report (both F), and a design for a common +50% surge (Section 10).