Vibration test records go to large language models: after six deterministic gates, all 20 configurations reproduce the gold notching decisions bit for bit
Synopsis
In sine vibration acceptance testing for LEO constellation satellites, test conditions live in Chinese records and were previously coded by hand into the analysis. A Harbin Institute of Technology team had DeepSeek V4.1 Flash and Qwen-plus read the records into a 76-field configuration, and placed six deterministic gates between that configuration and the numerical stage, checking units, physical consistency, verbatim quotes, agreement between the two models, consistency with measurement exports and applicability of the trained model. Under the final pipeline and consensus policy, all 20 LLM configurations reproduced the gold notching decisions bit for bit, and the one real error plus all 16 injected errors in consumed fields were stopped, of which 16 would have run without gates.
Interpretation
Analysis conditions for satellite vibration tests can now be read from Chinese test records by large language models and handed to a fixed numerical program, with the model supplying only the configuration and taking no part in the decision. The earlier analysis code hard-coded these conditions by hand; an audit found 20 of them, of which 14 can be read from the record text. Under the final pipeline and consensus policy, all 20 LLM configurations produced notching results bit for bit identical to the gold run; without gates, 16 of the 20 were identical, 3 differed only by rounding and 1 ran with a wrong consumed value.
Before entering the numerical stage, each configuration must pass six deterministic gates field by field, and any failure rejects or flags that field. The six gates cover unit conversion, physical plausibility, verbatim quote checking, comparison of the two models' outputs, comparison with the measurement exports, and the applicability range of the trained model. Over the 20 final configurations and 4 pre-re-ask outputs, the consensus policy stopped no correct configuration and stopped the real consumed error and all 16 injected consumed errors; the lenient policy without the fifth gate let 2 injected errors run.
Asked to forecast response curves directly, the large language models do no better than linear extrapolation, whereas a small network whose output structure encodes a frequency shift and attenuation cuts the over-notched bandwidth at calibrated coverage. Linear extrapolation needs a large margin because its error grows with the local slope of the low-level curve, and an amplitude margin cannot represent that frequency-dependent error. Training on one wing and testing on the other, WarpNet-CQR over-notched 33.1 and 23.5 octaves at nominal coverage c = 0.90 against 235.8 and 214.7 for linear extrapolation, a reduction of 86% to 89%; both LLMs without examples did not differ significantly from linear extrapolation on the same 48 points.
Perspective
The conclusion rests on one complete acceptance test campaign: a single product with two wings, its test records, a gold configuration prepared with the help of a coding agent, and synthetic perturbations; the numerical part has 59 and 32 relevant test channels in the +Y→−Y and −Y→+Y cross-wing directions. The intended readers are test engineers who enter test conditions and set notching decisions, and developers of analysis pipelines.
Both API endpoints are aliases whose models may change, so results are reproducible only through cached responses; the field-specific quote rules were written after inspecting errors, so their scores are optimistic; repeat calls show that temperature 0 does not make outputs deterministic, so single-run error counts are samples; the cross-model check rests on nine known errors; and the trained model fixes the 0.5 g to 8 g level ratio through its pretraining simulation, so the sixth gate stops campaigns at other levels rather than adapting.
