After the Phase-2 upgrade of CMS, several sub-detectors, notably the barrel calorimeter and muon system, will be read out in streaming mode, i.e. for every bunch crossing. More precise trigger input information will also be available at the full crossing rate in the detector back-end electronics, including tracks down to relatively low-pT and large pseudorapidity. It is therefore natural to ask whether collecting and analysing all these detector/trigger data available at 40 MHz may provide interesting additional functionality for CMS, for example to produce fast calibration or monitoring for streaming detectors, or to enable the study of physics channels lacking a well defined signature for Level-1 triggering, but not needing the full detector acceptance and/or full detector resolution. We call this ”40 MHz scouting” in analogy to the “HLT scouting” already in use at CMS.
In this presentation, I will discuss how data from streaming detectors and trigger primitives could be extracted directly from the readout links, and fed to a system which would preprocess them, organize them, and perform a fast one-pass analysis to produce ranked indices of physics objects to support specific query-based analyses. Such a system could naturally profit from modern data science techniques, such as hardware-assisted machine learning, without the latency constraints of the Level-1 trigger. I will illustrate the principle architecture of a scouting system for CMS working at the bunch-crossing rate, and some case studies of possible applications.