Well...I'd say that https://en.wikipedia.org/wiki/Therac-25 has become much more notorious than the Arianne launch; but it's also a UX fault, not a straight-out software error.
This is one of those ideas that sounds a lot easier than it likely is.
You can do real time dose monitoring using diodes (see [1]). The diodes are quite small and don't generate huge amounts of scatter.
It is quite common to use TLD's[2] to monitor patient dose which is then analyzed offline to give dose estimates after the fact.
edit: that said, I don't think anybody was doing real-time in-vivo measurements during the era of the Therac incident.
[1] https://www.aapm.org/pubs/reports/RPT_87.pdf
[2] TLD=thermoluminescent dosimeter, little hunks of metal whose electrons get stuck in an excited state when exposed to radiation, you then heat them up and measure the light they give off as they relax back to ground state. The amount of light they give off can then be correlated with the dose the patient received.
Obviously, it's too late at that point to prevent the beam from hitting the patient, but you'll know that something went wrong and can lock the system until the problem is found.
[1] http://tommytoy.typepad.com/.a/6a0133f3a4072c970b0147e2ed7f8...
As an aside, TLDs and IVD systems are falling out of favor for patient dose monitoring because AAPM TG 62 (referenced in the link in your other post) is not directly applicable to IMRT and VMAT modalities, which are pretty common these days.
[1]http://www.sunnuclear.com/medphys/patientqa/epidose/epidose....
[2]http://www.sunnuclear.com/snc_site/solutions/patientqa/perfr...
p.s. If you are employed in Med Phys allow me to make a quick plug for my open source Med Phys QA database project: http://qatrackplus.com/
That's how I'd do it from my understanding of the usage practices today.
As others have noted, the big miss is the Therac-25 bug, which is pretty commonly taught in Software Engineering/Computer Science classes the example of how intangible software can kill people.