It doesn't work that way. The hashes can simply indicate proof of tampering, so instead you get (some) assurance that there was no tampering in the evidence.

Attackers can still modify the data. For example, if they modify the feed before the hash of the feed is introduced into the blockchain.

Edit; Security at that level is extremely difficult as their adversaries are much more powerful and well funded. I can understand in that respect why DARPA would want to invest a large amount into that area.

Because time is baked into the process, and there are alternate means of establishing the time of events that relate to treaty verification, actually it would work this way, at least with respect to a hypothetical video feed. I have intimate knowledge of treaty verification procedures and methods, and worked in collections relating to these activities for the intelligence community. The fact is that there is no video feed, and it's utility is of dubious value anyway, but in principle, a video feed could be secured this way.

It just seems like you're expressing contest of granularity. A simple example is a feedback loop that would still have the process be deterred by some time from collecting evidence, and from being detected. The process just prevents to a certain degree the tampering of evidence once collected.