How to use relative file paths in JSON-LD - schema

Question

I want to design a JSON-LD based schema for describing datasets. Similar to COCO data structure. However I can't use COCO for this applications, because reasons. I also can't always use URIs/IRIs for paths to assets, because data will be moving around internally (security/export controlled situation).
Imagine a folder like:
dataset
├── images
│   ├── 01.png
│   └── 02.png
└── index.json
Typically, #id would be a URI. Can I just use a relative path instead, like this example index.json?:
{
"#context": {
"ImageObject": "http://schema.org/ImageObject",
},
"#id": "no_idea_what_goes_here_usually_a_uri",
"#type": "Dataset",
"name": "My image dataset",
"image_set": {
"#type": "ImageDataset",
"name": "My Cool Dataset",
"files": [
{
"#id": "images/01.png",
"#type": "ImageObject",
"name": "cat picture"
...
},{
"#id": "images/02.png",
"#type": "ImageObject",
"name": "dog picture"
...
}
]
}
}
Or is there other chicanery I have to do with context/vocab to get this to work? (I'm still a bit of a json-ld greenhorn) I mean, I could always just deviate from JSON-LD, but I'd like to stick with the spec as closely as possible and ideally leverage schema.org data structures wherever possible.
Edit 1: I'm going to be appending things as I go along fact-finding. It seems to further constrain my question, I think what I want is a jsonld file which, with the addition of some machine-specific context, can be expanded into a valid and fully qualified jsonld document. That probably means I want to inject or modify #base, #context and/or #vocab. Still not entirely sure which. Here is an interesting related discussion.
Sir Tim himself seems fine with the practicality and inevitability of relative URIs. Additionally, the URI RFC does mention relative behavior as well. I mean, clearly, relative URIs are a thing, they are used all the time on the web. It's just not entirely clear the correct idioms to use in the context of JSON-LD's expansion engine.