In my experience post-training mainly deals with "how" the model displays whatever data ("knowledge") it spits out. Having it learn new data (say the number of screws on the new supersecretengine_v4_final_FINAL (1).pdf) is often time hit and miss.
You'd get much better results with having some sort of RAG / MCP (tools) integration do the actual digging, and the model just synthesising / summarising the results.
I mean, consider The Boring Company sell a "flamethrower" despite being theoretically about… boring.
But they sold a blowtorch aka not a flamethrower. The difference being a flamethrower actually "throws flames" like 10+ feet.
I put quotemarks around "flamethrower" because that's what it was originally sold as before obvious and predictable legal issues with real flamethrowers and the fact it was obviously mimicing the prop in Spaceballs.
My point is: neither weed burners nor actual flamethrowers have anything to do with digging tunnels nor any adjacent aspect of civil engineering.
Just saying. It's kinda on brand by not being on brand because the whole network of companies... well I'm trying off topic.
It doesn’t matter how the person got access to dual-use info, like basically everything to do with large rockets, it’s 100% forbidden.