The Scenario
The ski-chalet thought experiment asks a continuity question:
If cloud access disappears, what kind of local setup is enough to keep useful domain work going?
The scenario is intentionally constrained — a niche, slightly absurd setup that is nevertheless plausible for a specific class of user. You are at a ski chalet. You have your machine. The internet is unreliable or gone. You need to continue Project Phoenix domain work.
The original framing implied the harshest possible version: one local machine, one local model, no cloud, no supporting structure. That version is interesting, but it mostly yields a useful "no." The stronger and more honest version is narrower:
The Setup
A 3090-class machine with Ollama and a local model already installed. A hobbyist who already has this setup is the specific target of the claim.
The Bundle
A portable USB kit carrying the domain repo, venv, deterministic datasets, databases, question banks, grounding artifacts, and scripts.
The Result
Prepared offline continuation — not spontaneous local creation from nothing. The harness makes the difference.
The Realism Standard
The scenario does not need to be ordinary consumer behavior. The point is not average life. The point is plausible offline continuity under prepared local conditions.
Consider: a hobbyist with a 3090 is much more likely than average to already have Ollama and Gemma installed. A USB travel kit carrying a repo snapshot, venv, and domain harness is strange — but fully believable for the same person.
That is the right realism standard. Not "could anyone do this?" but "could the specific kind of person who would attempt this scenario make it work?"
The current evidence says: yes, with a prepared harness.
The Key Insight
Raw local behavior is weak. The recent local-mode evidence is unambiguous about this. A bare local model without harness support gives the ski-chalet question a firm answer: no.
But that is not the end of the question. It is the lower boundary.
Once you add a prepared harness — deterministic substrate, grounding layer, control layer, audit layer — the picture changes materially. Grounded Gemma performs much better than raw Gemma on fixed domain question sets: stronger specificity, stronger alignment, lower omission.
The variable that matters most is not model size. It is harness level.
The important claim is not: 3090 + model = success
It is: 3090 + prepared harness = useful offline continuity
Navigate
The Harness
The four harness layers, the USB bundle contents, and the three configuration tiers that define how much continuity is achievable.
→Results
Four results from the ski-chalet investigation — from the lower boundary ("raw local is not enough") to the positive claim ("harnessed local is materially better").