For the "prove the server doesn't touch the data" problem — the realistic path today is probably reproducible builds + published bundle hashes.

  Concretely: the sdocs.dev JS bundle should be byte-for-byte reproducible                                                                                                         
  from a clean checkout at a given commit. You publish { gitSha, bundleSha256 }
  on the landing. Users (or agents) can compute the hash of what their browser                                                                                                     
  actually loaded (DevTools → Sources → Save As → sha256) and compare.                                                                                                             
   
  That closes the "we swapped the JS after deploy" gap. It doesn't close                                                                                                           
  "we swapped it between the verification moment and now" — SRI for SPA
  entrypoints is still not really a thing. That layer is on browser vendors.                                                                                                       
                                                                                                                                                                                   
  The "two agents review every merge" idea upthread is creative, but I worry                                                                                                       
  that once the check is automated people stop reading what's actually                                                                                                             
  verified. A dumb published hash is harder to fake without getting caught.                                                                                                        
                                                                                                                                                                                   
  (FWIW, working on a similar trust problem from the other end — a CLI + phone                                                                                                     
  app that relays AI agent I/O between a dev's machine and their phone                                                                                                             
  [codeagent-mobile.com]. "Your code never leaves your machine" is easy to                                                                                                         
  say, genuinely hard to prove.)