Compare every change with an unmodified node before keeping it

[hotpixelgroup]

What this changes

Turns Bitcoin Vibes from "an AI rewrites your node" into a lab that shows what a change does: every proposal is compiled in a sandbox and run beside an unmodified node on a private chain before you are asked to keep it.

• Two-node lab. A proposal is labelled by what was observed: no difference, local policy only, tightened, loosened, or split, with the block that was refused and the node's reason.
• Containment. AI-written code is compiled, tested and run inside sandbox-exec (macOS) or bubblewrap (Linux). The agent gets file tools only; the unrestricted mode is gone. A shared request is treated as untrusted input.
• The real network is a deliberate step. Typed confirmation, no starting a build whose comparison found a consensus difference, wallet and incoming connections off by default, and a plain-words alert when the node has refused a block.
• Checks the agent cannot grade. Compile before review; Bitcoin Core's unit tests per suite, with failures named as findings; edits to tests reported; applying is a fast-forward of the revision that was built.
• Ten ready-made experiments with a prediction first, a written explanation with sources, and a follow-up question. None needs an AI; seven need no compiler, and releases now ship the changed nodes for the other three. Ten true-or-false statements are asked before and after.
• Install and update. Shallow-clone updates work and report truthfully; the installer can be pinned to a release and checks downloads against their SHA-256; the Mac launcher no longer updates by itself; Docker runs unprivileged and keeps saved experiments.
• CI and releases. CI builds and runs the unit tests, checks src/ against the upstream base, and confirms each ready-made change is refused in the documented direction. Releases are gated on tests, published all-or-nothing, and carry checksums and provenance attestations.
• Docs rewritten to match, including what the checks do not prove.

Why

A research review of the project found that a lone modified node teaches the opposite of what is true (a loosened rule "works"), that every protection was advisory, and that the checks behind the "verified" label could be written by the agent itself. The full list is in the commit messages.

How it was tested

• 90 Python tests pass under Python 3.14 and the system Python 3.9, including the ones that drive real nodes inside the macOS sandbox.
• All ten ready-made experiments were run through the real console; the three source changes were compiled, tested and compared, then rerun from the kept nodes.
• A consensus change was applied, started on regtest, the reference node rebuilt from a temporary checkout, and the baseline restored.
• Browser and accessibility checks of the landing page and demo pass in Chrome.

Not tested

• A live AI session. The agent path was driven with a stand-in program only.
• Linux. The bubblewrap arguments have never run on a Linux machine; the CI job in this PR is their first real test.
• Docker, the Mac app, the release workflow, signing and notarisation. None can run locally.
• Shared networks. Everything ran on regtest; the mainnet rules are covered by unit tests only.
• Whether it teaches. No learner has used it, and the lessons have not been reviewed by a Bitcoin Core contributor or an educator.

contrib/vibes/QA.md has the detail.

Checklist

• I ran it
• python3 -m unittest discover -s contrib/vibes/tests passes
• UI touched: checked at 320, 390, 768 and 1440 pixels, and by an automated accessibility audit
• User-facing claims match what the code does
• src/ is not touched: a fresh install still ships stock Bitcoin Core consensus

Review

• Someone other than the author has read the diff. An independent AI review is being attached as a comment; it is advice, not approval
• CI is green on the exact commit being merged

🤖 Generated with Claude Code