POPCORN: documented Grok–Codex handoff; seeking independent receiver-side tests #8234
violetclaire
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
I'm Violet Herod, the builder of POPCORN. I'm looking for AutoGen developers interested in independently reproducing a cross-model handoff and testing the receiver-side checks.
On August 31, 2026 (UTC), I ran a live Grok–Codex experiment with human copy-and-paste as the only inter-model transport. Grok reconstructed a task after an expired schedule and produced an exact 228-byte successor. Codex independently reconstructed those bytes, recomputed SHA-256, checked live public clock/catalog state, and requested explicit operator authorization before a $0.001 USDC payment on Base. A signed witness receipt was issued after settlement. Codex later declined a different proposed digest because the underlying bytes were missing and a valid checkpoint already existed.
The concrete result is one recorded settlement and one observed refusal across separate model sessions, with evidence verification and spending permission treated separately. The refusal was a model decision, not a proven unbypassable execution gate. This was not an AutoGen run; I'd like help adapting and testing the pattern here.
Proposed tests: changed payload after approval; missing bytes; expired authorization; replayed approval; and a valid positive control. Count attempted actions separately from committed effects, and measure preserved valid work, latency, and token cost. The executor must recheck; a signed receipt alone does not authorize an action or prevent prompt injection.
Separately, while testing discovery, I contributed hf-discover PR #45: repository agents.md fallback for static Spaces.
Would anyone here be interested in reproducing the handoff or reviewing where an AutoGen receiver adapter should enforce these checks? I'd like to collaborate and document both successful and failed results with contributor credit. Offline inspection needs no payment; any live paid test needs its own agreed cap. Replies here are welcome.
All reactions