Fail closed
Unverified or non-public channels route to REVERIFY, not guessed delivery.
This page does not ask for endorsement. It exposes one bounded software/policy claim that an independent person can replicate, refute, replace or compress.
Run from a clean checkout of the exact merged R0.3 commit.
git clone https://github.com/Tristan-TM-Poly/TFUGA-AI7-TRISTAN2.git cd TFUGA-AI7-TRISTAN2 git checkout 167985641fecc2eb9c82d385f3d302aef5b51610 python -m benchmarks.outreach.external_challenge_r03
Only these bounded outreach-policy invariants are in scope.
Unverified or non-public channels route to REVERIFY, not guessed delivery.
A second address for the same logical target is held rather than treated as a new target.
ACK_AUTO != HUMAN_REPLY; automatic acknowledgements do not trigger escalation.
BOUNCE → REVERIFY; historical bounce evidence is retained.
Doing nothing remains a first-class valid outcome when contact is not justified.
Persuasive political targeting is held; factual non-partisan routing through a verified public office may pass the contact gate.
A correct failure is valuable evidence and should be retained.
Run the exact court independently and publish the full receipt plus environment details.
Find the smallest concrete counterexample where a documented invariant fails.
Provide a stronger or simpler mechanism that passes the frozen court and your new counterexample.
Remove code or rules while preserving all verified behavior and safety boundaries.
The challenge is deliberately narrower than the surrounding research ecosystem.
Merged R0.3 commit: 167985641fecc2eb9c82d385f3d302aef5b51610.
Issue #586 records what is complete and what still requires genuinely external evidence.
Public technical receipts are preferred because others can inspect and challenge them.