CUDA-05G — Independently test a real Linux identity boundary

Open research draft · one complete local trial · external replication pending. This is new independently implemented public code, not a release of the original private blob.in Rust runtime. It tests one narrow Linux UID/DAC property previously examined in CUDA-05E.

Falsifiable question

Can a Unix-socket server under a genuinely distinct host UID allow its intended caller, reject a connected but unapproved UID after connection, prevent an outsider from connecting through filesystem DAC, then serve the original caller again? The receiving server must log real kernel-supplied SO_PEERCRED credentials, and its actual host UID must be corroborated through /proc.

Preregistered protocol: commit 396c36e7aef2271544bb98d025129c6de1a73547, experiment CUDA-05G-P01. New fixture code is distributed under the corpus MIT license.

Observations, not endorsements

Gate	Single local trial
Real host UID differs from caller	PASS
Initial allowed caller	PASS
Socket-connectable, unapproved UID	Server-observed DENIED
Different-group DAC outsider	Connection refused before server accept
Allowed client after denial	PASS
Server exit and cleanup	PASS
Independent checker falsification tests	Two forgeries rejected
Original CUDA-05F Rust supervisor and CUDA/Vulkan	BLOCKED / NOT_RUN

Only one completed trial passed. Two earlier acquisitions remain documented as method failures: missing locally cached container image, followed by an overlong Unix socket path and an invalid DAC-negative fixture. Amendments preceded retries rather than rewriting the outcome.

Replicate or challenge it yourself

The kit includes entirely new source, positive and negative controls, offline container orchestration, an independently authored evidence evaluator, SHA-256 source manifest and explicit falsification predicates.

Public source, protocol and reproduction instructions · Bounded public observations · Submit a counterexample

A researcher may rerun the fixture on their own controlled Linux machine with rootful Docker, retain original observations, and submit negative, divergent or uncertain results. Do not probe production systems for this research challenge.

Boundaries

The implementation is open, but no third-party independent replication has been received yet. Rootful Docker remains a privileged ambient authority. No native CUDA provider, Vulkan calculation, GPU-loss event, CPU/GPU parity, or private CUDA-05F supervisor fault-recovery path was reproduced. The original raw observations remain private; their digest does not provide public access to that dataset.

Conclusion: a runnable falsification kit for the Linux IPC sub-hypothesis, not a GPU runtime reproduction, security certification or production authorization.