GATE-5G — From Prediction to Permission

What evidence should be satisfied before an AI output is permitted to cross from inference into an authorized operational state in increasingly autonomous network management?

Why This Problem Matters

Model evaluation and operational authorization answer different questions. A network classifier can produce a technically plausible prediction even when the evidence required to justify downstream operational use is incomplete, mis-scoped or uncertain. This gap becomes more important as network management moves toward closed-loop and agent-based automation.

Approach

The repository separates three stages of work: preserved Study 1 fault-classification evidence, a partially integrated independent Study 2 campaign, and the GATE-5G governance layer.

GATE-5G treats inference and operational authority as separate states. The governance layer evaluates evidence such as provenance, context, schema validity, model registration, calibration state, uncertainty or confidence evidence, distributional validity, applicable policy and requested authority before returning a governance outcome.

Architecture

Conceptually, GATE-5G sits between an AI/ML inference component and a network-management consumer, workflow or agent. It does not replace the classifier. Instead, it adds a deterministic decision boundary that can allow, cap, escalate, abstain from or block operational use according to available evidence and policy.

Evaluation

Preserved Study 1 evidence

The recovered Study 1 package contains:

• 33 complete simulation blocks;
• seven conditions per block;
• 231 accepted simulation CSV files;
• 462 accepted flow-level records;
• 12 quarantined CSV files and 24 quarantined flow records;
• grouping by combined seed and run ID to prevent flows from the same simulation block crossing model partitions.

Archived model outputs include:

The archived results also show severe class-specific failure for normal-operation records despite high aggregate accuracy, which motivates the governance focus.

Governance implementation

The implemented GATE-5G layer applies deterministic policy checks, explicit reason codes, authority controls, study/context isolation and tamper-evident audit evidence around model output use.

Results / Evidence

The preserved evidence supports analysis of the original Study 1 evaluation and its class-specific limitations. It does not establish a fully reconstructed end-to-end simulator reproduction. The governance implementation is maintained as a separate layer and does not retroactively alter the Study 1 experiment.

My Contribution

The current portfolio record identifies Yasir Siddiq as first author of the accepted WECE 2026 study with Sadaf Anwar and documents the GATE-5G reference implementation as part of the technical-governance work. The public policy memo is authored by Yasir Siddiq and Sadaf Anwar.

Limitations

The repository does not yet verify the exact original ns-3 version, exact original 5G-LENA version, canonical original simulation source, complete frozen Python environment or an independently rerun end-to-end reproduction for Study 1. Study 2 remains incomplete. GATE-5G does not prove that an AI system is safe, compliant or accurate.

Reproducibility

The public repository preserves accepted and quarantined raw CSV evidence, exclusion records, hashes, an evidence inventory, archived analysis material and GATE-5G implementation documentation. This should be described as recovered experimental evidence plus a separate governance prototype rather than complete simulator reproduction.

Current Status

Study 1 evidence is preserved and documented, Study 2 remains incomplete, and the GATE-5G governance layer is implemented and verified within the repository’s stated scope.