Proofline on Arm

Inspiration

AI agents can generate plausible status updates, but high-stakes work needs evidence that can be reproduced. Proofline makes completion a deterministic decision: requirements are mapped to evidence, contradictions are surfaced, external actions stay behind approval, and every decision produces a stable proof-packet hash.

What it does

Gemini and Google ADK turn a task into explicit requirements. Proofline then evaluates authoritative evidence and returns one of four states: NEEDS_EVIDENCE, CONFLICT, APPROVAL_REQUIRED, or READY. The public FastAPI demo exposes the product flow; the public repository on main and committed Arm64 artifacts are the authoritative source for the optimization claim.

How we built it

The original gate compared every requirement with every evidence record. During the Arm Create submission period, we replaced that repeated scan with a single index of authoritative evidence by requirement. Matching work changes from O(R x E) to O(R + E), while preserving the final decision and canonical packet hash.

We built a comparison harness that preserves the original evaluator as the baseline, alternates measurement order, records every raw timing sample, and fails if baseline and optimized proof packets differ. A guarded evidence runner captures machine metadata, tests, benchmarks, an archive, and its SHA-256 checksum.

Arm64 results

The measurement ran on a real Google Cloud c4a-standard-1 Axion VM reporting aarch64, with one Arm vCPU and 4 GB memory. Using 256 requirements, 1,024 evidence records, and five alternating-order repetitions:

• Baseline median: 62.539 proof packets/second
• Optimized median: 98.414 proof packets/second
• Speedup: 1.5736x
• Throughput improvement: 57.36%
• Decision and deterministic packet hash: identical
• Same-host repository tests: 15 passed, 0 failed

The independent optimized core benchmark reached a median 23,665.171 packets/second across three repetitions of 10,000 iterations. We do not use an x86 or cross-machine comparison for the 57.36% claim.

Challenges we ran into

The hardest part was making a performance claim that a judge can challenge. We needed compatible Arm hardware, same-host before/after execution, alternating order to reduce bias, preserved raw samples, and a strict semantic-equivalence check. We also separated the existing Cloud Run product demo from the new Arm64 validation path so the submission does not imply that the public demo itself runs on Axion.

Reproduce it

Clone the public repository, install the project on Arm64 Linux, and run PROOFLINE_PYTHON="$PWD/.venv/bin/python" ./scripts/run_arm_evidence.sh. The evidence archive SHA-256 is 80d59d302564216b4280bd2a3c8728500bc6af082d4cba7d5e438cce8cf6ff84.

Responsible AI

The project and presentation were developed with AI assistance and reviewed by a human entrant. Benchmarks use synthetic fixtures and no customer or personal data. Gemini interprets requirements, while the final state and packet hash are computed deterministically.

Built With

• arm64
• benchmarking
• cloud-ai
• fastapi
• gemini
• google-adk
• google-axion
• google-cloud
• python