Proofline

Inspiration

Autonomous agents are good at producing answers, but real work fails at the acceptance boundary: a deployment is assumed to be live, a marketplace task is treated as current, or a deliverable is declared finished without authoritative evidence. Proofline makes that boundary executable.

What it does

Proofline turns a task contract into explicit requirements, evaluates evidence for freshness and authority, detects contradictions, and produces a deterministic SHA-256 proof packet. Missing, stale, non-authoritative, or conflicting evidence blocks completion. Even when every substantive requirement passes, external actions remain behind an explicit human-approval gate.

How we built it

• Gemini 3.6 Flash interprets requirements and evidence through a real Google ADK agent.
• Google Agent Development Kit provides the application and narrowly scoped tool surface.
• A deterministic Python core evaluates evidence and hashes canonical proof packets.
• A reproducible Docker image runs the ADK API on Google Cloud Run.
• Firestore and Pub/Sub are documented extension points for future durable storage and asynchronous rechecks; they are not represented as active in this demo.

The verification core runs without cloud credentials so judges can reproduce the safety boundary locally.

Live evidence

Proofline is deployed in europe-west1 as Cloud Run revision proofline-00001-rqw. An external verification returned HTTP 200 for app discovery and session creation. A real Gemini 3.6 Flash run invoked evaluate_packet and returned READY with one authoritative evidence item and packet hash 973750f90ceffd925eba6716399f9064fcc789522ae8455efe764ef6c841eb5d. The exact URL, configuration, and result are recorded in the public repository.

Challenges

The main challenge was separating probabilistic interpretation from deterministic acceptance. Gemini can help understand the task, but it cannot silently decide that evidence is sufficient. Proofline therefore narrows the model's role and makes freshness, source authority, contradictions, proof hashing, and approval requirements explicit and testable.

Accomplishments

• Nine focused tests cover the decision surface, evidence quality, stable hashing, human approval, and the Cloud Run container contract.
• The real Google ADK 2.6.2 package imports the project as concrete App and LlmAgent instances.
• The public Cloud Run deployment was verified end to end with Gemini and the deterministic tool.
• The public repository contains reproducible commands, a four-state decision matrix, architecture documentation, and secret-safe deployment evidence.

What we learned

Trustworthy agents need more than good prompts. They need a narrow acceptance boundary, explicit evidence provenance, stable proof artifacts, and a clear point where human authority resumes.

What's next

Persist packets in Firestore, use Pub/Sub for scheduled evidence refresh, and add adapters for CI results, deployment health, marketplace state, and signed delivery receipts.

AI disclosure

Proofline and this demo were developed with AI assistance. Synthetic fixtures are clearly identified. Hosted deployment claims are backed by the live Cloud Run service, public source code, deterministic tests, and recorded verification evidence.

Built With

• cloud
• google