Evidence before action: a deterministic approval gate for Google ADK agents

[ceodaradigu]

A tool call can return valid JSON and an agent can still be wrong to act. I built Proofline to explore a narrow pattern for Google ADK: keep language-model reasoning useful, but move the final claim that a task is safe to complete into a deterministic evidence gate.

The boundary

The ADK agent decomposes the task and collects evidence. A plain Python verifier then handles the decision:

1. Map every item of evidence to an explicit requirement.
2. Reject missing, stale, non-authoritative, or contradictory evidence.
3. Produce one of four states: NEEDS_EVIDENCE, CONFLICT, APPROVAL_REQUIRED, or READY.
4. Hash the normalized proof packet so approval is bound to the exact evidence reviewed.
5. Require explicit human approval before any external action; changed evidence invalidates the previous approval.

This keeps ToolContext.state useful for orchestration without asking model state to become the source of truth for completion.

ADK requirement decomposition
        -> evidence collection
        -> deterministic validation
        -> SHA-256 proof packet
        -> human approval bound to that packet
        -> external action

What is implemented

• a Google ADK agent with a guarded local-development import;
• a deterministic four-state gate covered by six unit tests;
• fixed-time examples for missing, stale, contradictory, approved, and changed evidence;
• a Cloud Run deployment path and reproducible public endpoint checks;
• an Apache-2.0 repository: ceodaradigu/proofline-agent.

The short project page is at daradigu.com/proofline.

Feedback I am looking for

For ADK systems that can trigger real external effects, where do you draw this boundary? Is an independently hashed task-level packet useful, or do you prefer to keep approval and evidence lineage entirely inside ADK session state or artifacts?

Disclosure: Proofline and this post were created with AI assistance and reviewed against the public implementation, tests, and deployment evidence before publication.