Loops HouseTITAN EXCUBAT

NEAR HACKATHON
Links
Team
1 member- ALOwner
Alnitak34
Overview
AI agent developers need a clear verification boundary before model output reaches the agent loop. TITAN EXCUBAT is an experimental NEAR AI provider integration for DeepSeek Harness. It buffers completions and releases no text, reasoning or tool-call chunks until configured SDK checks pass and a provider_tee verdict matches the response ID.
NEAR AI supplies attestation, response-signature verification, TLS binding and supported-field encryption. EXCUBAT adds the release boundary, Harness translation, cancellation and size limits, sanitized diagnostics, and an exact reviewed proxy source-revision policy. The original public integration proposal is credited in NOTICE.md.
A public isolated run passed 52 unit and 48 integration tests. Actual Harness/SDK code runs with controlled transport and simulated attestation. New source-policy tests simulate Sigstore/TUF cryptography. The video includes actual gate execution with a simulated SDK; it does not execute tools. Separate integration tests exercise a harmless tool loop.
MIT source, pinned dependencies, test logs and reproduction steps are public. A historical live result is labeled separately. Signatures do not authorize tools or prove vulnerability-free software. Current post-incident NEAR deployment verification remains unestablished; no production security claim is made.