Appaloft Launches Open-Source Delivery Evidence Chain for AI-Generated Code
Key Takeaways
- ▸Appaloft's platform addresses a critical infrastructure gap in AI-driven development by creating auditable records of what code was approved and deployed
- ▸The content-addressed artifact design (manifest-of-hashes vs. tarballs) enables deduplication across agent runs and detects mutations mid-freeze
- ▸Fail-closed design is central to the architecture: broken evidence chains are explicitly detected and reported, preventing silent deployment failures
Summary
Appaloft, an open-source (Apache-2.0) delivery platform, has published technical details on its 'delivery evidence chain'—a system designed to solve a critical problem in AI-assisted development: proving exactly what code shipped and who approved it. The challenge is acute: AI coding agents can produce dozens of deployable changes daily, but traditional deployment practices lack a verifiable record linking the exact artifact approved to what's actually running in production. Appaloft's evidence chain addresses this through a four-step process: freeze (creating a content-addressed artifact), preview (showing what will deploy), promote (requiring explicit approval), and verify (confirming deployed reality matches approved state). Rather than hashing tarballs, the system uses a manifest-of-hashes approach—a canonical JSON list of file paths, sizes, and SHA-256 hashes—enabling deduplication, mutation detection during capture, and path safety validation. By design, the evidence chain fails closed: if any link breaks, the system explicitly reports why rather than silently passing incomplete verification.
- The four-link chain (freeze → preview → promote → verify) creates verifiable custody records linking exact bytes approved by a principal to the workload serving traffic
Editorial Opinion
Appaloft identifies and solves a real problem that traditional DevOps tooling doesn't address: the accountability gap created by prolific AI agents. When a human writes code, deployment memory is mostly sufficient; when an agent produces twenty candidates in an afternoon, proof matters. The fail-closed design is particularly sound—refusing to silently pass incomplete evidence chains prevents the deployment ambiguity that creates 2am incidents. This technical foundation could become a valuable reference pattern as the industry scales AI-assisted development.
