Evidence-Backed AI Operations Console

Status: Prototype. New screens on the data structures of a private evidence system I built. Data shown is illustrative.

Anyone running AI agents across repositories and chat histories ends up asking what is true right now and why they believe it. Decisions drift without anyone noticing. An agent says it finished, and that report gets taken as proof. Two notes contradict each other, and the newer one quietly wins.

Ultravision is a private, local-first app I built for one owner. It watches GitHub repositories, a chat export and AI agent runtimes, and stores every raw record as immutable evidence. Database triggers reject edits and deletes on raw records. A state snapshot will not seal while any item lacks provenance. Five deterministic rules sort what needs attention. When two assertions conflict, both stay visible and nothing picks a winner. Agent dispatch waits for my approval, every step goes to an append-only audit, and a provider's report that it finished is stored as evidence and never counts as proof.

These four screens are new. Field names, enums and relations come from the repo's schema. fernhill-labs (demo), its people and every value are illustrative.

Self-Hosted AI & Open-Source Deployment, from $600: one agent on your own machine or server, connected to one channel, locked down, with logs and a runbook. A private AI stack with local models, a chat interface and document search starts at $1,250.

Trail: a decision traced from the brief back to the records behind it.

Attention: five rules sort what needs me; conflicts stay listed with no winner.

Changes: what changed since my last review, as typed changes between sealed snapshots.

Control: every dispatch waits for approval, and a provider's report counts only as a claim.