A four-layer, commit-time gate that stops AI hallucinations before they enter your repo. Symbolic, syntactic, behavioral, and visual verification — all inside your existing git workflow.
Hallucinations don't all look alike. The IDE Package treats them as four distinct categories of error, each with its own gate.
Reads your stated intent before the commit. If the staged code drifts from the goal, Sovereign raises a veto. The Truth Engine does not allow goal-orthogonal commits to pass silently.
AST-level analysis of every staged file. Flags references to functions, modules, or APIs that don't exist in your project or its declared dependencies. The most common AI failure mode, caught at the symbol level.
Parses every changed file and runs the relevant test surface. If the staged code breaks parsing, types, or tests, Sentinel blocks. The agent says "tests pass" — Sentinel verifies.
For UI changes, Eyes loads the rendered output, fetches the asset, and verifies that the visual claim matches reality. The agent does not declare completion. Reality declares it.
No new tools to learn. No new workflow. The Truth Engine runs at git commit time and stays out of the way until something is wrong.
No marketing math. The IDE Package is governed by these four numbers and they don't change.
$0.01 AUD per commit, charged at the moment the first layer activates.
100 calls on signup. No expiration. No card required to start.
Hard gate. Zero balance blocks commits. The Engine fails closed, never open.
One sync auth check per commit. One async usage log. ~80ms total overhead.
100 free calls on signup. No subscription. No expiration.