approach

Autonomy is only useful if you can trust it.

The industry raced to unbounded agents. We took the opposite bet: constrain the machine so tightly that its output survives code review. Three ideas make that possible.

A fact model, not a chat log

We extract a deterministic Provenance Knowledge Graph — 7 node kinds, 9 edge kinds — from code and docs. Every claim carries a location, so answers are checkable, not persuasive.

Determinism where it counts

Comprehension has no LLM in the loop. The same commit produces a byte-identical report — something you can put in CI and diff like any other artifact.

Gates a human owns

Generation is autonomous; consequences are not. Nothing is pushed, merged, or written to your tracker without explicit approval — and every action is recorded in an append-only log.

See these principles in a real, shipping product — Spine is MIT licensed and its comprehension layer is free and read-only.