How
How it works.
Multi-pass deterministic analysis of the codebase, a model trained on what it produces, and a closed loop that returns evidence to your CI/CD gate.
The engine
Deterministic analysis, and the model it trains.
Everyone trains on code.
We trained on proof.
Before any model is involved, the engine runs a multi-pass deterministic analysis over the codebase. None of them guess.
- Structural dependencies — extracted from the code itself, the same view a compiler has, rather than a heuristic approximation of it.
- Arc-level coverage — finer than the line coverage most tools report, so a branch that was never taken cannot hide inside a line that was.
- Semantic diff — what a change actually did, so a reformat is not mistaken for a change in behaviour.
- Formal impact analysis — which behaviour the change can reach.
- Coverage-gap analysis — which of that impacted behaviour nothing exercises.
That is an answer computed, not inferred. It is also what keeps everything after it affordable, because the model is never asked to reason about a whole repository — only about the region these passes have already narrowed to.
The model is then trained on all of that, plus something nobody else collects: mutation evidence — a record of which tests actually caught a deliberately injected fault. That is a label for observability rather than coverage, and it cannot be scraped from public code. It only exists if you have already run the loop.
It is a foundation model, pretrained on open-source code, so a new codebase starts warm instead of cold. And it picks out the code that actually matters for a given test, which is what holds token cost flat as a repository grows rather than letting it climb with size.
The loop, end to end
Your change enters the loop alongside the existing codebase, its specs and its tests. SCOPE performs impact analysis, selects the regression set that change can actually break, and reports the coverage and observability gaps it leaves behind. Those gaps become a gap map: unverified behaviour, missing tests, affected paths.
GNOME takes that map and closes it — generating tests and checkers, validating that each one reaches its target, then mutating the code to confirm each test can detect a real fault. What survives becomes verified evidence, handed to the CI/CD gate. The loop closes: that evidence feeds the next change’s analysis.
