Engineering
Sortie
Independent model reviews, with a recorded decision about what matters.
Sortie sends a code change to several reviewers, brings their findings together and applies a severity policy before a merge. The retained findings make agreement and disagreement inspectable. Model agreement is evidence to examine, not proof that a change is correct.
Experimental review tooling. Implementation versions need reconciling before publishing a definitive stack.
Technical detail
Based on the published Python implementation account. Other material names a TypeScript/Bun version; this describes the documented design without asserting version equivalence.
Parallel review and synthesis
A configurable roster reviews the same change independently. A further model synthesises findings into a structured debrief. Code, tests and documentation have separate review modes so the question and prompt can match the artifact being inspected.
The decision boundary
In the documented configuration, findings independently identified by at least two reviewers are convergent. Convergent critical or major findings can block a merge; single-reviewer findings remain recorded as advisory. That policy makes agreement actionable, but can also leave a serious single-reviewer finding non-blocking. It is a tradeoff to inspect, not a correctness guarantee.
Binding review to the artifact
Attestations are keyed to the Git tree rather than only a commit label. A change to the staged content invalidates the match with the reviewed tree. The structured ledger retains findings, dispositions and cost data so the review process can itself be evaluated.
Stack / components
Parallel reviewers, synthesis, severity policy, retained findings