The problem
The strength of a review is its independence.
A review is worth trusting in proportion to how independent it is. When the model family that wrote a change is also the one grading it, the two share the same blind spots, and the review inherits them. Real assurance comes from judges that reason differently and reach the change from different directions.
Heterogeneity is the whole point. Findings that only one critic surfaces are often the ones worth having — and independent judges reach them precisely because they do not confer.
How it works
A panel of rival models, one deterministic verdict.
The orchestrator runs independent critics from rival model families in parallel against the same change. Each produces a structured verdict with findings tagged by severity. A pure verdict kernel composes those findings under a min-complete-quorum policy: a blocking finding holds the merge when a second independent critic reports it at the same place, higher-severity findings corroborate, and lighter findings are surfaced as guidance.
Because the aggregation is a deterministic function over recorded findings, the same panel and the same change resolve to the same verdict every time. If a critic times out, the panel proceeds with the critics that completed, and the verdict states exactly what was evaluated — so you always know what the change was checked against.
- Critics from different model families run in parallel against the same change
- Each returns structured findings tagged by severity
- Blocking is corroborated — a second independent critic at the same place holds the merge
- If a critic times out, the panel continues with the critics that completed, and the verdict records exactly what was evaluated
- An independent, multi-family panel — heterogeneity by design
Give every change a panel of independent judges.
Install the GitHub App for the hosted panel, or the free CLI to run the same review on your workstation.