Staged Review Pipeline

Three parallel reviewer agents each reading the whole diff was a reasonable first design. It is also three times the strongest model's cost, and each reviewer sees the diff with no shared view of what is risky in it. RFC-0052 replaces that with one review pipeline that assigns each kind of work to the cheapest thing that can do it.

RFC: RFC-0052 — Staged Review Pipeline · Status: Signed off October 1, 2026; implementation in progress · Depends on: Judgment layer, Proof of Execution

The stages

1. Code does what code can decide. Diff facts, changed symbols, callers, coverage of changed lines. No model call.
2. The judgment layer classifies. What the diff touches and how risky each hunk is, as a risk map.
3. The strongest model writes a short plan from the risk map and the task's acceptance criteria. Bounded input, bounded output.
4. Cheaper models execute the plan's probes in parallel. Read, trace, run, compare. Each returns evidence.
5. The strong model reads the evidence and writes the verdict. Every finding cites the probe that produced it.

Some probes are mandatory and no plan can remove them, so a plan-driven review cannot look away from something because the planner did not think of it.

What stays the same

The verdict envelope, the aggregation, the attestation and the iteration loop are unchanged. The pipeline is a third reviewer set, staged, beside the existing three and code-test-merged sets, selectable per repository.

How it earns its place

Adoption is decided by replay evidence against the known-defect corpus, under the relax-class bar from RFC-0049 (a judgment that could reduce review needs 95% precision and has no override path). The staged set is not promoted because it is cheaper. It is promoted when the replay shows it catches what the three-reviewer set catches.

What this unlocks

• One review that reads the diff and its context once, with the expensive model used only where it adds something.
• Findings that are grounded in evidence a probe produced, not in a reviewer's recollection of the diff.
• A measurable path to lower review cost that cannot be taken on assertion.

Where to go next

• Cross-harness review describes the independence rule the staged set still has to satisfy.
• Reviewer independence is the tier the staged verdict is attested under.