Published and updated 7 September 2026 · Altari Systems

Agent governance research

Exact-SHA Review for AI-Generated Code

Abstract

Exact-SHA review binds a review decision to one immutable Git commit. This prevents an AI coding agent from adding or rebasing code after approval while retaining the appearance of a reviewed branch. Nexus uses exact code identity as a prerequisite for evidence-backed, canonical integration.

01

A branch is not a review identity

Branch names are references that can move. If a reviewer approves a branch and a later commit changes it, the apparent approval no longer describes the code being integrated.

A commit SHA names the reviewed content and its history precisely. Integration can compare the accepted identity with the candidate it is about to advance.

  • Record baseline SHA
  • Record candidate SHA
  • Attach evidence to candidate
  • Invalidate review after change
02

Failure modes exact identity prevents

Floating approval permits post-review changes, stale test results and mistaken integration after rebases. Text summaries cannot detect these conditions reliably.

Exact identity does not prove correctness by itself; it makes the evidence and decision refer to the same object.

  • Approval drift
  • Evidence from older code
  • Late worker delivery
  • Wrong branch head
03

Nexus implementation context

Nexus separates Builder, Review and canonical integration authorities. A reviewer challenges the exact candidate and records findings; remediation creates a new candidate that must earn new evidence.

This invariant supports independent review for AI-generated code without trusting a worker's narrative.

Nexus launch access

Be first to know when Nexus launches.

Register your interest now and we will let you know when the autonomous AI software engineering control plane is ready.

Express Interest ↗