Review Is a Separate Program

Review needs a boundary stronger than “look again carefully.” The authoring session carries intent, shortcuts, and explanations that make its own choices feel inevitable. A separate reviewer executable starts cold, receives the frozen artifact and approved gate, and produces an attributable result.

The program boundary matters operationally. The workshop can record command identity, start and finish, output artifact, and availability failure. It does not need a web chat or a simulated persona. If a reviewer has no supported command surface, the system should say unavailable rather than inventing a round.

Independence is about context and role, not model mythology. A different model family can still receive a leading prompt; the same vendor can still provide a useful cold executable. The reviewer request should ask for concrete defects, locations, consequences, and a blocker/observation classification.

The author remains responsible for judgment. Review output is evidence, not an automatic patch queue. Every finding must be verified against the artifact and task contract before repair, because an independent reviewer can be confidently wrong.

Run, Capture, Attribute

Build the reviewer command from a known capability registry, pass a prompt file, capture output to a durable review artifact, and record failure distinctly from an empty result. Never hand-write a shell dispatcher from untrusted task text.

Then read the output as claims. A review that returns no actionable analysis may be a valid clean result only if the request and artifact were actually inspected; an unavailable or malformed run is not a clean review.

Independence begins with a cold artifact and a separate process. The reviewer receives the approved gate, exact files, and evidence claims without inheriting the author's live reasoning stream. That separation creates a real chance to notice contradictions the author has learned to read past.

Code

Capture a reviewer as a distinct subprocess·python

import subprocess

result = subprocess.run(
    ["python", "-c", "print('BLOCKER: missing acceptance evidence')"],
    check=True,
    text=True,
    capture_output=True,
)
review = {"executor": "python-demo", "status": "complete", "output": result.stdout.strip()}
assert review["output"].startswith("BLOCKER:")
print(review)

Review Is a Separate Program