---
title: 'In the News: September 21, 2026 (Extra 3)'
description: 'A code-review harness beat a single prompt in a new public benchmark, and a purpose-built fact-checker outperformed general LLM reviewers at catching bad claims.'
canonical_url: 'https://darkfactory.dev/news/2026-09-21-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-09-21-extra-3.md'
collection: news
date_published: '2026-09-21T23:20:00-04:00'
date_modified: '2026-09-21T23:20:00-04:00'
---

# In the News: September 21, 2026 (Extra 3)


Two new studies measure what actually happens when one AI checks another's work. One finds that a review harness beats a single prompt, and that an open-weight model running that harness matches a flagship model at a fraction of the cost. The other finds that a general-purpose reviewer model is not automatically a safety net, and that a small model built specifically to check facts beat much larger general reviewers at the job. Both publish their data and code.

## 1. A code-review harness beats a single prompt, and the cheaper model wins

**[Vibes vs. Evidence: What actually delivers AI code review quality](https://dsifry.github.io/harnesseval/)** · Dave Sifry · harnesseval, September 2026

Sifry ran eight models through three ways of reviewing the same pull requests, a single engineered prompt and two open-source review harnesses (Compound Engineering and metareview), at three reasoning-effort levels. Every finding was checked against tests that prove a bug is real: 42 bugs came from the benchmark's own human labels, and 105 more were confirmed by writing a test that fails on the pull request and passes on the fix, for 147 verified bugs across 2,416 review runs. The harness beat the single prompt in 39 of 42 matched comparisons, a mean gain of 13.5 percentage points of recall. The best-value setup, an open-weight model (GLM-5.3) running metareview at low effort, matched Opus 5 running the same harness at about a thirteenth of the cost. Raising the reasoning effort from medium to high showed no measurable quality gain in 17 of 22 head-to-head comparisons, while costing more in 20 of them. Sifry discloses that he wrote metareview, one of the two harnesses tested, and notes: "A review sends your source code to whoever hosts the model, and hosts' terms differ widely," a reminder to check a provider's data terms before routing code through an open-weight host.

**Why it matters:** This is a concrete argument for spending on the review workflow before the model tier. A cheap harness on an open-weight model outreviewed a flagship model working alone, and turning the reasoning dial up mostly just raised the bill.

## 2. A model built to check facts beat general reviewers at catching them

**[I Asked LLMs to Review Another LLM. They Still Got It Wrong](https://openteams.com/llm-review-reliability/)** · Khuyen Tran, OpenTeams · Sep 17, 2026

Tran tested five general-purpose models as reviewers of AI-written meeting summaries, hand-checking 35 claims against the source transcripts (23 supported, 12 not). Bigger reviewers caught more unsupported claims, up to 10 of 12, but the largest, qwen3:30b-a3b, also deleted 10 of the 23 true claims while doing it. A model trained specifically for fact-checking, bespoke-minicheck:7b, caught the same 10 of 12 unsupported claims as the best general reviewer, removed none of the true ones, and finished in about 10 seconds against a minute or more for the larger models. Only 2 of the 12 unsupported claims were caught by every reviewer tested, and one was caught by none. "I would not let a general reviewer silently delete claims on its own," Tran writes, after testing flagging and multi-reviewer voting as alternatives and finding each trades missed errors for lost content in a different way.

**Why it matters:** A general-purpose model asked to double-check another model's output is not automatically a safety layer. This is a case for choosing a reviewer trained for the specific judgment a verification step needs, rather than assuming a larger model is a better check.