---
title: 'In the News: October 12, 2026 (Extra)'
description: "QGIS now bars AI-written code contributions outright, and a paper trains a small model to rewrite an agent's harness at test time."
canonical_url: 'https://darkfactory.dev/news/2026-10-12-extra'
markdown_url: 'https://darkfactory.dev/news/2026-10-12-extra.md'
collection: news
date_published: '2026-10-12T10:04:32-04:00'
date_modified: '2026-10-12T10:04:32-04:00'
---

# In the News: October 12, 2026 (Extra)


Two items. The QGIS project has replaced its human-in-the-loop AI rule with a flat ban on AI-written code, citing reviewer load. A paper on training a model to rewrite an agent's harness at test time reports gains on reasoning tasks, not software engineering.

## 1. QGIS replaces its AI disclosure rule with a ban on AI-assisted contributions

**[Add QEP428: No ai contributions permitted](https://github.com/qgis/QGIS-Enhancement-Proposals/pull/387)** · nyalldawson and QGIS maintainers · QGIS Enhancement Proposals, GitHub, October 2026

The open-source mapping project QGIS has merged a proposal that, in the author's words, "formally changes the AI policy to a strict 'NOT ALLOWED' policy." It supersedes an earlier policy, QEP 408, under which AI-assisted work was allowed with a human in the loop. The approving votes were declared from rouault-on-strike (Even Rouault), alexbruy, signedav and elpaso; 3nids registered +0 and said he would open a separate proposal.

Commenter haubourg, who backed the ban, gave the main reason in the thread: AI use from external contributors puts "massive and unbearable pressure" on maintainers, and the volume is higher than human review can take seriously. Commenter uclaros reported that under the old policy about 50 pull requests were tagged AI Assisted and roughly 40 were closed without merging, and that the "extractive" tag the policy called for was never created. nyalldawson wrote that the earlier policy clearly did not work and had already cost the project at least one very valuable contributor. He framed the ban as temporary: "lets start with a closed door," to be opened case by case later.

The wording is contested. One commenter (02JanDal) asked whether LLM autocomplete, machine-learning completion ranking or brainstorming with a chatbot counts, and proposed limiting the rule to code "written" by an LLM or agent. Another asked for documentation, bug reports and comments to be covered too. timlinux's answer on enforcement was that the test is whether you wrote the code yourself in good conscience, since the project cannot inspect contributors' machines.

**Why it matters:** This is a maintainer group deciding that disclosure and tagging did not contain the review cost, and moving to a rule they can only enforce on trust. If you run agents that open pull requests against other people's repositories, this is a documented case where that stops being welcome, and the reason given is review capacity, not code quality. The quality claim is untested here; nobody in the thread measured it.

## 2. A small model trained to rewrite a solver's harness beats its larger teacher on average, on reasoning tasks

**[Harness Learning Enables Generalizable Test-Time Adaptation](https://arxiv.org/abs/2609.35738)** · Alvin Zhang, Xuecheng Liu, Zixuan Wang, Fahim Tajwar, Daman Arora, Ruslan Salakhutdinov, Daniel Khashabi, Yuda Song, Andrea Zanette · arXiv, submitted September 28, 2026

The authors train a proposer model with reinforcement learning to edit the executable harness around a frozen solver, using execution reports as feedback. No weights change at test time; the harness revisions play that role. On 21 Reasoning Gym task families the proposer never saw, a 4B proposer's single revisions averaged 0.62 against 0.32 for the untrained base. They also scored above the 35B teacher's 0.56 on average, though the teacher did better under oracle best-of-eight selection.

In multi-hop question answering, a proposer trained on HotpotQA carried over to MuSiQue and 2WikiMultihopQA. On MuSiQue, independent revisions from the single-step proposer averaged 0.15 exact match, barely above the seed harness, against 0.27 after ten rounds of building on the current harness. The authors report that training on whole revision sequences, as opposed to single revisions, helped in some settings and not others. On Reasoning Gym the learned harnesses were mostly interpreter loops, and generated helper components contributed to the answer in only a minority of harnesses that integrated them. The paper names its limits: a fixed solver, prescribed revision prompts and one teacher.

**Why it matters:** The tasks are reasoning puzzles and retrieval QA, not coding agents, so nothing here shows that a learned proposer improves a software factory's harness. What it does show is that harness edits driven by execution feedback can transfer to tasks the proposer was not trained on, and that repeated revision beat independent samples on the harder QA benchmark. I did not find a code release.