---
title: 'In the News: October 19, 2026 (Evening)'
description: 'Kiln says an autoresearch loop that rewrote the code around a classifier cut its errors on one task to 147, against 285 for a reasoning model.'
canonical_url: 'https://darkfactory.dev/news/2026-10-19-evening'
markdown_url: 'https://darkfactory.dev/news/2026-10-19-evening.md'
collection: news
date_published: '2026-10-19T19:30:00-04:00'
date_modified: '2026-10-19T19:30:00-04:00'
---

# In the News: October 19, 2026 (Evening)


Kiln AI reports that a coding agent improved a classifier by editing the code around it, not the prompt, and ended with half the errors of a reasoning model at a seventh of the cost. It is one task and every figure is the vendor's own.

## 1. Kiln: an autoresearch loop cut a classifier's errors by rewriting its harness, on one task

**[Auto-optimizing Jev: half the errors, 1/7 the cost](https://kiln.tech/blog/auto_optimizing_jev_with_autoresearch)** · Leonard Marcq, Kiln AI · kiln.tech, September 21, 2026

Kiln pointed a coding-agent research loop at Jev, a closed-set classifier from TypeSafe that returns one of the options you give it, with probabilities, and no text. The task was choosing the right pinyin reading for polyphonic Chinese characters, scored as mistakes on 61,750 positions. Kiln reports 316 mistakes for its first Jev request, 285 for DeepSeek V4.1 Flash with reasoning on ($11.00 a pass, 8,500 ms median latency), and 147 for the loop's final configuration ($1.60 a pass, under 300 ms). A phrase-table baseline (pypinyin) made 1,434. Kiln puts the whole campaign at about $18.

Hand-written, step-by-step instructions did not help, because Jev does not reason. What worked was code. The loop segmented each sentence with jieba, looked up every dictionary word covering the marked character in CC-CEDICT, and passed the candidate readings in as "word evidence". It also chained a second, cheaper Jev call for some characters. Kiln says the word-evidence change, developed on twelve characters, took twenty unseen ones from 97 mistakes to 11 with none made worse. The post says a change was kept only if it improved a held-out half by more than about fifteen positions, the variation Kiln saw between repeated calls.

The post is Kiln's own, and it ends by inviting readers to contact Kiln about optimizing their own agents. We read the post body; its expandable detail sections, the benchmark and the code were not examined, and nothing was re-run.

**Why it matters:** The reported gain came from deterministic code and a measurement rule, not from better wording, and the rule that a change must beat run-to-run noise on held-out data is worth copying. The result is for one narrow task with a model that cannot reason, so treat it as a pattern to test on your own evals, not a general claim about harness optimization.