We Measured a Values Fine-Tune. It Removes the Behaviour, Adds No Knowledge, and Costs Nothing on Our Translations.

On 6 November Nous Research released Arete-27B, Qwen3.8-27B fine-tuned to stop taking the Chinese state’s side. We ran it against its base on the Mac mini for a day, both as the same community Q4_K_M files, every answer graded blind. The question was the one anyone with a local Qwen asks: does a values fine-tune help, and what does it cost? Three things were proved, one was not, and one line answers the question. ...

November 8, 2026 · 14 min · Nestor

A Second AI Doubted 1 in 5 Test Enquiries. We Deleted Those.

Correction, 24 October 2026. Section 2 gave exp_040’s counts from an earlier filter run, and the original title (“Two AIs Disagreed on 1 in 5 Test Enquiries”) and summary called the deleted enquiries disagreements. Both now say doubted. The trained set. exp_040 trained on 2,267 items. Sampled six times each, the untrained model got 1,296 right every time and 349 wrong every time, which left 622 items with any spread. The 2,104, 1,281, 270 and 553 we gave are from filter run 2, before 163 abstention items were added (exp_040 BUILD.md, results/training/filter_manifest.json). The base still solved most items: 57 %. What the 85 were. In exp_043 all 85 deleted enquiries were ones the second reader flagged as doubtful. The two readers disagreed on the request in 6 of them, and on no label at all in 70 (exp_043 data/raw/, data/chk/, and HYPOTHESIS.md: “85 dropped on checker doubt”). In exp_044 all 98 were dropped on doubt. Records: exp_040 filter manifest, exp_043 pre-registration. ...

October 20, 2026 · 10 min · Nestor

We Trained It Three Times. Then We Stopped.

Correction, 24 October 2026. The glue met 5 of the 7 format floors on the 60 questions, not 7, and two of its floors failed on the fresh 30-question set this post promised to report. Format floors. The table gave the glue 7 of 7. Its judge file records 5. It missed answers citing (43 of 60, floor 90 %) and artefact rows (15, floor at most 5). Source: designA_A12_judge_20260922.json, field reference_recommendation.preregistered_floors_status. Against v2. v2 met both of those floors: 55 of 60 answers cited, and it had 2 artefact rows (b3_v2_judge_20260922.json). The glue did reach the two floors no fine-tune reached, but it did not take “every format floor the fine-tunes bought”: it lost these two. v2’s count. The record’s summary table gave v2 6 of 7. Its v2 judge file and verdict give 5: v2 missed out-of-scope (2 of 3) and Spanish term (89.6 %). The table and the v2 paragraph now say 5. The fresh set. The one re-score ran on 13 October, on 30 sealed questions the glue had never seen. Two floors failed: out-of-scope declines 0 of 2 and language 29 of 30. Both out-of-scope questions passed the scope gate and got invented answers. Plain text (30/30), hand-off (28/28), numbers (98.9 %) and citations in context (97.2 %) held. Under the pre-registered rule those two floors were fitted to the first 60 questions, and the claims now stand at 0/2 and 29/30. The judged half of the re-score is not in the record yet. v1 citations. 98 % was the share of v1’s citations found in the retrieved text. 50 of 60 v1 answers cited anything. Q8_0. The Q8_0 diagnostic ran on v1, not v2. The checkpoint and temperature diagnostics ran on v2. Record: HYPOTHESIS.md (blocks “Diagnostics after B3 v1”, “A12 judged” and “A12 on eval v2”) ...

September 22, 2026 · 16 min · Nestor