Ephemerent News https://ephemerent.com/news Research briefings, technology, and evidence. en-us Fri, 04 Sep 2026 12:00:00 -0400 https://ephemerent.com/assets/og-news-swc.png Ephemerent News https://ephemerent.com/news What a checked answer can promise. https://ephemerent.com/news/stochastic-witness-calculus https://ephemerent.com/news/stochastic-witness-calculus Fri, 04 Sep 2026 12:00:00 -0400 kt@ephemerent.com (Kenju Tomita) Research Dispatch Automated reasoning An intuitive guide to Stochastic Witness Calculus and the boundary between learned proof search, exact certificates, and empirical evidence. A model can help search an enormous space for a proof witness. It should not get to certify its own work.

Read the full Research Dispatch at Ephemerent News.

Read the public v6 preprint at Ephemerent Research. Updated September 5: exact-output receipts, complete coverage of 12,673 declared inputs, and 4,603 fresh generated inputs with zero failures. The availability result is specific to the recorded IID model and frozen solver; no physical deployment was evaluated. The paper, LaTeX source, and full evidence are downloadable.

]]> A simpler rule beat the hierarchy on GPT-2. Here’s what that means. https://ephemerent.com/news/drfsp-robust-compression https://ephemerent.com/news/drfsp-robust-compression Sun, 09 Aug 2026 12:00:00 -0400 kt@ephemerent.com (Kenju Tomita) Research Dispatch Neural network compression A controlled experiment rewarded a sophisticated compression hierarchy. A small GPT-2 test did not. The reversal became the real result. I built a sophisticated hierarchy to compress neural networks without erasing rare-domain behavior. It worked perfectly in a controlled setting. On GPT-2, a much simpler robust baseline did better at aggressive compression—and that became the real result.

Read the full Research Dispatch at Ephemerent News.

Open the peer-reviewed research record at Ephemerent Research.

]]>