Essays · commentary · synthesis · research notes
Long-form writing with explicit sources, scope and correction history. Articles may explain or develop research, but they are not certificate-backed Evidence Press releases.
commentary · 20 October 2026 · 15 min
Goodfire's October 2026 essay estimates that about 90% of Qwen3-32B neuron poles have one main job. A sourced account of the claim, the reaction and an Evidence Press audit of the released gallery: the labels are coherent by Goodfire's AI-reviewed rubric, but the gallery is not a whole-model sample and the developmental explanation is untested.
Contains bounded research claims · published
commentary · 10 October 2026 · 27 min
OpenAI's claim that the Riemann zeta function has no zeros with real part above 7/8: the Lean verification and its caveats, the anatomy of the proof, a graded reconstruction of how it was found, and what another system would need.
Exposition; no new research claim · published
synthesis · 3 October 2026 · 10 min
A synthesis of meta-learning and test-time training, with an unfamiliar-tool example and tests that distinguish adaptation, transfer and durable competence.
Exposition; no new research claim · published
commentary · 24 September 2026 · 8 min
What Anthropic’s array-associated reverse transcriptase discovery establishes about biology and AI research—and what remains unproved.
Exposition; no new research claim · published
commentary · 12 September 2026 · 6 min
A popular explanation of the candidate polynomial-composition rigidity proof, its critical-value argument, and what it does and does not settle for seven Evidence Press releases.
Exposition; no new research claim · published
commentary · 9 September 2026 · 24 min
An informed reaction to the September 2026 Navier–Stokes and Euler announcements: the exact claims, formal verification, disputed provenance, costs and implications for mathematical knowledge.
Exposition; no new research claim · published
essay · 4 August 2026 · 65 min
An evidence-backed essay on why verification capacity, provenance and accountable authorisation determine the useful scale of AI-assisted policy analysis.
Contains bounded research claims · published