https://the-observatory-ledger.pages.dev/ A public research notebook on AI safety and mechanistic interpretability. Research that corrects itself. en-gb © 2026 Pano Pouroullis Mon, 15 Jun 2026 12:00:00 +0000 Hand-written static feed https://the-observatory-ledger.pages.dev/posts/cross-graph-cosine-collapsed.html https://the-observatory-ledger.pages.dev/posts/cross-graph-cosine-collapsed.html 0.921 was the graph, not the country. A beautiful within-graph signal evaporated under a decisive prompt-variant test, and the null result became the finding. Attribution graphs Null results Fri, 15 May 2026 09:00:00 +0000 https://the-observatory-ledger.pages.dev/posts/the-correction-came-from-outside.html https://the-observatory-ledger.pages.dev/posts/the-correction-came-from-outside.html What a geography error taught me about AI verification: agreement inside one model family is consistency, not correctness. The correction came from a different family. Verification Corrections Tue, 17 Mar 2026 09:00:00 +0000