https://wtk-uplink.dev/blog/Dated research notes, factory logs, field notes, findings, and failure reports from the WTK engineering workbench.en-usSat, 01 Aug 2026 00:00:00 GMThttps://wtk-uplink.dev/blog/one-poisoned-source/https://wtk-uplink.dev/blog/one-poisoned-source/Sat, 01 Aug 2026 00:00:00 GMTResearch NotesOur daily sweep of new AI research turned up a number that should worry anyone building research agents: 54.7%.https://wtk-uplink.dev/blog/all-engineering-no-science/https://wtk-uplink.dev/blog/all-engineering-no-science/Thu, 30 Jul 2026 00:00:00 GMTResearch NotesIn two case studies, frontier agents received six days and substantial compute, completed the engineering, and made no substantial progress on the research questions. A second model and scaffold reproduced the failure pattern.https://wtk-uplink.dev/blog/held-out-doesnt-mean-clean/https://wtk-uplink.dev/blog/held-out-doesnt-mean-clean/Wed, 29 Jul 2026 00:00:00 GMTResearch NotesIn one multimodal fact-checking study, up to 29% of post-cutoff claims remained potentially contaminated, and the effect changed system rankings. That challenges how comparative value is measured, including in WTK.https://wtk-uplink.dev/blog/verify-once-answer-forever/https://wtk-uplink.dev/blog/verify-once-answer-forever/Tue, 28 Jul 2026 00:00:00 GMTResearch NotesA frozen 12B model claims perfect accuracy at zero tokens by replaying verified answers. Big if true, and "if" is doing a lot of work.https://wtk-uplink.dev/blog/controls-that-cant-travel/https://wtk-uplink.dev/blog/controls-that-cant-travel/Mon, 27 Jul 2026 00:00:00 GMTResearch NotesA new paper puts agent decision-making inside the model's hidden states. That's where WTK requires a tighter boundary, and we can now say why.