The Goats in the Machine: Why AI Agents Need Contracts, Not Personalities
A practical reading of two new agent papers showing why enterprise AI should be judged by observable behaviour and runtime contracts, not human-like performance theatre.
Cognaptus Insights
Industry thoughts, trends, and commentary.
Bond Before Brain: What Actually Drives Molecular MPNNs
A mechanism-first reading of a factorial molecular GNN benchmark showing why message construction deserves more attention than architectural nameplates.
Cheap Seats, Sharp Eyes: Reward-Hack Detection Without the Frontier Judge
A small trajectory encoder nearly matches a frontier LLM judge on reward-hack detection, but only when it can read the reasoning-rich trace.
Judge, Jury, and Calibration: Why AI Evaluation Needs Anchors
A practical reading of two arXiv papers showing why AI works best in high-stakes evaluation when it is anchored to human evidence and audited for real engagement.
Pre-Review, Not Peer Review: The Drafting Gate AI Actually Earns
A mechanism-first reading of AI-Paper-Review shows why AI review is useful as a pre-submission quality gate, not as a substitute for human peer review.
Split Before You Scale: Why Useful AI Starts by Sorting the Mess
A business-focused reading of three arXiv papers showing why scalable AI depends on decomposing structure, uncertainty, and supervision before optimisation.
Statecraft, Not Scorecards: Why Reliable AI Lives on the Path
A practical reading of two arXiv papers showing why AI reliability depends on the states models visit and the trajectories evaluators inspect.
The Chatbot Passed the Test. Then It Bowed Too Low.
NICE shows why aggregate social-intelligence scores can hide the communication failures that matter most in real deployments.
Feedback, Not Freefall: Why LLM Writing Tools Need a Teacher in the Loop
A large K-12 writing study shows that LLM feedback works best as a teacher-mediated workflow, not as a replacement chatbot with better grammar.
Frame Before You Aim: Why AI Needs the Right Reference Point
A cross-domain look at why useful AI systems need adaptation layers that translate models, protocols, and rankings into the realities they are meant to serve.