Same Causal Effect, Different Bill: Derivation Graphs and the Estimand Trap
A mechanism-first reading of derivation graphs, showing why equivalent do-calculus expressions can lead to very different estimators, costs, and operational decisions.
Cognaptus Insights
Industry thoughts, trends, and commentary.
Stale Gradients, Fresh Economics: CoCD’s Lightweight Route to Zeroth-Order AI
Coherent Coordinate Descent turns stale finite-difference gradients into a practical mechanism for lighter zeroth-order optimisation, with clear promise and equally clear scale boundaries.
Control, Alt, Generate: Why AI Needs Control Surfaces, Not Bigger Prompts
Two distant-looking papers show the same production lesson: generative AI becomes useful when teams can measure, constrain, and localise the behaviour that actually matters.
Furniture Has a Chain of Command: Why Dense Scene AI Needs Object Roles, Not One Bigger Generator
HetScene shows why dense 3D indoor generation improves when AI separates room structure from local object placement instead of treating every object as the same kind of token.
Judge, Jury, and Benchmark: Why LLM Evaluation Needs Fresh Cases, Not Bigger Leaderboards
CoEval shows how task-specific LLM evaluation can become renewable, contamination-resistant, and less dependent on a single judge model.
Lie Detectors Are Late: Why AI Oversight Needs Commitment Tracing
A mechanism-first reading of counterfactual localization, a method for finding when model reasoning shifts toward deception before the final answer exists.
No Easy A: Why AI Training Needs Hard-Case Routing
Two new arXiv papers show why production AI improves when scarce training budget is routed toward informative difficulty, not spread evenly across convenient data.
Raw Is Not Ready: Why Reliable AI Needs Evidence Architecture
A cross-paper analysis of why production AI reliability depends on structured evidence, calibrated uncertainty, and consequence-aware evaluation—not bigger models staring harder at raw inputs.
Source Code, Not Source Dump: Why Multimodal AI Needs Evidence Routing
A mechanism-first reading of MARS, a CASTLE Challenge system showing why long-horizon multimodal AI needs selective evidence control more than brute-force context stuffing.
Bidder Safe Than Sorry: Why Generative Auto-Bidding Needs a Fallback
A mechanism-first reading of Guide, a generative auto-bidding system that pairs exploratory Decision Transformers with conservative fallback actions and value-based selection.