Assert Less, Observe More: AICL and the New QA Stack for LLM Apps
A practical reading of AICL as a testability layer for LLM applications, where QA shifts from exact assertions to observable, replayable behaviour.
Cognaptus Insights
Industry thoughts, trends, and commentary.
From Chat Logs to Goal Logs: OnGoal’s Playbook for Goal‑Truthful LLMs
OnGoal shows why long LLM conversations need goal observability, not just longer context windows.
Prolog & Paycheck: When Tax AI Shows Its Work
A mechanism-first reading of why tax AI becomes more useful when LLMs translate rules into executable logic, defer when uncertain, and price mistakes like real liabilities.
Rollouts, Not GPUs: Why AWorld’s 14.6× Speedup Rewires Agent Training
AWorld shows that the practical bottleneck in agent training is not only model capacity or gradient compute, but scalable experience generation.
Vitals, Not Vibes: Inside the New Anatomy of Personal Health Agents
A mechanism-first look at why personal health AI needs specialised agents, grounded tools, orchestration, memory, and evaluation before it can move beyond wellness chatbot theatre.
Benchmarks with Benefits: What DeepScholar-Bench Really Measures
DeepScholar-Bench shows that research agents should be judged by coverage, source quality, and citation support—not by how convincingly they format a report.
Edge of Reason: Orchestrating LLMs Without a Conductor
A mechanism-first look at Symphony, a decentralised multi-agent LLM framework that routes reasoning work through capability matching instead of a central conductor.
Faking It to Make It: When Synthetic Data Actually Works
A practical map for deciding when generative synthetic data is useful, when it is theatre, and what must be evaluated before it touches production.
MoE Money, MoE Problems? FinCast Bets Big on Foundation Models for Markets
FinCast shows how a finance-specific time-series foundation model can improve forecasting accuracy, but not yet prove tradable alpha.
Who Watches the Watchers? Weak-to-Strong Monitoring that Actually Works
A practical reading of monitor red teaming for LLM agents, showing why scaffolding and escalation policy matter more than simply giving monitors more context.