FASLab publishes what its experiments, working systems and technical work are teaching us as the work goes allong the way.
The objective is not to publish only wins. FASLab's learning is most useful when it includes experiments that changed our thinking, questions that remain unresolved, and conclusions that contradicted the original hypothesis.
Portfolio reporting aims at giving a clear vision of what is happening. Our experiment explored a different problem: whether the value claims behind portfolio decisions can still be traced to evidence, assumptions and explicit decisions.
LLMs can interpret governance evidence, but the scoring model should not originate from model intuition. We experimented with keeping the framework, evidence and scoring logic inspectable.
Process optimisation with AI becomes unreliable when the current process is only partially known. We found that process assessment requires a factual baseline before inference can safely become redesign.
Our multi-model experiment gradually stopped being a model-routing problem. The useful question became **how each task should be executed, including cases where one model, several models or no LLM at all may be preferable.**
Long-running professional AI needs more than a large context window. We established several areas from knowledge, working context, engagement memory to behavioural memory and feedback because each needs a different lifecycle.
Instead of treating conversation history as simple learning, we experimented with extracting patterns from completed work and deciding which ones deserved to influence future work.