Every time we add complexity, one question: did it earn its place?
A new representation can be elegant. A new agent can be impressive. A multi-model architecture can produce a gorgeous diagram. None of it matters if a smaller system gets the same result faster, cheaper and more reliably.
Freeze the question.
Define what success means before looking at the result.
Keep the baseline boring.
Prefer the simplest credible established method.
Hold variables still.
Same underlying model, same evidence, same evaluation conditions where possible.
Count the cost.
Tokens, compute, latency, human intervention, infrastructure, complexity.
Record failure.
Average accuracy isn't enough — a useful system also has to fail in ways you can understand.
Publish the receipts.
Where licensing, privacy and commercial constraints allow: outputs, prompts, configs, data, code, procedure, analysis.
Numbers reflect the latest published runs. Status labels are honest — most of this is early.
RETRIEVAL
Can relational or field-guided retrieval improve evidence selection over a vector-only baseline?
REASONING
Does explicit structured state improve reasoning consistency on problems with changing context and dependencies?
AGENT COORDINATION
Can agents with explicit state, roles and relationship structure outperform an undifferentiated multi-agent population?
MEMORY
Can structured relational memory hold task-relevant information with less context than conventional conversation history?
Evidence currently being assembled.
Raw outputs are being normalised and checked against the registered procedure. We'd rather publish slowly than publish a result we can't defend.
We don't scale assumptions. We scale evidence.
We don't remove an experiment because the result was inconvenient. A failure can mean the hypothesis was wrong, the implementation was poor, the benchmark was badly designed, the advantage lives somewhere else, or the extra complexity simply wasn't worth it. All five are worth knowing.
Nothing published here yet.
That doesn't mean nothing's happening. It means we haven't reached the point where this work deserves a public result.