Things we measured, things we got wrong, and rules that turned out to
generalise. Written to be useful independently of our system — if a note is
only interesting to someone who already cares about Entropy Box, it does not belong
here.
01 Embedding similarity cannot decide duplication
We used 1,155 adjudicated near-duplicate pairs from our own registry
as labelled data. Once the gate has flagged a pair, precision is 0.069 and AUC is
0.629 — and raising the threshold does not repair it. Why a conservative
adjudication stage is a requirement rather than a safety margin, and why we could
only run the experiment because the decisions had been logged with their
reasons.
negative result · entity resolution · dataset published
02 Merge on reasoning, split on artifacts
Two multi-agent architectures failed before the third worked. The
rule that explains both failures: merge two stages into one context when what
passes between them is reasoning; split them when what passes between them is an
artifact. Multi-agent decomposition is usually justified by capability
specialisation — we think the binding constraint is what survives the
handoff.
multi-agent design · context engineering · two failures