Blog
Notes on what we measured
Every entry carries its date, its category, and how it came to be known — measured, observed, or stated. The same way the product stores anything.
arch26 August 2026measured
A word list beats embeddings at knowing a fact was reversed
We built the harness to detect contradictions with a similarity model. The reversal that must be caught scored 0.8593 and the negative that must not scored 0.8575, and all five derived statistics discriminated worse than the raw cosine. Here is the measurement, and what we shipped instead.
decision26 August 2026measured
The story of Knowl
A side project, a CLAUDE.md that would not stop growing, and five approaches that failed before one worked. The similarity model that lost by eighteen ten-thousandths, the subset test that fired backwards, and the guard that was obviously right and wrong on every case it touched.
fact21 August 2026observed
What is memory staleness in AI agents, and why decay doesn't fix it
A memory that was true when it was written and is wrong now is worse than no memory, because the agent has no way to tell. Here is why forgetting policies miss the ones that matter.
skill20 August 2026measured
A similarity threshold of 0.7 lets the clearest contradictions through
We found a supersession rule where the more completely a fact is contradicted, the likelier it is to survive. Here is the worked example that lands on exactly 0.7000, and what it cost on a benchmark.