Three cards, two brains
Reconstruction⚠ This is not a live capture: it is a condensed reconstruction, edited afterwards from the real transcripts of two sessions, on September 25 and from October 4 to 6, 2026. Ludo's prompts are quoted word for word, typos included; tool output is rewritten and shortened, the timing is compressed and names are sanitized.
One September evening, from a phone, a simple question: what takes time when you talk to the house’s voice assistant? Ollama’s log answers that every question re-reads its whole prompt, more than 5,000 tokens. The agent first blames the model’s hybrid architecture, “probable cause, not proven”. The dense model’s bench contradicts it: the cache works, and what changes at every question is three lines at the top of the voice prompt (the temperature, the time to the minute, the forecast).
The weather becomes a tool, the re-read drops to about twenty tokens, and the dense model replaces the mixture-of-experts model everywhere. Then the site’s chat is suspected of emptying the voice’s cache; the proposed remedy, two processing slots, runs into a warning in the log: this model only gets one.
Nine days later, while the articles are being written, the measurement script
shows what prompt_eval_count was hiding: the time at the end of the system
prompt is not enough, because a hybrid model can only resume at a checkpoint.
The agent corrects two things it had stated on September 25, then a third: the
cache was not emptying, it was warm for the wrong prompt, and the site’s chat
had evicted nothing.
The short version is in A Slower Brain for Bob; the theory, the logs and the confessions, in Three Cards, Two Brains and a Clock in the Wrong Place. The measurement itself is shown for real in The clock in the prompt, for real.