Journal · October 23, 2026
V2: what changed and what the tests showed.
The design sketch at the bottom of the v1 architecture page listed four things: match the training and serving prompt formats, log the full prompt, evaluate before claiming, and fix the corpus balance. V2 did all four. Training rows are now the verbatim prompt the server sends, every request logs the assembled context so a bad answer can be replayed exactly, and the evaluation battery exists: 147 rows, a 16-question semantics battery, and a fabrication sweep of ten questions where the answer is known to be unobtainable.
The part I did not expect was how much the runtime, not the weights, decided what it could say. Early v2 answered "what is my supervisor's name" with an invented professor, an invented university, then an invented first name, each time phrased as if its files said it. The fix was not a prompt asking it to be careful. The runtime now searches its files and writes down what it found, including the line "no line matches". Absence became something it can read. For questions outside the files it checks the outside world first, with sources labelled by reliability, and only then is "I don't know" the answer. A number leaves the machine only if that number appears in the fetched lines.
Two of my own mistakes are worth recording. First, the thinking cue I wrote told it things like "human words import human outcomes". That is a school of thought, not a process, and installing it would have made it recite my method instead of thinking. It is gone; the cues now say only to think first, use the facts, and not restate stored paragraphs. Second, I harvested its declines to teach honesty and got the opposite: 69 of those declines were on things it actually had, and the next training run collapsed arithmetic. Training the habit of hesitating is not the same as training the judgment of when to hesitate. The correction round keeps a decline only when its logged facts prove nothing was there. After that: the fabrication sweep went 0 for 10, and the battery returned to the base band with no fabrication in the failures.
What still fails, short and honest: arithmetic past two digits, tracking a fact across four turns, and answering an insult with anything better than a stance line. The opinion and smalltalk rows from v1 are going back into the corpus for exactly that reason, so this round's claims end here. When I run the extended evaluation session, I will write a separate entry with those results.