Deleting an example beat banning it across three rebuilds of a 680-line prompt

Nothing in the serving path tells a model that system tokens are rules. They are tokens in the same context window as the reply, and the next token is whatever the distribution favours. An exemplar phrase sitting a few hundred lines above the output raises the probability of that phrase, and a prohibition that mentions the phrase raises it too. That is the cleanest reading of why three rebuilds of an explicit ban changed nothing ^[4] and why the deletion landed on the next build ^[5]. It is also why the author warns that naming a bad example summons it ^[8]. The same echo produced a model saying "Morning." at an hour it had no way to check ^[3].

Natuworkguy writes that every finding in the piece falls out of that one idea ^[24]. The Python case says otherwise. Before he moved anything, he had already stripped every pipe union out of the prompt, and the model still emitted `str | Path` on every seed and every build ^[19]. Tokens that are not in the prompt are not being echoed from it. That behaviour comes from pretraining, where the union syntax is the current idiom, and it happens to raise `TypeError` on the Python 3.9 the prompt targets ^[18]. What fixed it was position: the same words, unchanged in meaning, moved from fourteen lines into the PYTHON section up to that section's opening lines, clean on all three seeds ^[20]. Two mechanisms, one writeup, and they imply different debugging moves.

Count the interventions on that one rule. Four rewrites, one prompt-wide strip of pipe unions, one added example signature: six edits that did nothing, then one relocation that worked ^[23]. If that ordering generalises at all, moving a failing rule is the cheaper probe, which is what the author now recommends ^[21].

The evidence behind each finding is not the same weight. The position effect has three observations: the Python case plus two more the author says he has watched since ^[25]. The delete-beats-ban finding is reported as fixed on the next build ^[5], a single rebuild, against a standing rule of three seeds before believing anything ^[17]. He paid for that rule: six revisions spent chasing a textbook-lecture format that only seed 7 was producing, while three other seeds had been clean prose for most of them ^[15][16].

For any of this to transfer you need his setup, which is a base model plus roughly 680 lines of prompt built into an Ollama tag with no fine-tuning ^[9], driven at a fixed prompt set with pinned seeds and read by eye ^[11]. The position finding needs a prompt long enough to have sections at all. The echo finding needs much less than that, only an example concrete enough to paste. If you keep full worked examples in a prompt, the shape that survives this evidence is fragments the model has to assemble, with any complete quoted answer attached to a question nobody asks ^[14]. That costs you something real: a style demonstration you can no longer read as a demonstration.