A Retraction Log: What Broke While Measuring Small Language Models on One Home PC

Across five days of measurement on a single consumer GPU — 2026-08-20 to 2026-08-24 — twenty-two claims I had written down were retracted. I kept the retractions in a fixed schema at the moment each one died, including a field that cannot be reconstructed afterwards: why it looked true. Nine of the twenty-two were caused by my own instruments: programs that reported something which had not happened, designs in which the answer leaked into the item, scoring keys and scorer vocabularies that were wrong, and a scorer and a generation budget that could not be trusted to treat two models alike. Five were caused by changing the framing of a task and watching the result reverse. Two were caused by stating a direction from a single-digit sample. Four belong to a class version 2 did not have, and it is the one I would not have predicted: the number was right and the sentence was not. In these the figures recompute exactly, and the prose around them names the wrong variable, the wrong population, or an alarm that the test had already answered. Nothing about re-running the analysis catches them, because the analysis is correct. The remaining two are an arm dropped without saying so, and a face added to a standing hypothesis on resemblance rather than on a measurement that separates it from its neighbours. The pattern is not that small-scale research is sloppy. It is that at this scale you are your own instrument-maker, and the instrument fails in ways that look exactly like findings. I am publishing the log rather than the lessons, because the lessons are short and the log is what makes them believable. If you operate retrieval-augmented or agentic LLM systems: cases in this log measure, with reproducible instruments, how a single retrieved line in the queried field's format turns abstention into confident assertion, and which side of the pipeline (data, not prompt) repairs it. The instruments run on one 16 GB consumer GPU. The author can be reached via the profile on this record.