Preregistration: Extreme Weight Quantization of Small Language Models

Preregistered hypotheses, decision thresholds and analysis plans for a study on extreme weight quantization of small language models. The machine-readable thresholds (gates.json for stages S1 and S2, gates_s2b.json for addendum S2b) are the source of truth; the accompanying documents explain them. SHA-256 checksums of all files are listed in SHA256SUMS.txt and SHA256SUMS_S2B.txt, allowing anyone to verify that the analysis was carried out against these exact, unchanged plans. Timeline disclosure: the S1/S2 files were frozen and hashed on 21 October 2026 and uploaded here as a draft the same day, but a failed publication step left the record unpublished until 24 October 2026, after stages S1 and S2 had been run. Their integrity rests on the SHA-256 of gates.json, which is embedded in every result file. The S2b addendum was frozen on 24 October 2026 and published here before any S2b run. Version 1.0.3 adds stage Q1 (fully compressed models at equal total memory), frozen 24 October 2026 before any Q1 run.