Work through them in order or pick the one matching your symptom. Every module is a failure we hit, what it cost, and the check that would have caught it in seconds. Evidence lives on the findings page rather than being repeated here.
Before you spend GPU hours: is refusal removal even possible on this checkpoint?
5 min · start hereThe run completed, the metrics looked fine, and it still will not answer. Almost always the same cause.
8 min · essentialMost reported degradations are smaller than the measurement can resolve. Including one of ours.
10 min · read this firstThe study completes, then the save dies. Both causes are visible in a config file before you start.
6 min · saves a dayThe word list you score refusals with decides how much damage the process does. Topic words are worth 4x to 13x.
7 min · biggest single winThe naive four-bit estimate is wrong in a way that decides whether a model fits your card.
6 min · before you buyRefusal removal — turning off a model’s trained tendency to decline requests. Quantization — storing the model’s numbers in fewer bits so it fits on smaller hardware; “4-bit” is the common setting here. KL divergence — a single number for how differently a modified model behaves compared with the original. Lower means less changed. Noise floor — how much a score moves between two runs of the same test, purely by chance. Differences smaller than it mean nothing. Checkpoint — the files a model is distributed as. Chat template — the formatting a model expects its input wrapped in.
Module 03. The noise floor argument applies to any benchmark you run, not just this pipeline, and it is the one where we published a claim and then had to retract it.