Case studies

Seven recorded runs on an unsloth/Llama-3.2-1B-Instruct fine-tune, each with a defect planted on purpose. Of the seven, 4 confirmed what they set out to test, 2 were partly confirmed, and 1 was inconclusive. The strongest single result is the counterfactual retrain: a capability fell from 0.94 to 0.44, Gradian attributed 75% of the negative influence to a planted cluster of 51 examples, and removing those examples returned the capability to 0.94 exactly.

The runs that did not work are published with the reason. Two of them fail in the same way, and it is a limitation worth knowing before you trust any attribution result: a log-likelihood proxy metric can rise under a defect that genuinely damages the model, because a fine-tune mostly teaches answer format. Where that happened, the audit still named the real cause and the influence scores were meaningless. Use an exact-match metric.

Experiment D, unsloth and HuggingFace backend parity, has no recorded pass and so has no page here. The contract-level tests pass on CPU; the end-to-end comparison has not been run on a GPU. GPU validation covers how to run the suite yourself.