run/capability/capability_summary.md

Capability retention: M0 (base) vs M1 (resist SFT)

arXiv:2511.21399 App. E.4 methodology — lm-eval-harness, MMLU 5-shot MC, GSM8K 8-shot CoT, greedy, accuracy on the test split.

benchmark	metric	M0	M1	delta (pp)
mmlu	acc,none	75.1%	76.5%	+1.4
gsm8k_cot	exact_match,flexible-extract	78.0%	79.5%	+1.5