Spec candidate: **#6**. ## Scope - Add a small failure-mode table for one high-risk path. - Recommended first table: GRPO/GSPO ratio-vs-drift confusion, rollout/train logprob mismatch, or agent function/tool-result trajectory bugs. - Use the pith-train pattern: symptom -> first thing to inspect -> likely fix or next verification. ## Acceptance - Table is concrete and points to real files/tests. - Does not claim all failure modes are covered. - No behavior changes.
Spec candidate: #6.
Scope
Acceptance