agami-evallisted

# agami eval You are running the golden evaluation harness. Goal: take a dataset of questions whose answers are already agreed, have the model answer each one again from today's semantic model, execute both statements, and report where the two disagree. A dataset is the profile's own regression suite — the run says whether the model still answers the questions this team already signed off on, and each disagreement is a concrete thing to fix (a drifted metric, a missing filter, a join that changed shape). This skill orchestrates: 1. **List** the profile's datasets, before anything expensive runs. 2. **Choose** one — the only one, the one named, or the one the user picks. 3. **Run** it via `scripts/run_golden_eval.py`, which generates, executes and scores every case. 4. **Present** the verdicts in the order the script emitted them, failures first. 5. **Point** at the run's report — and the JSON beside it — for the drill-down the terminal deliberately withholds. Spec for the deterministic half: [`scripts/run_golden_eval.py`](../../scripts/run_golden_eval.py) (dataset choice + schema rendering + the printed payload). The scoring itself is agami-core's. ## Conversation style - **A run with failures is a SUCCESSFUL run** — the failures are the whole value. A 9/12 run is the run that just told this team three things about their model. Lead with what did not pass; never soften it, never bury it under the passes. - **Tight loops.** This skill is a tool, not a tutorial. One quest