What problem does it solve? After an ML experiment writes a skore report, extracting its checks and metrics for narrative review usually means ad-hoc scratch scripts that risk mutating the project. This Skill standardizes that step: one durable, read-only audit file per experiment, executed by a bundled in-process runner that streams a markdown digest of each cell's output. ## Core Features & Use Cases - Read-only report auditing: Opens the skore Project, loads a report by id, and renders report.checks.summarize().frame() and report.metrics.summarize().frame() — never calling skore.evaluate(...) or project.put(...). - In-process cell runner: scripts/run_cells.py executes jupytext # %% files via IPython's InteractiveShell.run_cell, capturing stdout, stderr, and last-expression reprs into a markdown digest streamed to stdout or written to scratch/audit/<stem>/audit.md. - Four-way stem pairing: Each audit file aligns 1:1 with journal/NN_*.md, experiments/NN_*.py, and tests/smoke/test_NN_*.py, keeping the audit trail durable in git. - Use Case: After experiment 02_target_transform finishes, ask the agent to audit it — it places audit/02_target_transform.py from the template, runs the bundled runner, and surfaces the checks and metrics digest for the JOURNAL update. ## Quick Start Ask the agent to audit experiment 02 and show its checks and metrics digest from the stored skore report.
