doubao-data-analysis

Attach the files and state your question, such as filtering records, checking thresholds, or explaining a metric change. The Skill reads each attachment, registers a delivery contract, and validates every conclusion against the source data before answering.

Use the bundled analysis_toolkit.py ab_paired_binary command with the four-cell counts (both correct, treatment only, control only, both wrong). It returns an exact two-sided McNemar p-value and a paired bootstrap confidence interval, plus Holm correction via p_adjust for multiple comparisons.

Yes. The cashflow_audit.py script mechanically recomputes inflow subtotals, outflow subtotals, and net cash flow for each activity section in an .xlsx cash flow statement, reporting any FAIL rows where source values diverge from recomputed values.

Yes. The time_window_audit.py script classifies records as point events, non-overlapping periods, overlapping periods, cumulative snapshots, or unknown. Overlapping or cumulative data is never split into fake daily records or used for before-after attribution.

The Skill separates provable hits, provable misses, and unknowns, then reports a strict count interval instead of inventing a precise number. Proxy fields may be listed as a separate labeled scope but are never renamed as the user's original criteria.

When the actual current configuration is unknown, the decision_guard.py script blocks absolute targets and numeric relative changes. Only verified configuration evidence, user risk budgets, platform minimum steps, or response-curve data can justify specific adjustment magnitudes.