simba-results-analysis

Analyse saved results

Reuse guidance sections already supplied in context. Fetch only missing sections.

Use any supplied model identifier directly in get_model_results. Do not call list_models or another discovery/status tool first: the result read checks the identifier. Discover only if no identifier was supplied or that read rejects it, and only when discovery is authorised.

Choose sufficient sections:

• Revenue/ROI: channel_summary; coefficients for individual periods.
• Diagnostics: model_stats and r_hat for a screening rule. If the question contrasts that screen with a coefficient or posterior summary, also request posterior before describing the summary. Do not infer the summary from r_hat names.
• Decomposition: contributions, model_config and channel_map. Verify which components are media before classifying them; names alone are insufficient.
• Current/historical marginal ROI: mroi_summary/mroi_periods respectively.
• Add channel_map for unverified identity or media classification.

Use requested dates, never guessed bounds/year. Inspect dated rows when needed. When the answer names a channel, include channel_map in that same read. Pass granularity only when the question asks for a calendar bucket. The token is week, never weekly. When the question names JSON fields, the answer is only that JSON object. No prose before or after it. A requested section absent from the returned results is not an unknown quantity. For availability, use false and reason not_returned unless the payload already gave a documented reason. For convergence, use unknown and reason not_returned. Null is only for a numeric field the returned evidence does not establish. Before answering, identify every requested output, including combined totals as well as individual values. Supply each once; check none is missing. Keep the requested explanation within 100 words unless more detail is requested or needed for coverage. Check that each formula describes the metric it labels; omit unrelated formulas. Stop once answered: no unsolicited methodology, tool-use or safety footer. Before sending, check every explanatory clause against a returned field, a shown calculation or the documented contract. Remove plausible but unestablished interpretations. If actions are explicitly requested, check the actual call record, including refused attempts; never reconstruct motives.

Apply interpretation; consult the full contract for other sections/parameters.

Role availability: all profiles can inspect saved results and native views with JSON fallback. Executable view examples preserve missing points and KPI units. Campaign and actual-data reporting use their focused Skills or the campaigns/reporting guidance topics.