Skip to content

SKILL.md

Latest commit

 

History

History
53 lines (44 loc) · 2.65 KB

SKILL.md

File metadata and controls

53 lines (44 loc) · 2.65 KB
Raw
name simba-optimizer-runs
description Run Simba budget optimizations correctly over MCP — payload conventions (percent bounds, laydown/CPM arrays), revenue vs profit objectives, polling by run_id, interpreting decision vs comparison columns, and curating run history. Use when optimizing budgets or reading optimizer results through the Simba MCP tools.

Simba optimizer runs

Payload rules (the four that reject or mislead)

  1. Channel keys = ACTIVITY-COLUMN names, case- and space-sensitive. Get them from channel_summary or get_scenario_template (avg_cpu_by_channel doubles as baseline CPM).
  2. bounds are PERCENTAGES of total_budget (0–100), not currency.
  3. laydown_weights and period_cpm are ARRAYS of length num_periods ({"TV": [10, 10, 10, 10]}, never {"TV": 10}); CPMs strictly positive; the same channel keys must appear in bounds, laydown_weights, AND period_cpm.
  4. gamma is uncertainty aversion (objective = mean − gamma·spread): 0 = maximize expected return; higher = more conservative. Typical dashboard range 0–0.1.

Objectives and margin

  • objective="profit" needs a margin: the model's stored operating margin is used automatically; otherwise pass forward_margin or the API errors. Result Revenue/ROI columns are then on the profit basis.
  • group_bounds (joint % constraints over channel sets) forces the slsqp engine; a BINDING group's members legitimately sit off the global marginal.

Poll and read

  • run_optimizer returns 202 + a run_id ("opt_..."). Poll get_optimizer_results(model_hash, run_id=...) — the model-level form reflects only the LATEST run and a newer run overwrites it.
  • Column conventions must not be mixed in one summary:
    • Revenue/ROI = the solver's decision math (removal-lift counterfactual).
    • OptimizedEvalRevenue/ROI, HistoricalRevenue/ROI = fitted-convention comparison columns (match the Contributions panel).
    • ObjectiveMarginal (solver's equalized marginal) ≠ MroiAtOptimized (posterior mROI at the optimized spend, with 94% HDI) — they can differ by several times; quote the one matching the question.

Run history and curation

  • list_runs(artifact="optimizer"|"scenario", ...): pinned-first then newest-first. count is the PAGE length, not the total — page until a short page. The objective is NOT in summaries: fetch the run's inputs (profit runs carry objective: "profit"; revenue runs omit the key).
  • update_run renames/annotates (sets auto_named false permanently); set_run_pinned pins declaratively and idempotently. Scenario runs work the same via get_scenario_results(run_id=...).