Skip to content

How it works

Decisions and impact ranges

Recommendations that cite findings, carry a computed range, and name the check that will judge them.

The planner proposes candidate actions; code decides which survive. Each candidate must:

  • cite at least one finding (because);
  • request an impact scenario — recover_share, price_change, remove_excess_rate, retain_at_risk — whose low and high ends are computed in code from facts and bounded parameters (no number from the model);
  • name a check: a metric, a comparator, a threshold fact, and a date after the data.

Candidates are ranked by |impact midpoint| × confidence ÷ effort, and the top 2–5 are kept. A recommendation can also carry a not advised line — the tempting wrong move (“Don’t cut retail prices; retail is growing”).

From the example: Offer wholesale accounts a volume-tier discount to win back paused standing orders — expected impact 3,995 to 7,191 next quarter, from recovering 50–90% of the lost 8,500 net of a 6% discount; check revenue ≥ the March level by 30 September.

If the model can’t produce valid actions even after validation retries, the report ships with its findings and no recommendations, and says why.