Policy Engine

A policy engine is the rule-definition surface of a profit firewall: the place where the merchant's finance team writes down, in machine-readable form, what is and is not an acceptable transaction. Without a policy engine, margin governance reduces to either ad-hoc code (custom checkout scripts, brittle and developer-bound) or out-of-band controls (after-the-fact review, slow and incomplete). With a policy engine, finance owns a declarative policy registry (usually a UI-based set of rules but version-controlled in the background) and the enforcement layer reads from that registry on every order. In Agentis, those rules are evaluated on each order within 60 seconds of it being placed, and decide whether the order clears, is flagged, or is held before fulfillment. The architectural pattern is borrowed from network firewalls (rules-engine + match-action), authorization systems (Open Policy Agent / Cedar), and feature flags (LaunchDarkly): a clean separation between the policy authors (finance) and the enforcement runtime (the firewall). Policy types typically supported: (1) Margin-floor policies: minimum acceptable gross margin per category, channel, customer segment; can be expressed as percentages or dollar floors. (2) Discount-stacking policies: which combinations of automatic and code-based discounts may run simultaneously, with margin-aware overrides. (3) MAP-floor policies: per-SKU minimum advertised price floors with effective dates, sourced from vendor agreements. (4) Velocity policies: caps on how many orders may use a given promo before it is auto-disabled, regardless of stated limit. (5) Freight-zone policies: minimum margin requirements that vary by destination zone, accounting for actual freight cost variance. (6) FX-aware policies: margin floors that automatically adjust for current FX rates on cross-border orders. The financial-policy-engine variant adds audit-grade controls: an evidence record per policy evaluation (in Agentis, the Evidence Ledger), separation between the people who write policies and the system that applies them, and review steps for policy changes that affect a large share of orders. Mature policy engines support testing in 'shadow mode' where a new policy runs against live traffic without enforcing, generating an outcome ledger that can be reviewed before promotion to enforce mode; this is critical for confidence that a policy change does not over-restrict and hold orders that should have shipped. The policy engine and the profit firewall are inseparable: the engine without enforcement is a wishlist; enforcement without an engine is hardcoded business logic that cannot evolve.