How ARPI keeps its numbers honest

Volume alone never ranks a person. A high-volume salesperson may show weak gross retention, poor follow-up, heavy discounting, or simply favourable lead routing. Any view that displays employee results must state how it shows lead quality, store traffic, tenure and vehicle mix alongside the number.

A lead is credited to the source that produced it, once. Multi-touch journeys are not modelled, so a campaign that assisted a sale without originating the lead receives no credit. This is a deliberate simplification of a genuinely hard problem, and marketing return figures must be read with it in mind.

The synthetic generators contain correlations because a plausible dataset needs them. Those correlations were written by the generator, so no finding drawn from this data may claim that one thing causes another - in the data or in the industry. This is why the project draws no conclusion about automotive retail at all.

A versioned YAML rule file evaluated against the published datasets. No language model, no learned model, no scoring heuristic: every word comes from a rule template and every number from an exported column, so any action can be recomputed by hand. Evaluated once at export time, so the queue is identical for every visitor of a dataset version, and stateless: nothing is assigned, acknowledged or completed.

Editing a review threshold changes the queue even though no business fact moved. The manifest therefore records the ruleset's hash, and the export check re-derives the queue and fails if the committed one differs. Every threshold the rule file owns is a project default for a fictional dealer group, never an industry benchmark, an OEM standard or a compliance requirement.

F&I penetration, customer retention, service-to-sales opportunity and target attainment all depend on facts that have not been built. No result requiring a deferred fact may be published, which rules those four subjects out of any finding until the facts exist.