Categorisation Rules

Categorisation rules automate the classification of bank import transactions. When a transaction's description, reference, amount, or contact matches a rule, the system suggests (or auto-assigns) the target ledger account.

Rules operate at three provenance levels, evaluated in strict priority order:

  1. Local rules (ρ=3) — created by the user for their organisation. Highest priority.
  2. Verified rules (ρ=2) — seed rules created during organisation setup. Mid-priority.
  3. Global rules (ρ=1) — cross-tenant consensus rules from the Global Rules Repository (GRR). Lowest priority.

When multiple rules match, the winner is determined by a 7-dimension lexicographic lattice score: provenance, operator strength, field strength, pattern specificity, value fit, user-assigned priority, and creation order. This guarantees deterministic, explainable outcomes.

Each categorisation decision produces an evidence object explaining why the outcome was reached, including the score vector, matched field, and social proof from global rules. Decisions and corrections are logged to an audit trail for the feedback loop (TMADD Phase 4b).

Rules support seven match operators: contains, equals, starts_with, ends_with, greater_than, less_than, and regex. Fields include description, contact_name, amount, reference, and metadata keys.

→ The Categorisation Rules object

Endpoints