v4.5.0 9 February 2026 Feature

Lexicographic Lattice Scoring

Lattice-Scored Rule Matching

The categorisation engine now uses a formal lexicographic lattice to select the best matching rule, replacing the previous “highest priority wins” approach. When multiple rules match a transaction, the engine scores each across six dimensions and selects the mathematically strongest match.

Score Dimensions

Each matching rule is scored as a 6-dimension vector, compared left to right:

  1. Operator specificity — equals (5) beats starts_with (4) beats contains (3). Regex is scored dynamically based on literal character density.
  2. Field strength — reference and contact_name (4) outrank description (2) and amount (1). More informative fields win.
  3. Pattern specificity — longer match values are more specific. A rule matching “BRITISH GAS” (11 chars) outranks one matching “GAS” (3 chars).
  4. Value fit — how much of the input the rule covers. An equals match scores 100%; a short contains on a long description scores proportionally less.
  5. Priority — user-assigned priority (0–999), as before.
  6. Tie-break — earliest creation date, then lowest ID. Guarantees determinism even when all other dimensions are equal.

What This Means

  • Specific rules win automatically — an equals "TESCO STORES" rule beats a contains "TESCO" rule at the same priority, without manual priority tuning.
  • Regex penalty — lazy patterns like .* are penalised. A regex TESCO.*LTD with 8 literal characters scores well; a regex .* scores near zero and loses to any literal match.
  • Visible scoring — the test panel now displays the score vector with labelled dimensions, so you can see exactly why one rule won over another.
  • Backward compatible — existing rules and priorities continue to work. The lattice adds specificity resolution on top of priority, it doesn’t replace it.

Evidence Improvements

  • The test panel now shows the score vector: [3, 2, 12, 100, 10, -8] with dimension labels (op · field · pattern · fit · priority · tie).
  • When multiple rules match, the evidence explains which dimension decided the winner (e.g. “won by operator specificity”).
  • Ambiguity detection now compares full score vectors rather than just priority.