ACRE (Adaptive Cascade Reconciliation Engine)
ACREReconPe's proprietary 6-stage probabilistic matching pipeline that profiles data, cascades through blocking levels, scores candidates with Bayesian confidence, and enforces domain invariants.
ACRE is the matching engine behind ReconPe. Rather than running simple exact-key lookup like traditional rule-based reconciliation tools, ACRE runs a 6-stage pipeline: profile, block, score, assign, validate, and classify.
The blocking stage is where ACRE's 'cascade' comes in: candidate pairs are generated through three successive layers — exact key match, relaxed fuzzy match, and LSH nearest-neighbour search. This catches matches even when order IDs have typos, date formats differ, or amounts are off by rounding.
The scoring stage applies Fellegi-Sunter probabilistic weights to every candidate pair, producing a 0–100 confidence score with field-level agreement metadata. Scores above a tuned threshold are auto-approved; scores in the grey zone are surfaced for human review.
Because ACRE is deterministic (no LLM calls, no inference), it produces the same output for the same input — which is what auditors and finance controllers require. The AI layer sits downstream, consuming ACRE's structured output to explain exceptions and suggest resolutions.