Prefiler enforces strict mathematical and structural constraints. Rather than relying on soft, probabilistic models, every compliance finding is generated by auditable, rule-based algorithms.
Digital documents use layout-aware deterministic table parsing to extract raw values. Scanned pages or photo uploads activate a bounded vision/OCR transcription pass. All extracted raw values are initially flagged as unverified and carry exact page-coordinate provenance.
The core engine runs strict mathematical validation. Balance continuity verifies that computed running balances match statement-reported values exactly, enforcing a 0-paise absolute tolerance. Transactions must pass complete ledger-to-bank reconciliations.
Ruleset compliance checks and statutory computations are executed strictly via deterministic rules (surcharges, deductions, marginal reliefs). Non-authoritative AI models are isolated and only assist in clarifying gaps or drafting finding explanations.
The core Intelligence engine is continuously evaluated against a standard 20-case adversarial corpus (`canonical-evaluation.test.ts`). This corpus injects corrupted files, missing documents, overlapping transaction dates, mixed currencies, and transaction anomalies.
* **Evaluation Results**: Our standard test execution logs confirm 100% mathematical reproducibility and correct safe-failure triggering (raising correct `currency_mismatch` conflicts, blocking unreadable pages, and identifying sequence gaps) under synthetic equivalent data benchmarks.