Verify every rupee.
Escalate what can't be proven.
Deterministic matching first, AI only where rules genuinely can't resolve it, honest exceptions for the rest.
Show exact values
AI confidence: 0.60 · Fee tolerance: 3% · Date drift: 3 days
Advanced settings & bring your own data
Required columns — orders: order_id, amount, razorpay_payment_id, created_at. settlements: settlement_id, payment_id, utr, amount, settled_at (optional payment_ids column: a quoted, comma-separated list of payment IDs for one settlement row that covers more than one payment; optional is_instant column: true/1/yes for an on-demand Instant Settlement, which gets a wider fee tolerance since it carries a real, higher payout fee). bank_lines: line_id, narration, amount, value_date. refunds (optional): payment_id, amount (add refund_id for reliable outbound cash-position matching against a refund's bank debit line -- without it, matching falls back to a payment_id+amount key). chargebacks (optional): payment_id, amount, plus an optional fee column — a chargeback is a forced, customer-initiated reversal, not a voluntary refund, and is never netted against a settlement the way a refund can be; it's reported as its own exception with the reversed amount plus fee. No ground-truth accuracy is shown for your own data (no seeded truth label to score against) — match rate, amount at risk, and the exception list still are.
Accuracy across corruption levels
Runs fresh batches at increasing corruption rates (not the same batch shown above), several independent seeds per rate averaged together with the spread shown, so the accuracy number isn't just true for one cherry-picked run.
Confidence threshold calibration
The AI confidence threshold (default 0.6) is a starting point, not an empirically justified one. This sweeps the auto-accept bar across a range of values against one fixed batch and one fixed round of LLM verdicts — only the accept/reject line moves — so the default can be defended with a number instead of asserted.