Synthetic demo data · Read-only · Run the full agent with approvals locally: npx closeloop-mcp

Month-end close · 1 Oct 2026 – 31 Oct 2026

Gateway settlements reconciled against the bank statement and the ledger as of 31 Oct 2026. Every number on this page was computed by the deterministic engine; nothing here was typed by hand.

Results · held-out evaluation

8 seeds × 25,000 payments, unseen noise variants included · full report (EVAL.md)
Credits linked to the right settlements
90.5%
n=1467 · 1 wrong
Batch credits decomposed correctly
76.2%
23.8% sent to a human
Corrupted UTRs recovered
100.0%
incl. correct suggestions
Exception recall / precision
91.8% / 94.1%
macro, 13 types

Gateway report vs amount-only

The gateway's settlement recon report says which settlements each bank transfer paid, so a batched credit is matched from its item-level rows instead of an amount-only subset search. Amounts must still agree to the paisa.

Synthetic report
Amount-onlyWith gateway report
Demo data · open items as of 31 Oct 2026
Batched credits matched from the report's rows014
Batched credits matched on amounts alone111
Ambiguous batches waiting for a human10
Match conflicts waiting for a human30
Settlements parked on a batch decision100
Report ↔ bank mismatches raised—0
Held-out eval · 8 seeds · 85 batched credits
Batched credits matched to exactly the right settlements40.0%100.0%
… sent to a human as ambiguous320
… blocked by a match conflict180
Wrong links (all bank credits)10
Human review items on bank credits15888

4,864 report rows covering 293 settlements paid under 227 bank transfer UTRs. The report is generated from the same synthetic data, so it never disagrees with the bank here; disagreements raise a mismatch exception instead of a match.

Bank credits matched, by kind

Exact UTR
135
UTR found after normalising case, spaces, hyphens
Fuzzy UTR
7
Edit distance ≤ 1, exact amount and date, clear margin
Contiguous batch
8
One credit = a run of consecutive settlements
Arbitrary batch
2
Proven-unique subset of settlements

67 settlements sit inside batch credits · 3,406 payments settled · 164 ledger receipts linked (9 by amount + date + party) · 180 payments still inside their T+n window

Exposure by class

Exposure classes measure different things and are never summed into one figure.

Amount at risk
₹8,90,846.06
21 exceptions · Money that may be lost or misstated until resolved
Awaiting settlement
₹3,55,344.46
13 exceptions · Money still owed by the gateway or arriving late
Structural
₹12,87,904.74
17 exceptions · Books or links that need a human decision; not money at risk on its own

Bank vs books tie-out

Σ bank credits − Σ ledger receipts must equal the reconciling items the engine identified.

Bank credits
₹1,72,66,222.50
Ledger receipts
₹1,70,67,102.88
Difference
₹1,99,119.62
Explained by reconciling items
₹1,99,119.62
Unexplained residual
₹0.00Ties out

Largest exceptions

51 exceptions in Oct 2026

Open queue →

AI: LLM explanations were requested for the DEMO.md cases and the top 20 exceptions by amount; all other explanations are deterministic templates.