Three Looks, Then Settlement: Deadline-Bounded Escalation of Uncertain UPI Payments under Rationed Status Checks
Year: 2026
Type: preprint
Abstract: In April 2025 UPI had an outage of several hours that press reports, citing NPCI, attributed to excessive status-check traffic. NPCI responded by rationing status checks on uncertain payments. Each payment may now be checked at most three times, the first check no earlier than 90 s after the payment (45–60 s once revised response timers apply). What is still uncertain after that is settled through settlement files, or, for a transaction reported as not found, by one check through NPCI's dispute system (UDIR). Our implementation uses the settlement files and does not implement the UDIR route. We ask what a participant (a payment service provider, gateway, merchant or lender) should do within this ration. Our answer is an uncertain-outcome coordinator. It is event-sourced and treats UNKNOWN as an explicit state. After 24 hours an unresolved payment is escalated to an operator; the deadline bounds escalation, not resolution. Money in limbo is recorded as a ledger hold that is never voided without evidence; a second debit is blocked while the first outcome is unknown; intraday settlement files are ingested as they arrive. On top of these known parts we add shared-fate deferral. When a bank stalls, the coordinator holds back the rationed checks of every payment waiting on that bank, probes it with a single canary, and releases the held checks once the bank answers again. Two TLA+ models are checked with TLC. One covers a single order in depth: no double posting, no wrong final state, no double charge, a per-payment escalation deadline, holds released exactly once. The other covers several payments on several banks, with deferral, canaries and release. Eleven planted defects, counting a disabled duplicate guard, are each caught. We evaluate in a deterministic simulator calibrated to published bank-wise decline rates and the NPCI rules. The pre-registered main grid has 14 policies: the full coordinator, 4 baselines, 7 ablations that each remove one component, and 2 add-on variants. It runs them on 7 workloads with 30 seeds each, 2,940 runs. Running a subset of the policies on 7 further workloads adds 1,260 runs, for 4,200. With the post-hoc experiments, including two fairer baselines added after review, there are 8,536 runs in all and 0 invariant violations. Against common industry practice (our baseline B3), the coordinator reduces double-charged orders by 60–98% in the customer-facing workloads. During a bank stall it brings the 95th-percentile time to the first real evidence of a payment's outcome (P95 time-to-truth; guesses do not count) down from about 19 hours to about 63 minutes. These headline figures are against B3 only, not against the strongest baseline. Fairer baselines show where the gains come from. Most of the tail-latency gain comes from ingesting intraday settlement files: B3 plus those files reaches a P95 of 103 minutes in a stall. Most of the double-charge reduction comes from never guessing failure plus the duplicate guard: the best published schedule plus our guard comes within 2 double charges per 10k orders of the coordinator, and beats it after crashes. Against that strongest baseline (B4 plus the guard), the coordinator's own mechanics still lower P95 by 12–35 minutes in six of seven workloads. In the bank-stall and retry-storm workloads, and only there, they also add 1.2 pp of fulfilment and use 27–36% fewer status calls than B4 plus the guard. Shared-fate deferral on its own is worth 3–13 minutes of P95 when an outage ends inside the two-hour check window and nothing otherwise. We also report costs and negative results. The coordinator makes more status calls than a baseline that throttles itself into silence and then guesses. The duplicate guard costs 0.02–1.5 pp of fulfilment; a "warn-only" variant recovers up to 1.35 pp of it at up to 1.25 extra refunded double charges per 10k orders. Two schedule variants (a learned schedule after crashes, and a per-bank bandit) did not beat simpler ones. Our Go implementation on Postgres survives repeated kill -9 of the leader with zero API errors. With group commit it accepts about 3,100 payments/s on a 2-vCPU machine, and it completed an end-to-end test-mode payment against a commercial gateway. These are measurements on a single test machine, not production figures. Every other result comes from simulation under stated assumptions. No real payment traffic was used. The most important next step is validation with a real participant's data. Code, data and reproduction scripts: https://github.com/sujan-mitta/threelooks
Author: Mitta, Venkata Sai Sujan
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235800768
