Every month, subscription payments fail.
Recovery Loop evaluates retry timing against a documented day-one, day-two, day-three baseline. The UPI retry allowance applies to each mandate execution sequence, not its entire lifetime. Razorpay also offers intelligent retries, which this experiment does not benchmark.
The problem
Recovery Loop investigates whether failure reasons and issuer health can improve a fixed-calendar retry baseline. Public documentation does not establish every signal Razorpay uses internally.
Razorpay's Subscriptions retry model lists T+1, T+2 and T+3 for UPI and cards; the subscription enters halted if all fail. This is our evaluation baseline, not a description of every Razorpay product. Razorpay also advertises intelligent UPI retries, which we have not benchmarked.
The NPCI guideline limits UPI AutoPay to one original attempt plus three retries per mandate execution sequence. It is not a lifetime limit across every recurring debit. Our evaluation models one failed debit per mandate with its own remaining allowance; saved retries are not pooled across customers. Provider-specific execution windows also apply.
Razorpay's Priority-based Routing documentation describes temporary gateway downtimes lasting twenty minutes when success rates drop, with traffic routed to another gateway. This is a downtime duration, not a response-time guarantee. Its Recurring Payments warning says Optimizer rules apply to registration, not subsequent debits.
The finding
How long the agent may defer before it is forced to act. Drag it. Every number below is measured on 10 held-out seeds the cap was never chosen on.
At 14 days, deferral first spans a salary cycle. The agent can wait for the credit that actually funds the payment — this is the frozen, validated cap.
Grey figures compare against Razorpay's fixed T+1 / T+2 / T+3 ladder.
Counts are means across 10 seeds, so they carry a decimal.
Grey figures compare against Razorpay's fixed T+1 / T+2 / T+3 ladder.
Counts are means across 10 seeds, so they carry a decimal.
Reported as inconclusive — the seed range crosses zero.
How it was measured
Each was found, documented, and corrected. The numbers before and after are in the repository.
Success was defined as a random draw falling below the model's own prediction — the agent graded its own homework. Rebuilt so outcomes come from hidden world state the model never sees. The advantage collapsed from 2.5× to nothing.
The simulator emitted an old field shape after the diagnosis layer changed. Every policy silently recovered ₹0. A generator assertion now fails loudly instead.
Deferred retries were scheduled with a valid time and then thrown away by a one-line filter. Recovery was being deleted, not measured.
The budget was modelled as one shared pool. NPCI's cap is per mandate and non-transferable — a category error that inverted the result.
Real NPCI baselines were fed to a gate designed for outage spikes, so it read every bank as permanently down and made 0 attempts. A policy that never acts has not been measured.
Economic refusals were treated as terminal, so 98% of mandates were abandoned at t=0 and never revisited. Separating hard stops from deferrals is what exposed the deferral cap itself.
Honest limits
The console runs the same policy on a live decision stream. Take a bank down and watch pending retries turn into deferrals.