Repository navigation
Releases
Releases · REQUIMDA/Suspense-Revenue-Recovery-Agent
v1.0.0 — Track 3 submission
A revenue recovery agent that knows when recovering is unsafe.
The LLM diagnoses and explains; a deterministic, version-controlled policy table decides the money action.
Results
Evaluated across 30 independent seeds / 3,600 simulated cases against a naive always-retry baseline on identical cases with identical hidden ground truth:
| Naive | Suspense | |
|---|---|---|
| Revenue recovered (mean) | ₹5,638.85 | ₹6,137.67 |
| Lost to duplicate charges | ₹118.36 | ₹0.00 |
| Net position | ₹5,520.49 | ₹6,137.67 |
Mean advantage ₹617.18, 95% CI [₹338.50, ₹895.87] — entirely above zero. Suspense won 25 of 30 runs, and the 5 it lost are reported rather than hidden.
What's in it
- Bucket A — exposure analysis across all 97,277 real Olist orders
- Bucket B — controlled naive-vs-Suspense policy evaluation with a multi-seed robustness sweep
- Bucket C — live Razorpay test-mode execution (real Payment Links)
- Replayable JSONL audit trail, per case per node
- 330 tests, CI green on Python 3.12 and 3.13, ruff clean
Metrics are never blended across buckets — the rule is enforced by the metric key names themselves (real_ vs sim_).