Outcome Contract — Cash Recovery Engine

Request

Finance and revenue-operations teams sit on large balances of overdue receivables and a fixed amount of collector time. The conventional worklist — chase the oldest invoice, or the biggest balance — wastes that time on accounts that would have paid anyway and on accounts that will not pay regardless of how many calls they get.

Build a decision engine that, given an accounts-receivable ledger and a weekly collector-hours budget, produces a prioritized collections worklist that maximizes the cash actually recovered per collector-hour, and quantifies the business impact (cash accelerated, days-sales-outstanding effect).

Scope & honest constraint

No real company accounts-receivable ledger is present in this workspace, and none can be retrieved from here. The contracted, verifiable outcome is therefore a complete, runnable receivables-prioritization engine plus a documented seam for real data — not a recovery number from a specific company's books.

What is delivered LIVE and verified:

• An uplift (causal) model that separates invoices that will *self-cure*

from invoices a collector touch will actually convert (the "moveable middle").

• A capacity-constrained optimizer that fills a collector-hours budget with

the highest expected-cash-per-hour invoices (0/1 knapsack, greedy-by-ratio).

• A risk segmentation (Persuadable / Self-Cure / At-Risk / Critical) with a

recommended action per segment.

• Downloadable deliverables: a ranked worklist CSV, a structured JSON

result, and a one-page executive summary.

• A live web interface that re-optimizes the worklist as the capacity dial

moves and charts recovered cash vs. collector-hours against the conventional baselines.

• End-to-end verification on a synthetic AR benchmark with a known causal

ground truth, plus an auto-detecting real-data evaluation path.

Success criteria (MUST PASS)

1. node run.mjs trains the engine and writes a ranked

reports/collections-worklist.csv, reports/engine-output.json, and reports/executive-summary.md.

1. node verify.mjs passes 100% of its checks, asserting:

• the logistic learner recovers a known separable signal;
• synthetic potential outcomes are monotone (working never reduces payment);
• the optimizer never exceeds the hours budget;
• self-cure propensity ranking (AUC) and calibration (ECE) clear thresholds;
• the engine beats the strongest simple baseline on cash recovered by a

clear margin, and recovers at least 2× the cash of FIFO at equal hours;

• results are reproducible (seeded);
• the worklist schema is complete and budget-feasible.

1. Deterministic, dependency-free Node (no network, no native deps).

Out of scope

• Real cash collection, payment processing, or accounting-system integration.
• A claim of recovered cash for any specific company (a disclosed seam pending a

real ledger and a holdout/A-B validation).

Acceptance

Accepted when both MUST_PASS commands succeed on a clean checkout, the proof package answers the ten IRS_AUDITOR questions, and the delivery package documents exactly how to swap in a real AR export to produce a live worklist.