Run & Deploy — Cash Recovery Engine

Requirements

• Node.js 18+ (tested on Node 24). No npm install — zero runtime dependencies.
• No network access required.

Run locally

cd delivery-package/cash-recovery-engine
node run.mjs        # writes reports/ + public/ledger.js
node verify.mjs     # writes verification-report.json/.md + evidence copy

Open public/tool.html directly in a browser (it loads public/ledger.js).

Deploy the web tool

The tool is a single static HTML file plus a generated ledger.js data file. Serve the public/ folder from any static host (or use the Forge publish pipeline, which copies it to published/cash-recovery-engine/tool.html). No server-side runtime is needed — the engine's scores are precomputed by node run.mjs and the optimizer re-runs client-side.

Bring your own AR data (closing the real-data seam)

The verification suite auto-detects a real dataset. Provide two CSV files:

data/official/history.csv — resolved invoices (to learn from)

One row per *closed* invoice, with the eventual outcome and whether a collector worked it:

invoice_id,customer_id,account_segment,amount_usd,days_past_due,aging_bucket,prior_late_ratio,avg_days_to_pay,disputed,size_vs_median,reminders_sent,credit_health,relationship_months,effort_hours,days_to_pay,worked,paid_within_horizon

• worked ∈ {0,1} — was this invoice actively worked by a collector?
• paid_within_horizon ∈ {0,1} — did it pay within the collection horizon?
• disputed ∈ {0,1}; credit_health ∈ [0,1]; ratios as decimals.

data/official/open.csv — live open invoices (to act on)

Same columns; worked/paid_within_horizon may be blank for currently-open items (they are only used for the control-AUC check on rows where worked=0).

Then:

node verify.mjs     # runs the identical checks on your data + an official block

Field mapping from common AR systems

Engine field	Typical AR/ERP source
amount_usd	invoice open balance
days_past_due	today − due date
prior_late_ratio	share of the customer's past invoices paid late
avg_days_to_pay	customer historical DPD-to-pay average
credit_health	normalized credit score / risk rating (0–1)
worked	dunning/collections activity log flag
paid_within_horizon	settled within N days of the snapshot

Important deployment note

This engine recommends *where to spend collector time*; it does not move money, post to the ledger, or contact customers. Treat its output as a decision-support worklist. Validate uplift on your own data with a holdout or A/B test before attributing recovered cash to it — see proof/LIMITATIONS.md.