Three operational answers, one model.

Use 1Sizing a bonus pool

A primary-care clinic with $5M/month in realized revenue wants to allocate 5% to a physician/staff bonus pool. Over the regenerated 60-day evaluation window, the perfect-foresight pool would be 10,000,000 × 0.05 = 500,000. The operational question: how much of that can be committed safely before remits arrive?

Of every dollar billed, only ~40% lands as cash within the bonus window — Texas-Medicaid's 0.52× index, payer mix dragged toward self-pay, and adjudication lag together strip 60¢ off the top. A bonus pool sized against billed revenue commits the clinic to an outflow it has not yet collected. Cost Predictor replaces billed with p10 realized.

$5M-monthly clinic — 60-day window, 5% bonus pool

Approach	Median under-allocation	Over-allocation risk	Row breach p10
Perfect foresight · 5% of actual realized	—	0.0%	—
Approach A · 5% of p10 realized	~$433,684 left vs perfect	0.0%	0.090
Approach B · 5% of p10 realized (by-payer)	~$462,848 left vs perfect	0.0%	0.020

Reading: under Approach B by-payer, the clinic leaves ~$462,848 on the table per 60-day evaluation window (vs a counterfactual perfect-foresight pool of $500,000) — and in zero origin/horizon windows did the p10-sized pool exceed the actual 5% of realized. That second number is the point: the clinic structurally cannot over-allocate at this sizing.

The trade-off is real — under-payment of bonuses is recoverable (a true-up at quarter-end), over-payment is not. Cost Predictor's design preference is conservative-first.

Use 2Cash-flow planning

Below: cumulative realization over 90 days for each payer, using the LogNormal lag parameters from config/adjudication_params.yaml. Curves are the survival CDF F(t; μ, σ) — the fraction of paid dollars that have arrived by day t.

The operational read: at day 30, Medicare FFS has paid ~95% of its committed dollars; Medicaid MCO is at ~38%; self-pay is at ~7%. A clinic projecting 30-day cash from a $5M billed month will see materially different cash on hand depending on which payer those dollars came from — which is the whole reason the model is decomposed by payer rather than pooled.

Use 3Payer-mix sensitivity

Three planning scenarios on a $5M-monthly clinic. Numbers are point-estimate expected realized (deterministic denial × E[paid-claim realization] per payer; equivalent to the mean of the demo's Monte-Carlo trials at 5,000 N).

Expected realized, three payer-mix scenarios

Mix scenario	Medicaid MCO	Commercial	Self-pay	E[realized]	Realization rate
HRSA UDS Texas (default)	25%	27%	28%	$1.96M	39.2%
+5pp commercial, −5pp self-pay	25%	32%	23%	$2.16M	43.2%
+5pp Medicaid MCO, −5pp commercial	30%	22%	28%	$1.84M	36.8%
−10pp self-pay, split commercial / Medicaid MCO	30%	32%	18%	$2.27M	45.4%

Two business reads stand out. First, dropping self-pay 10pp (split into Medicaid MCO and commercial) lifts realization ~6pp — bigger than any plausible operational efficiency gain. Second, the direction of a +5pp Medicaid MCO move depends on what it displaces; against commercial, it's a net drag, against self-pay it's neutral-to-positive on realization but speeds up cash (lag mean 35d vs 60d+ for self-pay).

Try this in the live demo — drag a slider, watch the headline numbers and the per-payer bar chart shift in real time.

Open the live demo

§Limitations

From §5.2 of the system-design document, in the model authors' own words. Don't oversell.

• Synthetic data only. Population is Synthea v4.0.0 with custom Texas Medicaid MCO encoding. Demographics are seeded to Houston, but Synthea's eligibility-driven mix under-weights Medicaid/uninsured vs HRSA UDS Texas FQHC mix. Demographic re-weighting is deferred.
• Single-remit only. Initial pay → recoupment → secondary ladders are not modeled. Multi-remit support waits on real 835 ERA data, where the structural change can be made against a measured ground truth.
• Texas-specific priors. Lag and paid-claim realization priors are seeded from Texas TDI prompt-pay statutes, KFF Texas fee index, and Premier 2024 hospital denial rates. Deploying outside Texas requires re-seeding the YAML — not a model change, but a calibration step.
• Public-source-bounded accuracy ceiling. Per-payer lag and realization-rate priors are priors, not measurements. Until a partner-clinic 835 feed lands, the accuracy claim is bounded to "directionally useful for cash-flow risk management against the configured payer mix; not a substitute for clinic-specific historical adjudication data."
• Phase 2 (denial classifier) not implemented. The Bernoulli denial draw inside the convolver substitutes for the documented XGBoost / LightGBM classifier. Acceptable while terminal-denial rates are bracketed at 4–8% and dwarfed by lag and realization-rate uncertainty in pooled forecast variance.