This repository contains the Monte Carlo financial model behind savelocalbroadband.com — the Okanogan County, WA ratepayer campaign for public review of the PUD's pause on its rural wireless broadband network.
When the campaign says the ~$1.2M wireless upgrade is likely to pay for itself, that claim is the output of the code in this repository — published here in full so anyone can inspect it, challenge it, or re-run it.
python montecarlo_v2.pyThe only dependency is NumPy. The random seed is locked (20260629) and every configuration runs 200,000 iterations, so a re-run reproduces the published numbers exactly — there is no lucky run to fish for. The script prints each ledger's summary and rewrites the six output JSON files, which should come back byte-identical to the committed copies. Since v2.2 every output embeds the model version, seed, NumPy/Python versions, and the model file's SHA-256.
test_v2.py runs the guard suite: source-level guards on every audit correction, the runtime county = household + district identity, monotonicity checks (stress must hurt, upside must help), a 3-seed stability sweep, and an n-ladder convergence check. sensitivity_levers.py regenerates the published lever table (sensitivity_levers_output.json) from the live model.
| Ledger | Question |
|---|---|
| A · Cost recovery | Does the wireless line's own discounted cash earn the ~$1.2M back within the 15-year asset life (one wear-out refresh charged; the stricter before-wear-out test is published alongside)? |
| B · Ratepayer benefit | How much do households avoid overpaying because the cost-based public rate disciplines the market? |
| C · County value | Counting A and B honestly (the public rate cancels out): does the county come out ahead? |
| D · Scenarios | How do A and C look under three coherent worldviews — pessimistic, balanced, optimistic? |
| E · Build vs run-to-failure | How does building now compare against spending nothing and letting the end-of-life network decline? |
Results from different ledgers are never summed — they answer different payees' questions.
The model is documented at three depths — plain English, a decision-maker's briefing, and a full technical reference covering every distribution, formula, and audit correction:
- Live: savelocalbroadband.com/model-docs.php
- In this repo:
docs/model-documentation.html(a self-contained snapshot — download and open in any browser)
There is also a live calculator that runs a lightweight version of the model in your browser, with a slider for every major assumption.
| File(s) | Role |
|---|---|
montecarlo_v2.py |
The model (v2.2). Ledgers A–E plus the cost-of-waiting delay sweep; writes the six output JSONs. |
*_output.json, scenarios_output*.json |
The published results this code produces (version + seed + environment + SHA-256 embedded). |
test_v2.py |
The guard suite: audit-correction guards, ledger identity, monotonicity, seed stability, convergence. |
FINDINGS.md, REVIEW_v2.md, PROMOTION_BUNDLE.md |
The adversarial review records — the audit trail of corrections. |
montecarlo_robust.py, recovery_model.py |
Superseded earlier versions, kept public so the improvement is auditable. |
financial_appendix.py |
The deterministic DCF companion analysis. |
sensitivity_levers.py |
The one-lever-at-a-time sensitivity table. |
sources_and_uses.py / .json |
The funding-plan arithmetic. |
parse_all.py, *.csv |
Parsing of the District's billing/budget records (obtained by public-records request) into the datasets the model's inputs are anchored to. |
A tuned model's headline only ever improves. This one's history runs the other way: adding the labor and contingency a board member asked about dropped the headline from the 90s into the 70s; two adversarial audit passes flattened the growth assumption, discounted the recovery stream, and moved the second-largest reseller's exit to a timing that strands capital instead of saving it. The v2.2 release (2026-07-10) continued the pattern after a third-party review and an internal consistency audit: the sensitivity-table generator was found still pointed at the superseded model (every stale lever number had drifted in the favorable direction — all corrected downward), the year-2 exit stranding was extended to the county ledger, the macro-stress factor and overbuild tail were carried into every ledger, sales tax was extended to growth radios and the refresh, a fixed-vs-variable operating-cost bracket was published (−9 to −15 points at its ends), and the literal 15-year rate freeze — harsher than the model's own floor scenario — was published as its own sensitivity (~25%). Every markdown was kept and every markup is documented with its reason — the review records are in this repository, and the full correction log is §3.15 of the documentation.
Every input is tagged DOCUMENTED (from the District's own workbook, the Washington State Auditor's 2024 audit, or billing records) or INFERRED (a labeled modeling judgment, bounded wide, leaning against the campaign's conclusion). The two numbers only the District can publish — a firm tower-labor figure and the wireless-only cost split — are modeled as wide ranges and are the subject of an open records request.
If you find an error — in the code, the inputs, or the documentation — tell us. Corrections get made and logged, whichever direction they move the number.