Methodology: How Every Number Is Calculated
Lodestar makes one architectural promise above all others: the language model never does math. Every dollar figure you see is produced by deterministic code, from published data, under versioned rules. This page documents exactly how, because a tool that tells you what to do with money should be able to show its work.
The Four Layers
An analysis passes through four separate layers. Only the first and last involve AI at all, and neither of those is permitted to produce or change a number.
Parse — the only place a model reads your documents
An uploaded statement is converted into structured data: accounts, balances, transactions, rates, and holdings. A language model does this because statement layouts are endlessly varied and text extraction is genuinely what models are good at. Its output is a typed data structure validated against a schema — not prose, and not a conclusion.
Raw statements are processed and discarded rather than retained, and bank credentials are never stored at any point.
Compute — deterministic code, no AI involved
The structured data goes into the calculation engine: a financial snapshot (everything owned and owed), the metrics derived from it (savings rate, debt service, runway, allocation), spending analysis, net-worth projection, debt amortization, and Monte Carlo simulation.
This is ordinary code with test coverage. The same inputs always produce the same outputs, and any figure can be reproduced by hand or in a spreadsheet.
Diagnose — versioned rules produce findings and dollar values
A catalog of diagnostic rules is evaluated against the computed picture. Each rule that fires produces a finding and, where there is a move to make, a recommendation with a dollar impact calculated from your actual numbers — what an unclaimed employer match is worth, what a high-rate balance costs per year, what a savings-rate change does to a 30-year trajectory.
The ruleset is versioned, and every analysis records which version produced it, so a past result can always be explained by the rules in force at the time.
Explain — the model phrases and orders, nothing more
The ranked list of findings and recommendations, with their computed dollar values already attached, is handed to a language model. Its job is to write the explanation: plain-English phrasing, sensible ordering within equal-impact groups, and the connective tissue that makes a list of rule outputs read like a plan.
It cannot introduce a figure, change a figure, or add a recommendation that the rules did not produce. Model, prompt, ruleset, and calculation-engine versions are all stamped onto the result.
Where the Data Comes From
Net-worth benchmarks. The Federal Reserve Survey of Consumer Finances, dataset version scf-2022.1. Medians are the Fed's published figures; percentile cutoffs are SCF-derived and rounded, which is why they are always shown with a tilde. You can check your own position on the net worth by age page.
Tax figures. Published IRS brackets, contribution limits, and phase-out thresholds for the applicable year. Tax output is educational and always directs you to confirm with a CPA — Lodestar does not provide tax advice.
Market assumptions. Yours to set. The tools ship with conventional defaults and deliberately present a range of outcomes rather than a single projected number, because a single average return hides the sequence risk that actually determines whether a plan survives a bad decade.
What Lodestar Will Not Do
Lodestar Wealth is not an investment adviser and is not registered as one. It does not act as a fiduciary and does not recommend specific securities to buy or sell. Recommendations stay at the level of asset classes, account types, interest rates, and amounts — which is where the large, quantifiable wins usually are.
It also does not let commercial relationships touch the ranking. Affiliate payout is never an input to the engine: recommendations are computed and ordered by dollar impact before any product is matched to them. A move with no affiliate relationship ranks exactly as high as one that has one.
For a fuller treatment of how this category is regulated and what the “AI financial advisor” label actually means, see AI financial advisor apps: what they do and what they can't.
Check It Yourself
The free calculators run the same engine as the full analysis, with no account required. If you want to verify that the projection math is what this page claims it is, put your own numbers into the net worth calculator or the debt payoff calculator and reproduce the result in a spreadsheet. It will match.
Questions People Ask
Does Lodestar use AI to calculate financial figures?
No. Every dollar figure comes from a deterministic calculation engine written as ordinary, test-covered code. The language model has exactly two jobs: converting uploaded statements into structured data, and ranking and phrasing recommendations that were computed before it was involved. It is handed final numbers and has no ability to alter them.
Why does that distinction matter?
Language models fail silently at arithmetic. Asked to compute the interest cost of a balance, a model will return a confident, specific, wrong figure with no signal that anything went wrong. For a tool whose entire output is dollar amounts you plan around, that failure mode is disqualifying — so the architecture removes the possibility rather than trying to catch it after the fact.
What data sources does Lodestar use?
Published, citable datasets only. Net-worth percentiles come from the Federal Reserve Survey of Consumer Finances (dataset version scf-2022.1). Tax brackets, contribution limits, and phase-out thresholds come from published IRS figures for the applicable year. Market return and volatility assumptions in projections are yours to set — the tools ship with conventional defaults and show a range rather than a single number.
How are recommendations ranked?
By computed dollar impact. The diagnostic engine evaluates a versioned catalog of rules against your financial picture, each rule producing a finding and, where applicable, a recommended move with a calculated dollar value. Those values determine the order. The language model receives the ranked list and writes the explanation — it does not reorder it.
Does affiliate revenue affect what Lodestar recommends?
No. Affiliate payout is never an input to the ranking engine. Recommendations are computed and ordered by dollar impact before any product is matched to them, and a move that carries no affiliate relationship ranks exactly as high as one that does. Where a product is mentioned and a referral relationship exists, it is disclosed.
Can I reproduce the numbers myself?
Yes, and that is the point of publishing the method. The projections use standard monthly compounding with end-of-period contributions; debt payoff uses standard amortization; the Monte Carlo bands replay the same plan through 500 randomized return sequences and shade the 10th-to-90th percentile spread. The free calculators expose the same engine, so you can check any figure against your own spreadsheet.
Is any of this investment advice?
No. Lodestar Wealth is not an investment adviser, does not act as a fiduciary, and does not recommend specific securities. It provides education and tools. Recommendations stay at the level of asset classes, account types, rates, and amounts. Anything with tax consequences points you to a CPA.