About Bracketsight

What this site is

Five tools, each answering a decision rather than an arithmetic question. A payment calculator tells you what one plan costs. These tell you which option costs least, which of them cannot be undone, and where the rule that decides it is written down.

• Student loans — Which of the 9 federal repayment plans costs you least over 30 years?
• Paycheck — Which OBBBA deductions is your pay owed — and does your W-2 show them?
• Health cover — How close is your household to the 400% subsidy cliff?
• Property tax — Is your home over-assessed enough to be worth appealing?
• Trades — What should this job cost, and what must the contract say in your state?

How a number gets onto this site

A rate, threshold, bracket, deadline or statutory term is never written into the code that uses it. It goes into a versioned rules file carrying the date it takes effect, the date it stops, and at least one citation with a URL and the date somebody last opened that URL and read the value off it. The engine reads the file; the page reads the engine. When a rule changes, one file changes and every page that depends on it recomputes — which is why a correction can be published the same day it is confirmed rather than hunted through prose.

The arithmetic itself is plain TypeScript with no dependencies and no network access. Money is held in integer cents and rates in basis points, because 360 iterations of floating-point drift produces a wrong answer at the end of a 30-year simulation. The five engines carry 399 tests between them and the app another 69. Where an agency has published worked examples — the eight for the Repayment Assistance Plan, for instance — those examples are encoded as tests, so the engine has to reproduce the government’s own arithmetic before it ships.

Where AI is used, and where it is not

No language model touches a calculation, an eligibility test, a ranking or a clause selection. That is enforced in the build, not promised in prose: the engine packages have zero dependencies, so there is nothing for an AI call to be made through. The planned uses for AI are reading an uploaded document to fill a form in, and explaining a result in plain language — and in the second case every numeral and date in the generated text is checked against the engine’s own output before it renders, with a deterministic template shown instead on any mismatch. Neither feature is live today.

What is not verified

This is a pre-launch build and it says so on the pages it affects. 55 items across the five engines are recorded as unverified, grouped by what would unblock each one. No gap is filled with an estimate, a remembered figure or a plausible substitute: where a value could not be read off a primary source, the engine keeps what it had, the gap is recorded, and the page carrying it warns the reader. A gap closes only against a primary source — a secondary source may cross-check a value but may never close it.

Two of the tools refuse to answer rather than answer badly. Contract generation is blocked for any state whose statutory notice text has not been transcribed word for word, because a paraphrased statutory notice is not a notice. And a New Jersey assessment verdict is withheld when the Director’s Ratio that governs the county is unavailable, because the ratio is what decides whether the appeal has a case at all. In both places the tool says what is missing instead of producing something that looks like an answer.

No credentialed reviewer has signed off yet

The editorial policy for each section names the review it requires before launch — an enrolled agent or CPA for the tax and subsidy engines, a construction attorney for each state’s contract clause language, working contractors for the pricing data. None of those reviews has been completed. Until they are, every affected page carries the warning and the numbers are illustrative. Naming this here rather than only in a repository file is the point: a reader deciding whether to trust a figure should not have to go looking for the caveat.

Who is behind it

The verification process that stands behind every figure — what a primary source is allowed to be, what a pass over 315 individual values found, and which professional reviews have not yet happened — is set out on who writes and checks this.

Bracketsight is independent. It is not a lender, a servicer, an insurer, a marketplace, a payroll provider, a county assessor or a contractor, and none of those pays for placement, ranking or mention anywhere on it. No tool recommends a named commercial product, and no result changes based on who is reading it.

How it is paid for

The tools are free and require no account. Planned revenue is advertising, an unbranded document upgrade in the trades tool, and clearly labelled software affiliations. None of those can reach a computed figure: advertising is served by a network that receives no input from the engines, and there is no code path by which a sponsor could alter a ranking, a clause list or a recommendation. No advertising runs on the site today.

What happens to what you type

Nothing you enter is sent to a server, because there is no server to send it to. There is no account, no signup wall and no database. Work is saved to your own browser’s storage, and a shared scenario link carries its numbers in the URL fragment, which browsers never transmit. The full detail is in the privacy notice.

How to get something corrected

Write to info@bracketsight.com. A report is checked against the primary source before anything changes; if it holds, the rule file is updated, the pages recompute, and the change is recorded in that tool’s changelog with the date and the source. Corrections are never taken on a secondary source, however confident it sounds. What to include, and what this address cannot do, is on the contact page.