Editorial policy

What decides a number here, and what only explains it?

A dependency-free TypeScript engine decides; prose explains. Money is held in integer cents and rates in basis points, so no floating-point drift can reach a result, and the calculation path has no AI, no network access and no third-party packages in it. Content pages are written against the same versioned rules files the engine reads, which is what stops the prose and the arithmetic from drifting apart. If a sentence on this site states a figure, the figure came out of the engine or out of a cited rules file — never out of a language model.

How does a rule get onto this site?

Every rule is encoded from a primary source — statute, regulation, IRS guidance, or an HHS publication — never from a blog or a competitor's calculator. Each rules file carries its citations and a last-verified date, published on /sources. A rule we cannot cite does not ship.

Rules are also dated rather than constant. Each file records an effective range and a ruleset version, so a change in the law is a new dated file rather than an edit to code. That is not a stylistic preference: when Pub. L. 119-21 §71305 struck the advance-credit repayment cap for tax years after 31 December 2025, the fix was to record in data that no limitation is in effect — and the reconciliation code now refuses to run at all if a future ruleset claims a cap exists but ships no bands, rather than reaching an uncapped answer by accident.

What closes a verification gap?

A primary source, fetched and read — nothing else. A secondary source may cross-check a value but may never close a gap, and any row resting on one is labelled as secondary until the authoritative document is read. Where a source could not be reached, the value is left alone and the gap is recorded rather than quietly improved: no row in the register carries a guessed figure or a plausible-looking substitute. Closing a gap means editing the code and the register entry together; deleting a row without a fetched URL and a verification date does not count.

One episode explains why the bar is set there. While the applicable-percentage table was being checked, a summarising fetch layer misreported Rev. Proc. 2025-25 as “2025-21” and returned a completely fabricated percentage table. The engine's table was already correct, and trusting that summary would have turned a right answer into garbage. Every figure is therefore read from extracted PDF or statute text, never from a summary of one.

Who has reviewed the tax logic?

Nobody yet, and that is a launch blocker rather than a detail. Bracketsight launches only with a named enrolled agent or CPA reviewing the subsidy and tax logic and the content built on it. No such reviewer is engaged for this section today. A verification pass checked whether encoded values match their cited sources, which is a much narrower question than whether the output is sound tax advice — and nothing published here substitutes for that review. When it happens, this page will carry the reviewer's name, credential and review dates. Until then the site says it is a pre-launch build, everywhere it matters.

How are corrections handled?

Errors are corrected within 48 hours of confirmation, logged on the changelog with what changed and why, and affected pages carry updated verification dates. Report an error to info@bracketsight.com; a wrong threshold on a cliff page can cost a household its entire credit, so corrections outrank features. A correction that changes a number also re-derives the affected tests by hand, from the authority, rather than snapshotting whatever the engine now prints.

How is the site funded, and can that reach a number?

Bracketsight plans to earn from clearly separated display advertising and an optional paid planning PDF — never from insurance commissions. We do not sell or recommend specific insurance plans, we are not a brokerage, and no revenue source can influence a computed result: the engine has no knowledge of monetization, takes no input from it, and would produce the same figure with every ad slot removed. Any future enrollment-partner affiliation will be disclosed on the page where it appears, at equal visual weight.

How often is each rule re-checked?

On the schedule the publishers actually keep. HHS poverty guidelines arrive in mid-January and drive the following coverage year; HSA limits around May; the §36B applicable-percentage table in mid-summer; retirement and IRA cost-of-living figures around October; the CMS marketplace public use files for the next plan year around October and November; Medicaid expansion status quarterly. Legislation amending §36B is watched continuously, because that is the failure mode this discipline exists for — the verification pass found a repayment cap that Congress had already repealed underneath a rules file nobody was watching. Indexed values drift predictably; statutory structure does not.

What we never publish

• AI-generated numbers or auto-published AI content.
• Outcome promises — everything is an estimate under current rules.
• Advice to time income — that conversation belongs with your tax professional.
• A figure without its source. Sample or placeholder data stays labelled as sample data on every surface it reaches.