Methodology

Every number on this site comes from a dataset committed to the repository that builds it. No figure is written into prose by hand. This page sets out where the data comes from, what has to pass before a page can be generated, and — the part that matters most — every place where our arithmetic and a published source disagree.

CPI-W months committed

630January 1974 onward

Part B premium years

22

Benefit ladder rungs

46$100 steps, from the source bands

Birth years, full retirement age

24

1 · Where the numbers come from

Social Security figures come from the Office of the Chief Actuary’s own tables and the annual fact sheet. Medicare figures come from the annual Federal Register notice — the legally operative text — retrieved through govinfo. CPI-W comes from the Bureau of Labor Statistics flat file for series CWUR0000SA0.

Two sources are deliberately not used. cms.gov blocks automated retrieval outright. PDF mirrors of CMS fact sheets have been observed to parse into plausible, well-formed and entirely wrong tables carrying figures from an earlier year — a wrong table that parses cleanly is more dangerous than one that fails. Search-result snippets are also never used for index values: a snippet was observed assigning one month’s CPI-W value to the adjacent month.

Every dataset, with its URL, retrieval date and licence note.

2 · What has to pass before a page exists

• The adjustment oracle. All 51 published adjustments are recomputed from the committed CPI-W series using the three statutory regimes and matched against SSA’s table.
• The Medicare oracle. Every income-related tier is reconstructed from the statutory cost-share multipliers; every Part D amount is reproduced exactly from the national base premium; the Part A coinsurance ratios are asserted exactly.
• Fact-sheet reconciliation. Every 2026 parameter is matched against both the SSA fact sheet and the Federal Register determination notice — two official sources that must agree.
• SSI derivation. The entire published SSI series is reconstructed from unrounded annual amounts and matched to SSA’s monthly rates year by year.
• Rendered-page gates. After the site is built, every currency and percentage token in the output is checked back against an engine output or a committed value; projected figures are checked for their labelling in the built HTML; and the year convention is checked on every year page.

3 · Rounding, at each step

Everything that can be integer arithmetic is integer arithmetic. A published adjustment is carried as an integer number of tenths of a percent and applied to an integer number of cents, because the floating-point route can land a hair below a whole dollar and floor to the wrong one — a real $1 a month error on a site whose whole promise is exactness. The computation in full detail.

4 · Where a source and our arithmetic disagree

These are reported rather than quietly reconciled. In each case the reasoning is set out so it can be checked, and in three of them a widely repeated figure is wrong and this site is right.

4.1 · 6 adjustments cannot be reproduced from the current index

SSA computed the early adjustments from the CPI-W as published on a 1967 base. BLS rebased the index in 1988 and now serves one-decimal conversions of those months, which discards roughly two thirds of the significant digits SSA worked from. For 6 determinations that is enough to flip the final tenth. Affected payable years: 1976, 1978, 1979, 1980, 1984, 1985. SSA’s published figure is what those pages state, each carries the disclosure, and the build gate allow-lists precisely those years so the list cannot grow silently. The full table of differences.

4.2 · The quarterly rounding rule is era-dependent

A single rounding rule for the quarterly mean is wrong. Determinations through 2006 round it to the nearest 0.1; from 2007 they round to 0.001. The proof is SSA’s own footnote that the 1999 adjustment “was originally determined as 2.4%” — only the tenth rule reproduces that, and a flat three-decimal rule would make the statutory override that followed look like no override at all.

4.3 · The hold-harmless premium on a $600 benefit is $201.00, not $201.80

The capped premium is the prior standard premium plus the beneficiary’s dollar adjustment. That dollar adjustment is the difference between two floored benefit amounts, because monthly benefits floor to the whole dollar before anything compares them. On a $600 benefit the floored adjustment is +$16, so the capped premium is $201.00 — not the $201.80 that using the unfloored increase produces. Derived numbers beat asserted ones; where a computation and a piece of prose disagree here, the computation wins and the prose changes.

4.4 · The hold-harmless boundary is between $642 and $643, not at $639.29

The continuous break-even is $639.29 and is correct as published derived arithmetic. The integer boundary sits higher, for the same flooring reason: a $642 benefit still binds and a $643 benefit clears. Both figures are shown on the hold-harmless page.

4.5 · The benefit ladder steps in $100 because the source does

SSA publishes the retired-worker benefit distribution in $100 bands and no finer table exists. The ladder step follows the published band width rather than being chosen to produce a page count — 46 rungs from $500 to $5,000. If the source were ever republished at a different granularity, the ladder would change to match it, not the other way round.

4.6 · One CPI-W month where BLS and SSA disagree

Across the 628 overlapping months checked, BLS’s series and SSA’s republished table agree exactly on all but one. BLS is the producer of record, so BLS’s value is committed. The month feeds no adjustment computation under any regime, so it changes no published figure — and the ingest gate is strict where it matters: a disagreement in a month that does feed a computation fails the build outright rather than picking a side. It is logged in the corrections record.

4.7 · Medicare coverage is narrower than it could look

22 years of Part B premiums and 14 income-related bracket tables are published here. Some years are absent, for two documented reasons: the 2007 and 2008 tiers are phase-in figures that cannot be validated against the statutory multipliers, and several later notices state the tier percentages in prose without publishing a table at all. This site does not hand-key a table it could not read from a primary source. The missing years are listed with the reason in each case.

5 · Scope limits, stated as limits

• National all-items CPI-W only. BLS affirmatively measured the effect of its 2025 collection reductions on the national all-items index and found it negligible, and it explicitly states it did not measure the effect on subnational or component indexes. Staying on national all-items keeps every index value here inside the one zone BLS has vouched for. This is a binding limit, not a preference. The detail.
• October 2025 does not exist. It is stored as an explicit null, and the function that averages a quarter throws rather than skipping it. Twelve-month comparisons crossing it render a dash, never a number. Why it is permanent.
• No projected figure is ever rendered bare. A record that is not announced can only reach a page through a component that requires its assumption as a string, and the built HTML is re-checked for the marker after the site is generated. The projection.
• No claiming advice, ever. Rates, history, arithmetic and mechanism. Nothing about when to claim, whether to file at 62 or 70, or which Medicare plan to buy. The editorial policy states it as a permanent boundary.
• The following year’s Medicare premium is never estimated. It is announced about a month after the adjustment, and until it is published no honest net figure exists. The calendar.

6 · What happens when a source changes

A revised figure is added as a new dated record rather than overwriting an old one, the affected pages are rebuilt, and the change is logged. The correction log is the accountability record — this site publishes under a brand rather than a personal byline, and a dated, public list of what changed and why is what stands in its place.

Related

• COLA computationthree regimes
• Sourcesevery dataset
• Correctionsthe dated log
• Editorial policyscope and standards
• Aboutwho publishes this
• Data qualitythe index itself