Model changelog

This is the dated record of every review and release of the engine behind your number. Releases change the model. Reviews check it against new research and sometimes change nothing, and a review that changed nothing is still listed, because "we looked and it held" is part of the record.

The engine's design, every source it draws on and the reasoning behind each release live on the methodology page. This page is the timeline.

Live model: engine 2.2.2 · calibration 2026.09.0 · last reviewed 3 October 2026. Every saved estimate records the engine and calibration version that produced it. When the model changes, your history is re-scored from your original answers under the new version, so an old measurement and a new one are always compared on the same math.

How to read this page

• Engine versions change how the math works. A new engine version can move an estimate.
• Calibration versions change the numbers inside the math, usually one factor's adjustment after new evidence. A new calibration version can move an estimate too, and the entry says in which direction.
• Reviews are the monthly read of the research behind every factor and every action in the habit library. Most reviews change copy, sources and the library, not the model. When one does change the model, it ships as a release and is listed as one.
• Errata are our own mistakes, found and corrected. They are listed as what they were.

2026

3 October 2026 · Monthly science review (October). No engine or calibration change. Every number touched was checked against its primary paper. Four corrections across the articles and the report. The sense-of-purpose figure now reads as the paper reports it, about 24% lower mortality (HR 0.76), on every surface including the paid report; the step-count bands by age in the exercise article were the wrong way round and now follow Paluch 2022 (8,000 to 10,000 steps under 60, 6,000 to 8,000 at 60 and older); the Mediterranean-diet figure is 8% (Sofi 2010) everywhere; the alcohol article's short answer now matches its body.

In the habit library, "Cut caffeine after noon" was promoted to established evidence on Gardiner 2023, a new action was added (swapping table salt for a potassium-enriched salt substitute, from the SSaSS trial, Neal 2021, with a kidney and medication safety note), and the secondhand-smoke action gained its source (He 1999) and a safety note. The library stands at 99 of 167 actions carrying a graded primary source.

2 October 2026 · Engine 2.2.2 released. The potential ceiling and each habit's "years available" now use the best answer a person can still reach. A current or former smoker can never become a never-smoker, so the ceiling for a smoker is now "quit long ago", not "never smoked". Body weight gained its own years-available figure through the same transform as every other habit, so it ranks in the one order with them. No factor's adjustment, endpoint or current estimate changed. Design notes in the version history.

9 September 2026 · Calibration 2026.09.0 released. The body-weight adjustment above BMI 40 keeps rising instead of flattening, and below 18.5 it rises as BMI falls. One value had been standing in for every weight in each tail, so BMI 40 and BMI 54 had scored the same. Nothing between BMI 18.5 and 40 changed, and no estimate became more favourable. The alcohol slider's starting position moved from two drinks a week to zero, because about a third of adults do not drink.

5 September 2026 · Engine 2.2.1 released. A consistency fix. The potential ceiling and every per-habit figure now pass through the same transform as the headline estimate (the same survivorship discount, cohort adjustment and floor). Before this, one habit's "years available" could exceed the whole gap between your estimate and your potential. No adjustment, endpoint or estimate changed.

5 September 2026 · Monthly science review (September). The first review run on the monthly routine, with 26 findings and 15 recommended edits, all applied. The diet figure in the paid report was corrected from 9% to 8% (the 9% belongs to Sofi 2008, the model cites Sofi 2010). Three Blue Zone rationale sentences in the library were replaced with cohort evidence (Wang 2023, Sutin 2026). The ultra-processed-food share of U.S. adult calories was corrected from 57% to 53% (NCHS 2021 to 2023). The 75-minutes-a-week exercise gain was corrected to 1.8 years (Moore 2012). The smoking article's quit-attempt claim was corrected ("quadruples attempts" rather than "triples success"). Citations were advanced to their newest editions (Lindson 2026, Livingston 2024, Kaminski 2026 for NordICC's 13-year follow-up). No model change.

16 August 2026 · Erratum, the life table. An accuracy audit found that the conditional life-expectancy table inside the engine had drifted from the CDC source it cites (United States Life Tables 2023, NVSR volume 74 number 6). Men aged 40 to 60 were overstated by up to 1.0 year and women at 30 and 35 were understated by about 0.5 years, the result of a partial update from an older vintage. The table was corrected cell by cell to the CDC spreadsheet, and a build check now pins it to that source so the drift cannot recur. The same audit corrected fourteen articles, among them a misattributed colonoscopy citation, the alcohol years-lost table (rebuilt on Wood 2018) and a smoking sample size (202,248, not 216,917).

1 July 2026 · Engine 2.2.0 and calibration 2026.07.0 released. The quarterly citation refresh. Two evidence-driven corrections. Light drinking (one to three drinks a week) moved from a small benefit to neutral, on Zhao and Naimi 2023 (4.8 million people, bias-corrected) and two Mendelian-randomization cohorts, which show the protective J-curve was a sick-quitter artifact. The sense-of-purpose adjustment was shrunk because reverse-causation analyses show the raw association is inflated. Sources for smoking, activity, social connection, stress, preventive care, sleep and diet were advanced to larger and newer cohorts, and the research base behind the model passed 40 million participants.

Before July 2026. The model's foundations, from the centering rebuild on, are described in the version history on the methodology page.

What a review checks

Each month the research behind all ten factors is read again, along with the source behind every health claim in the habit library. The check is the same every time. Does the cited paper still say what we say it says, with the effect size we quote. Has a larger or newer study superseded it. Has anything been retracted or corrected. Does every surface that quotes the number (the results screen, the paid report, the library, the articles, this site) quote the same one. Findings become edits, edits become an entry here, and a change to the model becomes a release with a version number you can see on your own saved result.