Projects
Match it to the digit.
Then put a finger on why.
Proofs and results anyone can check.
Proof Grid
No missing cells. What each rebuild was checked against, in which language, and what is proved.
| Project | Checked against | Python | R | Lean proves | Result |
|---|---|---|---|---|---|
| GapYears | Paper's published gaps | – | ✓ | – | 4,070 country-yearsShape holds; exact values moved when WHO revised all 4,070 country-years |
| BloodVsMethylation | Graf et al., HRS 2016 | – | ✓ | – | 2.8 years vs ≈0The disagreement reproduces, wider: 4.3 years against the paper's 2.8 |
| NeighborsMailer | The authors' Stata log | ✓ | – | – | 130/130 matchedBoth main tables, to the last printed digit |
| AncovaThreeWays | The 2006 SAS output | ✓ | ✓ | The table: each exact value rounds to the SAS number | 70/70 matchedPython and R agree to 12 decimals |
| PlaceboVsObserved | The 1955 Francis report | ✓ | ✓ | The verdict, below 1 in 1,000, in whole numbers | 1,064/1,068 matchedThe four misses are the report's own |
| ColonAdjuvant | Python against R, then the paper | ✓ | ✓ | The log-rank verdict, as an exact fraction | 74/74 agreeMismatches with the paper left in view |
| HeadlineExpectedLoss | The archive's data paper | ✓ | – | All 99 two-version verdicts and their totals | 511 checks, 0 mismatchesSimulation checked against exact formulas |
| BcgLatitude | The two source papers | ✓ | ✓ | The flip in the latitude decision, in exact fractions | 164/164 matchedPython from scratch, R with metafor, Lean with no statistics library |
✓ a language the rebuild is written in. Where Python and R both appear, each was written on its own and the two are compared. Lean is used only where the decision can be stated exactly; it does not fit models that come out of an optimiser.