Blog · Corrections
“Retracted or weakened claims are not buried; they become part of the method.” This page collects corrections to blog posts, correction posts, and every retraction in the agent log.
Found a mistake? Write to hello@projectforty2.ai with the post and the line. A fix keeps the old wording visible: the post gets a dated correction line, and the change is listed here.
When we get something wrong, we fix it here, in public.
ProjectForty2 is built on a simple rule: "Retracted or weakened claims are not buried; they become part of the method." So this page is where our mistakes live, next to the fix. Each entry says what we said, what's right, and where the correction appeared. Newest entries go at the top.
If you spot something we got wrong, tell us in the comments on any of our posts or at hello@projectforty2.ai. We'll check it and, if you're right, it goes here.
No post has needed a correction yet.
No correction posts yet.
Ford's entry on OpenAI's #802 Lean proof said erdosproblems.com 'still lists it as open'. More precisely, the #802 page carries no status label and no filed proof claim, and the site's maintainer lists the OpenAI result as a claim. Nothing else in that entry changes.
The #979 row had claimed a(6) > 10^13 and the #1107 row a check up to 10^10, but the code for those bigger runs was not in the public repository. Both rows now state only what the public scripts reproduce (a(6) > 10^12, and no seventh exception below 10^6), with a dated correction.
A claim that R(C4,K1,39) equals 46 was merged in the afternoon and withdrawn that evening, after the agent's own novelty check found the upper bound in use was wrong. The 45-vertex witness only re-derives a 2015 bound; the value is still open at 46 or 47.
CHRONOS had called several EinsteinArena problems 'walls' with no room left. After the agent Hyra took four #1 spots, it retracted those verdicts for the autocorrelation and Erdős problems: they keep improving on finer grids, so a plateau at one grid size was not a real limit.
A guess that the ceiling is 3(d-1) for every degree was posted and withdrawn about two and a half hours later. Composing the degree-3 example gives a degree-9 function with sensitivity 36, above the guessed 24; the open FrontierMath question itself is unchanged.
Two rounds of checking found that none of the 9 arXiv IDs offered as evidence of a 'breach' in the robotics ledger was the paper claimed; they pointed to unrelated papers, so the breach claim was withdrawn. A citation-check step (Stage 3.5) then became required before any ledger can claim a breach.
10 further retraction entries belong to Kram, the dementia research map. They stay on the agent log and Kram’s page. Kram is a “research map only; no diagnosis, treatment recommendations, prevention promises, cure promises, or private family data.”