Corrections

What I got wrong, and when I fixed it

A record that cannot be corrected cannot be trusted. Each entry says what was published, what was wrong and what replaced it.

Log

1. 11 October 2026

   The first version of this site was withdrawn and rebuilt

   The first version went live on 11 October 2026. Later that day I checked every figure and identifier on it against the results files and removed what I could not trace.

   • Identifiers. A DOI and an organisation registry link were shown for work that had never been deposited. Neither existed. They are gone, and these records carry no DOI.
   • “Merkle root”. The digest shown was the SHA-256 of a single audio file, not a root over the dataset. There is no combined digest yet, only one per file.
   • Word error rates by codec. Five values with confidence intervals appeared in no results file. They are replaced by the means in the E2 ledger, as point estimates.
   • Agent coordination panel. The “12,400 holds” and “0.00% collisions” figures had no source, and an 84.7% consistency figure was another experiment’s routing accuracy. The record now shows the 16-attempt diagnostic.
   • Error correlation. A value of 0.42 matched no statistic. The simulated correlation is 0.75, and the record now says it is a simulation.
   • Energy per token. A figure of 1.66 J came from dividing an assumed power by an assumed speed. Wall-meter runs read 2.54 to 3.81.
   • Dataset description. The files are 16 kHz with 33 conditions, not 8 kHz with 30.
   • Labels. “Proven”, “empirically verified”, “gate cleared”, “6/6 pass” and an RFC-style header implied a reviewer and a standards body. There is neither. They are removed.
2. 10 October 2026

   Write-up of the silence experiment disagreed with its results file

   The prose summary of the first experiment said an energy gate removed 90 to 100% of invented words. The results file shows 19.51 to 25.0%, and no effect on noise or squelch. The results file is right. This site uses it.