sean.theinformed.org / training / Dynamic compositional representations / news

All runs finished, 54.09 node-hours used

All 1046 runs across 12 grids finished. The study used 54.09 node-hours and is now complete.

What ran

The following grids ran to completion: grid1, grid1b, ctx1, ctx2, fold1c, ctx3, ctx3-transplant, grid2-role, fold1, fold1b, fold1c-lm, ctx3-lm.

What finished

All 1046 runs finished. No runs are running or waiting to start. The last jobs to finish were the timing tests and the final chunks of the grid2-role, fold1, fold1c-lm, and ctx3-lm grids.

The numbers

Failures and what was done

There were no run failures. Five jobs stopped early due to time limits but completed their work in subsequent chunks. No action was needed.

What was learned

A notebook entry confirmed that grid2-role used 14.73 node-hours (3,770 core-hours), correcting an earlier projection. Another finding confirmed that R4 fails under both the original and corrected readings, supporting the negative result. The fold line closed as negative: R1, R2, and R4 all fail under corrected readings, and R3 is not readable. A design review found six issues requiring fixes before publication, including statistical corrections and larger seed counts for key claims.

What was decided

The study team decided to close the first line (grid1, grid1b, ctx1) as negative, with no condition beating the table baseline under corrected tests. They decided to apply a shared statistics module with proper minimum-effect screening and Holm correction for all results. They also decided to require 15–20 seeds for any claim-bearing cell and to report context-swap and rebinding results as untested until further work.

What is next

The next steps are:

Additionally, the team will carry out post-hoc reading passes per grid and update public pages with final results and methods.

Written by a language model I run at home from the day's records; checked by code against those records. All studies' news.