All runs finished for grids grid1 and grid1b
Dynamic compositional representations ยท October 3, 2026
The grid1 and grid1b runs are complete, with 270 of 270 runs finished. The study has no next steps queued.
What ran
The study ran two grids: grid1 and grid1b. grid1 included 192 runs across conditions like sparse context, dense context, static sparse, and table variants. grid1b included 78 runs focusing on refined versions of sparse and dense conditions, including J1 and J4 variants.
What finished
All 192 runs in grid1 finished. All 78 runs in grid1b finished. The rescoring job (id 3427641) and timing test (id 3423233) also completed.
The numbers
- Total runs: 270
- Finished runs: 270
- Failed runs: 0
- Running runs: 0
- Not started: 0
- Node-hours used so far: 7.03
- Node-hours projected: 7.04
- Core-hours per node-hour: 128
- Mean run time: 5 minutes
Failures and what was done
One job (id 3423287, grid1 chunk 2 of 6) failed with exit code 1. The failure was logged but did not stop the chain; subsequent chunks completed. No corrective action is noted in the facts.
What was learned
The notebook entry from the independent review found that H1 fails (1 of 4 cells pass), with role binding passing (0.565 vs 0.200) but combo and role-20k not. rebind-20k is at ceiling for all conditions (0.998), so it cannot decide hypotheses. H3 fails on both cells in grid1b: sparse arm loses 0.072 with more refinement, dense arm gains only 0.012; on rebind-20k, dense gains 0.253, sparse gains 0.006. No constructor beats the plain table; dense context hurts role binding. The reference line failed because the quantized baseline scored 0.001 on scan-addjump and had dictionary collapse.
What was decided
The review concluded that affected runs and reporting must be corrected before interpreting grid1 as a matched-budget study. The notebook noted that the reference line incorrectly compared training sets and must be reworded; a cluster job with published code is packaged. The task form was fixed before cluster use: combo and role keep hard held-out splits, depth is a floor, and rebinding and role have a 20k low-data arm.
What is next
There are no next steps listed in the queue. The study has no jobs waiting or scheduled. A prior notebook entry named grid2-role as a proposed second grid, but it is not submitted or queued.