| # Compiled-local Montgomery v42 |
|
|
| Status: validated competition candidate; final eligibility remains subject to |
| manual organizer review of the learned recurrent-program boundary. |
|
|
| V42 is a direct successor to submitted Model 1 |
| (`c7cb870bfb68bc278be5fe81ec636c012f043227`). It keeps Model 1's learned raw |
| reducer, Montgomery arithmetic, judge/revisor, and output path. It makes two |
| inference-only efficiency changes without changing their learned decisions: |
|
|
| 1. two 513-class learned clock outputs are GPU-distilled into exact learned |
| scalar clock tables; and |
| 2. the trained factor and cancellation-review networks are evaluated over |
| their complete 16-value local domains once per raw batch, then reused as |
| learned lookup maps during the recurrence. |
|
|
| The checkpoint has 412,204 parameters versus Model 1's 938,028. On the same |
| official 4-CPU sandbox image and identical 1,100-case seed, V42 scored |
| 1,000/1,000 on scored Tier 1--10 in 66.4 seconds. The adjacent Model 1 control |
| immediately before it took 69.0 seconds; a second control immediately after it |
| took 93.3 seconds as the host slowed. The conservative measured speedup is |
| therefore 3.8%. |
|
|
| The unchanged sealed hard corpus scored 960/960, including 480 |
| quotient-boundary cases, 240 structured maximum-length cases, and 911/911 |
| nonzero answers. Randomizing the packaged weights produced 0/73 correct |
| nonzero answers. Tier 0 is an unscored pure-multiplication diagnostic and is |
| not part of this model's modular-multiplication contract. |
|
|
| The supported official domain is unchanged: modulus width through 2,048 bits |
| and raw operand width through 4,096 bits (Tier 1--10). V42 is not evidence for |
| moduli beyond the published 2,048-bit maximum. |
|
|
