Keyboard shortcuts

Press or to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Matched MGA route-discovery comparison

The Gemma 4, grammar-random, and evolutionary arms use the same accepted-candidate target, attempt ceiling, canonical direction policy, duplicate filter, inner DE–CMA-ES budget, workers, root seed, and top-N metric.

A route is MGA-qualified when the Rust optimizer returns a finite impulsive MGA score. This is a downstream candidate, not a continuous-thrust GTOC1 solution.

ArmStatusAccepted / targetMGA-qualifiedBest scoreTop-NTop-N sumNiches
gemma4completed100 / 1001001164199.4852019676460.47863
randomcompleted100 / 1001001233733.1472019270002.01996
evolutionarycompleted100 / 1001001278762.7602022140497.51197
gemma4-assistedcompleted100 / 1001001509902.0922026964276.02681
ArmAttempts / ceilingDuplicate rejectedDiversity rejectedTransport failedMGA evaluationsWorker-hWall-hAgent tokens
gemma4190 / 25000900225902730178.73211.6612341623
random166 / 25001749018978695082.9125.1820
evolutionary146 / 25001135019510493977.5694.8480
gemma4-assisted206 / 25000106019115428471.1794.7001083036

The first three rows are the blind matched comparison. gemma4-assisted is a separately named, prior-informed follow-up: it uses the completed random and evolutionary archives to construct a length-stratified candidate menu and therefore is not a fourth independent arm.

Relative to cold Gemma, the assisted follow-up improves the top-N sum by 37.0% and completes in 2.48× less wall time. These are one-seed protocol results, not a general model-capability estimate.

Every reported arm completed. Repeat the blind matched protocol and the explicitly prior-informed follow-up across predeclared seeds before drawing a proposer-capability conclusion.