The Bonded Pair is an honest lab that trains two small models coupled on each other's state and forced to disagree, then measures them against an equal-capacity monolith and an independent ensemble across many random seeds. You run the experiment in-browser and see the accuracy of each system and their decision boundaries. The reproducible finding: two small models beat one small model, but the coupled bonded pair beats neither a plain ensemble nor a same-size monolith.
the bonded pair §17, actually built & measured
Two small models co-trained on each other's state, forced to disagree and resolve — measured honestly against a plain ensemble and an equal-capacity monolith. Run it yourself. This one isn't a product pitch; it's a reproducible result, reported straight.
test accuracy — mean over the seeds (higher is better)
Press run. Each system trains from scratch on every seed — nothing is cached, the numbers are live.
what each model learned — decision boundaries on seed 1