A new model generation should earn authority, not win a contest

A continual learner keeps meeting newer versions of itself: a model retrained on new data, a fact that has changed, an expert added for a new task. The intuitive picture is a contest. Old and new compete, and the winner takes over. Our new synthesis argues that competition is the wrong word and the wrong mechanism. A new generation should supplant the old one the way trust is earned: by demonstrated competence, in degrees, until its authority is dominant, and without ever gaining the power to rewrite the rules it operates under.

Our own results show what goes wrong when the decision is left to a contest. Filtering what an old model passes on by its own confidence made things worse than no filtering. In a toy knowledge base left to make itself consistent, two stale templates out-voted the updated one, and 44% of updated facts ended up coherently wrong. On Split-CIFAR-100, the error in choosing among frozen experts grew from 0.53 to 0.79 as the field grew from 5 to 20 (ledger 8.18). Confidence, numbers and crowded contests are exactly what the word competition invites.

Graded succession separates three questions that competition blurs: who is right, who acts, and who makes the rules. An evaluator independent of both generations decides who is right, from held-out outcomes. The successor's authority, its share of decisions in each context, rises in steps as evidence accumulates. Anytime-valid statistics such as e-processes let that evidence be checked continuously while bounding the chance of promoting a successor that isn't better. A second test for “worse” demotes it. Industry already ramps new models through shadow, canary and staged traffic with rollback, as in AWS SageMaker and Google's SRE practice, but those ramps advance on a timer and check only that the new model is not worse.

The math of mixing old and new points the same way. Bayesian model averaging, the obvious evidence-weighted choice, ends up selecting a single model and is slow to let a better newcomer catch up. Fixed-share mixtures redistribute a little weight every round, so a better successor takes over in bounded time and the old generation keeps a small floor from which it can be restored. Plain averaging never hands over at all.

Old knowledge also needs protection, but not a uniform kind. In our test of elastic weight consolidation, a single stiffness setting cut correct updates from 62.5% to 13.5% and left stale answers at 81.7% (E4). With no anchor at all, beliefs drifted into the coherent errors above. The design therefore anchors each item to verified evidence, and releases an item's anchor when contradicting evidence for that item is strong enough.

Above all of this sits a constitution: hard constraints and core rules that no amount of earned authority can override. It changes only by explicit, recorded amendment, which takes effect at once rather than gradually. Both OpenAI's Model Spec and Anthropic's constitution already rank authority this way, though neither publishes an amendment procedure. The counterintuitive conclusion is that the two layers should move at different speeds: knowledge gradually and on evidence, the constitution abruptly and deliberately.

This is a design proposal. The synthesis lays out five pre-registrable experiments, S1 to S5, that test whether graded succession beats both a hard switch and a plain average: whether authority follows evidence rather than tenure, whether staged rollout catches a degrading successor, whether protected rules hold, and whether real judges promote successors that game them. None has been run, and every threshold is a placeholder. If the gradient doesn't beat a good canary plus a good evaluator, we will say so.