TraceDuo: Provenance-Aware Continual Learning Beyond Label Aggregation

Continual learning from weak supervision must account for shared source errors, reliability drift, and partial corruption. TraceDuo combines temporal sourcecluster label estimation, an exactly conserving source-deletion gradient decomposition, component-wise gradient clipping, and certified update routing. Full and adapter-only candidates are frozen before protected-risk and current-domain utility bounds govern commitment. Our analysis characterizes the independentcluster support and score-separation regimes in which the pseudo-label bound becomes informative, and gives local alignment conditions for beneficial deletionbased component control. Under fresh conditional IID sampling, the V2 certificate bounds the probability of any invalid commitment by 0.05 per stream and limits net risk increase relative to a fixed initial anchor to 0.02 on the simultaneous coverage event. Across the V1 experiments on CIFAR-100 and the binary Amazon, Elliptic, and TON tasks, domain final balanced accuracies are 98.10%, 99.20%, 98.80%, and 99.30%, respectively. Their macro average is 98.85%, compared with 97.55% for RobustAgg+DERpp. A common task–time definition gives 99.17% current-task accuracy and 0.40 percentage points of forgetting. Factorial routing and optimizer comparisons quantify component effects, while a separate numerical check tests no-clipping equivalence. Known-risk V2 certification experiments record 95 erroneous five-window trajectories out of 100,000 with joint correction, versus 11,236 without it, and quantify when 512 observations provide useful endpoint power