Auditable Continual Learning: Measured Transactions for Sequential Weight Edits

Continual weight-level learning is usually treated as unsafe because its damage is unbounded and unaccounted. We ask whether a sequence of weight edits can be made transactional — each candidate measured against an explicit trusted base, admitted or rolled back by a mechanical rule, recorded in an append-only ledger — and whether the enforcement does any work. On Qwen2.5-3B-Instruct with MLP-window edits and a 641-item forced-choice margin panel: forecast gating fails (a first-order certificate misses 63.2% of live-loop casualties on its primary run; 59.1% pooled with the seed replication, cluster CI [0.490, 0.673]); measure-then-rollback succeeds (5/5 lessons, flat cost); knowledge-grade transfer appears only under paraphrase-diverse training (22/39 held-out vs 1/39 at base, relation-heterogeneous); and a paired gated-vs-shadow design shows the gate reduces final panel damage on the stream run (21 vs 54). The protocol transfers across a four-family battery and to 7B. Against six measured training configurations on the same 48 facts, no standard fine-tuning configuration paid less interference on the worst general-benchmark slice on the same instruments (on other slices the ordering reverses — reported at full weight); the audited editor was the only method with a measured gate, a cumulative halt (which fired), and an append-only ledger. Honest two-sided result: learning succeeds, collateral remains material. Six rounds of external adversarial review are folded in and credited. Research conducted in close collaboration with an AI assistant (Claude, Anthropic) under a sealed pre-registration discipline; disclosure and contribution statement in the paper.