diff --git a/app/services/prefetcher.ts b/app/services/prefetcher.ts
index d2b0391..aa3a8e3 100644
--- a/app/services/prefetcher.ts
+++ b/app/services/prefetcher.ts
@@ -506,7 +506,10 @@ export class Prefetcher {
    * Up to corpus.ahead capture ranges whose chunks aren't all in the corpus, by priority: planned ones not started
    * (not leased soon), and done ones (a window of every group, ranges:capture, takes those indexed already too);
    * as many of a group as its connections allow (perGroup / perRange): in id order the 16 were one group's, which
-   * took 4 of the loops and left the rest idle (30k articles/s; LOG.md 2026-10-23 (100)). A range found complete
+   * took 4 of the loops and left the rest idle (30k articles/s; LOG.md 2026-10-23 (100)). The candidates are taken
+   * a range of every group at a time (the n-th unstored range of each group before any group's n+1-th): a group's
+   * ranges have consecutive ids, and once the small groups of a window were stored the 512 lowest ids were all one
+   * group's, two targets on two connections (10 ranges an hour; LOG.md 2026-10-25 (110)). A range found complete
    * gets stored_at and isn't read again.
    */
   private async corpusTargets(rows: (q: any) => any, taken: Set<number>): Promise<Target[]> {
@@ -524,7 +527,7 @@ export class Prefetcher {
           .where((p: any) => p.where('r.status', 'planned').where('r.spooled_chunks', 0))
           .orWhere('r.status', 'done')
       )
-      .orderBy([{ column: 'r.priority' }, { column: 'r.id' }])
+      .orderByRaw('r.priority, row_number() OVER (PARTITION BY r.group_id ORDER BY r.id), r.id')
       .limit(ahead * 16)
     const targets: Target[] = []
     for (const row of candidates) {