diff --git a/docs/architecture.md b/docs/architecture.md
index a3b6f23..2d56351 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -71,6 +71,30 @@ durable relay cursor.
 | conversion + validation | `convert.zig`, `wire.zig` |
 | verification + repair | `verify.zig`, `repair.zig` |
 
+#### Four worker pools, four different counts
+
+"32 workers" in one document and "100 workers" in another are both correct and
+describe different pools. Nothing here is stale:
+
+| pool | workers | bounded queue = `workers * 2` | source |
+|---|---|---|---|
+| initial backfill crawl | **100** | **200** | `ingest/backfill/engine.zig` `default_workers` |
+| live scheduler (steady state) | **32** | **64** | `ingest/pipeline.zig` `worker_count` |
+| live repair | **32** | **64** | `ingest/repair.zig` `worker_count` |
+| failed-repo retry | **16** (4 per host) | — | `ingest/backfill/retry.zig` `default_workers` |
+
+Only the first is operator-tunable in practice (`--backfill-workers`); the live
+counts are matched to upstream Atmos defaults rather than chosen here.
+
+**`workers * 2` is one expression serving three pools.** It bounds backfill
+dispatch, the live scheduler's per-DID pending capacity, and the repair job
+queue — so `stream_backfill_queue_depth` reads 200 while `live-scheduler.md` and
+`live-repair.md` both describe a "64", and all three are the same rule. It is
+also the bound whose interaction with a 100,000-repo batch caused the experiment
+4 deadlock (`invariants.md`, "a bounded producer must not outrun its consumer"),
+so raising `--backfill-workers` moves a deadlock-relevant bound, not just
+concurrency.
+
 ### Storage — the segment format and the metadata store
 
 Two places hold state, and the durability ordering between them is the