diff --git a/docs/ingest-pacing-evidence-2026-08-05.md b/docs/ingest-pacing-evidence-2026-08-05.md
new file mode 100644
index 0000000..6d3d805
--- /dev/null
+++ b/docs/ingest-pacing-evidence-2026-08-05.md
@@ -0,0 +1,57 @@
+# live ingest is self-paced at ~360 events/s — evidence bundle
+
+Measured 2026-08-05 ~22:30Z on the experiment 6 box (ccx53, 32 cores, 128 GB),
+image `6a511fa`, `--no-verify` active. For whoever owns the consumer/pipeline.
+
+## the number
+
+`rate(stream_upstream_seq[1h])` ≈ **360-366 seqs/s**, flat across 10m/30m/2h
+windows, zero firehose disconnects. At ~360/s against the relay's ~300-400/s
+production, catch-up slope is ~1.1-1.2 — closing a 54h lag takes weeks.
+
+## what it is NOT
+
+- **Not the relay.** A bare Python websocket client on the same host, same
+  relay, nearby backlog cursor (`32417500000`) drained **190.9 MB / ~3,700
+  frames/s over 10s**. The pipe offers ≥10x what stream consumes.
+- **Not disk.** iostat during ingest: ~0.2 writes/s, 0.04% util. (Merge wrote
+  ~70k events/s through this same archive on this volume.)
+- **Not CPU or memory.** 32 cores mostly idle; consumer threads parked in
+  `hrtimer_nanosleep`; RSS flat at 53 GB of 122.8.
+- **Not verification.** `--no-verify` is on; the plc.directory throttle
+  (~1.5/s, ~50% key-cache miss) was the previous bottleneck and is gone.
+
+## where to look
+
+`src/internal/ingest/pipeline.zig:17-21`:
+
+```zig
+const worker_count = 32;
+const key_queue_capacity = worker_count * 2;
+const result_capacity = 4096;
+const batch_size = 50;
+const batch_timeout ... fromMilliseconds(500);
+```
+
+360/s with all resources idle and threads sleeping is the signature of a
+pacing/batching structure, not saturation. Suggestive arithmetic: the observed
+rate ≈ 9 events per 25 ms tick. Prior art: the bootstrap-era measurement that
+4x workers bought 1.23x and "production is slower than a plain fetcher on half
+the cores" — the serialization was already visible then.
+
+## reproduction
+
+```bash
+# consumption rate
+rate(stream_upstream_seq[30m])                       # via prometheus on the box
+# offered rate (same box, bare client)
+# see the 12-line python in this doc's git history / session notes: raw ws
+# handshake to bsky.network with a backlog cursor, count frames for 10s
+```
+
+## why it matters
+
+The archive serves (receipt: `experiment-6-serving-receipt.md`) but sits ~54h
+behind live and closes the gap at ~0.1-0.2h per hour. Every day of glide costs
+~€33 (ccx53 pricing). A consumer that merely kept pace with the bare client's
+drain rate would catch up in under 8 hours.