diff --git a/.gitignore b/.gitignore
index f062debc..ce1cda54 100644
--- a/.gitignore
+++ b/.gitignore
@@ -53,6 +53,7 @@ CTestTestfile.cmake
 recordings/*
 !recordings/ota_capture_*/
 !recordings/ota_full_session_*/
+!recordings/ota_noise_bed_*/
 
 # ota_simulator runtime outputs (when scenario writes relative paths)
 /out_*.wav
diff --git a/recordings/ota_noise_bed_2026-05-18_20m_14100/README.md b/recordings/ota_noise_bed_2026-05-18_20m_14100/README.md
new file mode 100644
index 00000000..47f59bf2
--- /dev/null
+++ b/recordings/ota_noise_bed_2026-05-18_20m_14100/README.md
@@ -0,0 +1,62 @@
+# 20m noise bed — 14.100 MHz, 2026-05-18 morning
+
+Real-HF reference capture for the OTASim noise-overlay workstream (task #71).
+
+## Source
+
+- Date / time: 2026-05-18 ~11:30 AST (15:30 UTC)
+- Frequency: 14.100 MHz USB
+- Band: 20m
+- Radio: Yaesu FT-891, USB data out (level 30 default)
+- Recording chain: FT-891 USB → Mac audio input → ultra_gui `-rec`
+- Sample format: 16-bit PCM, mono, 48 kHz
+- Duration: 600.00 s (10:00 exactly)
+- Source bundle: ultra_gui diagnostics report 0ad79b
+- Author callsign: redacted (privacy.callsigns_included=false in the bundle)
+
+## Levels
+
+| Metric | Value |
+|---|---|
+| Overall RMS | 0.0363 |
+| Peak | 0.4738 |
+| Per-5s segment RMS | 0.027 – 0.042 (median 0.037, stddev 0.003) |
+| Clipping (>0.9) | 0% |
+| Dynamic range from 16-bit floor | ~30 dB |
+
+The audio level sits a hair below the modem-decode calibration target
+of RMS 0.05-0.25 (CLAUDE.md), but for noise-bed playback purposes the
+absolute level is normalized at load time. The capture is solidly
+within the dynamic-range envelope and has no clipping.
+
+## Spectral character
+
+| Band | Energy share (mid chunk) |
+|---|---|
+| 0-300 Hz | ~1% |
+| 300-1000 Hz | ~16% |
+| 1000-2000 Hz | ~56% |
+| 2000-3000 Hz | ~27% |
+| 3000+ Hz | <1% |
+
+~98% of energy in the SSB passband (300-3000 Hz). No DC contamination,
+no AC-mains bleed, no high-frequency junk. Real-HF noise shape.
+
+## Content
+
+- Steady background band noise (broadband, structured — not Gaussian)
+- NCDXF beacon network rotation (14.100 hosts the IBP beacons; one
+  station transmits every 10 seconds). Over 10 min you get 60+
+  beacon transmissions interleaved with quiet periods.
+- No dominant strong nearby signal (per-segment RMS stddev only 0.003).
+
+## How to use
+
+Load this WAV at OTASim service startup, normalize to unit RMS,
+then scale per-tick by `kModemReferenceRms * 10^(-snr_db/20)` (same
+formula as AWGN) to feed continuous real-HF-character noise to all
+session receivers. Loop on `position % length` — at 10 min there is
+no audible periodicity.
+
+See `docs/AGENT_TASK_BACKLOG.md` (task #71) and prior `ota_simulator
+v2` real-HF noise overlay for the implementation pattern.
diff --git a/recordings/ota_noise_bed_2026-05-18_20m_14100/noise_bed.wav b/recordings/ota_noise_bed_2026-05-18_20m_14100/noise_bed.wav
new file mode 100644
index 00000000..d1de9e0d
Binary files /dev/null and b/recordings/ota_noise_bed_2026-05-18_20m_14100/noise_bed.wav differ