# Test results (2026-11-18)

Built with g++ 12 on the supernode: `tests/test_core.cpp` against `core/TakeStampCore.h`
(the same code the AUs use).

```
[jitter] frames=1224 (incl dups) crcBad=0
[placement 48k->48k] samples=960000 mismatched=0 maxAbsErr=0 uncovered=0
[xcorr] placement error take1 @5s = 0 samples, take2 @12s = 0 samples
[take-order] late duplicate of take1 ignored, take2 still at offset 0
[stamp] ns round trip worst error = 0 samples (44.1-192k, 0-2.3h)
[ping] true loop    37 smp -> measured    38 (err +1, peak ratio 8.2x)
[ping] true loop   512 smp -> measured   513 (err +1, peak ratio 624.2x)
[ping] true loop  1189 smp -> measured  1190 (err +1, peak ratio 546.0x)
[ping] true loop  2777 smp -> measured  2778 (err +1, peak ratio 530.8x)
[ping] true loop  9001 smp -> measured  9002 (err +1, peak ratio 599.6x)
[print] loop 1189 smp, printed take placement error = 0 samples
[relay e2e] frames=400 crcBad=0 covered 1776123..2160123 (expected 1776123..2160123) mismatched samples=0 placement error=0 samples
ALL PASS
```

- The jitter test sends 20 s at 48 kHz stereo in 20 ms frames. Every frame gets a random 50-500 ms
  delay and the delivery order is shuffled, 2% of frames are duplicated, and a punch-in take covers
  10-14 s. The receiver renders in random 32-2048-sample blocks the way a host does. Every one of
  the 960,000 output samples was bit-identical to what Marcus sang at that position.
- On the ping, the simulated preamp includes a one-pole HF roll-off with roughly 0.7 samples of
  group delay, so the +1 is real chain delay that the ping is meant to catch. The print loop is 0
  error by construction.
- The relay e2e ran through `relay/server.py` over HTTP. Each 100 ms POST got its own random
  50-500 ms delay (60-497 ms actual), so frames reached the relay out of order. A long-polling
  receiver pulled them, and every received sample matched (CRC 0 bad).

## Session mode (producer-controlled record), 2026-11-18 23:00 CT

See `PROOF.md`. Through the public relay, across 5 network scenarios (0, 300, 1,859 ±150 ms reordered,
+250 ms mid-take delay jump, +250 ms performer clock jump), 250/250 claps placed on the exact sample they
were sung on (≤0.415 smp = 0.0086 ms off the fractional 130.81 BPM grid), monitor vocal-to-beat offset 0,
performer auto-started every time, CRC bad 0. `test_core` ALL PASS after the header change (payload_len
now written, optional take_id at offset 56). `test_session_hdr` C++ <-> Python round trip OK.
