Releases: dweekly/cyrinx
Release list
v2.0.0 — measured 65.875 kbps acoustic goodput
v2.0.0 — measured 65.875 kbps acoustic goodput
Cyrinx 2.0 substantially raises measured MacBook Pro-to-Pixel 7a near-field
goodput while retaining explicit accounting and resilience boundaries.
Measured results
- The schedule-comparable flagship measured 65.875 kbps mean strict ordered
byte-verified goodput, recovering 4,215/4,280 blocks (98.481%). It is 1.801x
the accepted 36.571 kbps benchmark and 49.0% above its prospective paired
44.199 kbps control. - The candidate won all eight order-balanced pairs (
p = 1/256) but failed its
predeclared baseline-equivalent reliability gate: the paired control
recovered 2,999/3,000 blocks (99.967%). - A separately classified zero-gap profile measured 69.652 kbps and recovered
6,129/6,760 blocks (90.666%). It is not schedule-equivalent to the flagship
and also failed its reliability gate. - Replaying identical flagship captures through frozen receivers isolated a
+347-block gain from frequency-local known-pilot reliability: 4,215 versus
3,868 recovered blocks, with all eight runs improving.
Engineering changes
- Canonical portable-C 64-QAM rate-2/3 profiles with CP96 and route-qualified
pilot schedules. - Frequency-local known-pilot LLR reliability with deterministic edge and
nonfinite-input behavior. - Payload-independent held-out-pilot selection between the primary microphone
and two-microphone maximal-ratio combining. - Fixed receiver-contract validation and thin Swift bindings for the new
profiles and diversity diagnostics. - Evidence-bound campaign accounting, provenance, retained-failure semantics,
and paired statistical gates. - A separate 11-page Cyrinx 2.0 follow-on paper documenting improvements,
rationale, rejected branches, limitations, and future 3 ft work.
Scope and limitations
These are host-decoded measurements for one near-field route, one MacBook Pro,
one Pixel 7a, and one pose. The flagship improves throughput but does not match
the conservative control's block recovery. Android JNI/live decoding, a
reliability-qualified 3 ft profile, an integrated ultrasonic mode, and a
pleasant-audible mode remain future work.
Tracked ledgers and content hashes are included in the source archives. Full
raw captures, execution manifests, frozen binaries, and detailed replay reports
remain local ignored artifacts and are not part of this release.
Release artifacts
cyrinx-2-goodput.pdf— focused Cyrinx 2.0 follow-on papercyrinx-acoustic-link.pdf— revised Cyrinx 1.0 system paper and historical
measurement record
v1.0.0 — data over sound, measured
First public release.
What this is: a research acoustic modem that moves real data through the air as sound between commodity devices — an ordinary laptop speaker to an ordinary phone microphone. No radio, no pairing, no network.
Measured, ordered-byte-verified goodput:
| Link | Rate |
|---|---|
| Mac → Pixel 7a | 36.6 kbps (39.3 kbps decoded by the shipped C library) |
| Pixel 7a → Mac | 27.3 kbps |
| Mac → iPhone 17 Pro Max | 36.57 kbps |
| iPhone 17 Pro Max → Mac | 16.87 kbps (speaker-limited) |
Graceful degradation, OTA-validated: 48 kbps (clean contact) → 11.6 kbps (reverberant — 0/75 blocks decodable on either mic alone, 75/75 recovered by two-mic MRC) → 138 bps RS-coded MFSK floor (shadowed) → never zero.
In the box: portable C bulk PHY (CCyrinx, Apache-2.0, KISS FFT default / vDSP backend) with Swift bindings; golden-vector test contract across C/Swift/Python/Kotlin implementations; adaptive EVM-probe sounder with MRC-aware MCS selection; iOS/Android HIL apps; the full reproducible measurement harness; the 27-page whitepaper (attached); and the negative-findings log of dead ends already paid for.
Website: https://cyrinx.org — hear an actual frame synthesized in your browser.
🤖 Developed almost entirely by AI coding agents driving a physical audio bench; the paper's §Benchmark documents which agent era built and measured what.