Bluetooth timing
research.

PToB researches timing across Bluetooth devices, USB clients, Linux gateways and Starlink over WiFi. Explore the hardware, what the experiments show, and what each result means for your setup.

17 papers & guides · Updated

Start with the whole system.

The refreshed 30 October viability preprint connects source-scoped clocks, timing mesh, hardware, USB/macOS and sampled holdover. Start with this 17-page synthesis of the research and its tradeoffs.

Read the overview

Earlier viability paper · 16 October 2026 (PDF)

Affordable timing from your Starlink

Already have Starlink and WiFi? Explore how an ESP32-S3 can combine terminal NTP with WiFi beacons. Two runs reached 19.226 µs and 8.240 µs RMS after prior GNSS phase calibration; portable UTC accuracy remains unqualified.

Explore Starlink timing over WiFi

Papers & reading paths

Begin with the foundations, follow an implementation, or explore a proposed application. Every paper has an HTML summary, its research scope and the original PDF.

Foundations

Understand the system, then examine the protocol and its measured basis.

Implementation & time transfer

Explore Starlink timing over WiFi, improve event capture on Zephyr and ESP32-S3, and follow time into macOS and a Raspberry Pi NTP/PTP gateway.

Applications & future hardware

Explore proposed uses and the validation they still need.

What the experiments show

These records measure different clocks, paths and references. Read each error statistic with its observation count and method; the records are not a combined benchmark.

Pi 5: from M-HAT PPS to Linux time

The fixed 600.026-second window retained 600/600 new PPS assertions after 120.041 seconds of warmup, with one source cohort and no observed pulse gaps or service restarts. GPIO phase RMS was 872.002 ns, absolute P99 4.809 µs and maximum 7.017 µs; 46/600 observations exceeded 1 µs. Every tail remains included.

These are GPIO IRQ timestamps against the same Pi system clock disciplined by chrony. The separately sampled chrony tracking RMS median was 72 ns; neither metric independently measures UTC or electrical output accuracy. Loopback NTP and Wi-Fi PTP egress passed separate protocol checks. Ethernet was disconnected; wired endpoint timing and reverse PTP ingress remain unmeasured. A separate exploratory Linux USB/SOF diagnostic follows.

Pi 5: an exploratory USB/SOF comparison

On , a T096 connected over USB fed a shadow-only Linux start-of-frame association on the same Pi 5, while chrony stayed disciplined by the M-HAT PPS. The USB clock and the HAT both followed one grandmaster over Bluetooth. The fixed 600-second run was a diagnostic, not a qualification: its readiness gate had not passed, the grandmaster restarted once, creating two source cohorts, and 1,799 of 4,800 USB attempts were unavailable.

Each PPS assertion was joined to the nearest same-cohort USB observation, so adding the USB and GPIO phases cancels the Pi system clock both share. In the first cohort (280 of 281 pulses matched), USB and GPIO phases each showed about 7 µs RMS but moved as opposite-sign mirror images (r 0.989). Their sum had 1.099 µs RMS, 1.048 µs SD, absolute P99 6.093 µs and maximum 6.381 µs, with 6 above 5 µs. The shorter second cohort (94 of 111 matched) includes source warmup and a 1 ms association tail; its 65 report-ready rows gave 2.505 µs RMS. The cohorts are not pooled.

We found no Linux interface that timestamps host USB frames. The observer instead bracketed reads of the xHCI frame index, with a mean bracket of about 19.8 µs and maximum 63.8 µs. A fixed −25.6 µs USB offset was fitted from these data, so absolute USB bias is unknown. 66 frame-label disagreements cluster just below the 1 ms frame boundary; their cause is unresolved and no sample was removed.

In a separate 60.5-second window, an RP1 PIO state machine timed the same PPS input. Its period SD was 42.6 ns, against 831.2 ns for Linux IRQ timestamps of the same 60 periods; a repeat gave 43.9 ns and 448.5 ns. PIO periods are relative to the uncalibrated RP1 oscillator. These dependent, shared-source comparisons establish neither USB/GPIO parity nor absolute UTC or electrical accuracy.

W12 sampled sub-microsecond holdover

· W12 v0.2.33. Frozen clock models are evaluated against later authenticated BLE witnesses. P99 is taken across each run prefix’s worst sampled error.

Longest fully replayed device frontier

Sampled P99 <1 µs duration
64 s
Run prefixes
128
P99 / maximum sampled error
760 / 879 ns

This C-1 rolling population is fully reconstructed from accepted pairs. It is the longest replayed device-grid frontier, rather than the final value in every window.

Separate fixed mature cohort

Last passing / first failing age
75 / 76 s
Fixed run prefixes
128
P99 at the passing / failing age
983 / 1,028 ns

The one-second replay holds the cohort fixed. It does not replace the device’s rolling history. Later capture-end windows on both W12 units end at 24 seconds; failures, unknowns and unfinished runs remain in the evidence.

Four finite captures retain 25,639 accepted pairs and 10,450 numerically replayed START/END pairs. Their witness gaps longer than 250 ms number 212 / 338 / 292 / 213, with maxima of 1.688 / 1.375 / 2.188 / 1.000 s. Gap interiors are unmeasured. Temperature remains shadow-only; these are sampled BLE-relative errors, without a continuous bound, calibrated UTC/output or physical GM-transition qualification.

CoreSE: clock and shadow diagnostics

CoreSE 0.1.30 separately reports sampled P99 <1 µs HOLD ≥1.75 s across 128 trials, with one unknown. At a 103.4 s frozen-run age, its latest 128 BLE witnesses have P95/P99/MAX 909/936/948 ns. The 165 witness gaps longer than 250 ms leave their interiors unmeasured; the model age is not a continuous accuracy duration.

Explore the Precision Clock, screens and specifications

ESP32-S3 timing on W12

· Nordic source → W12 receiver / estimated relay → T096 witness. Two 300-second windows follow a 120-second warmup, with W12 0.2.10 and T096 0.15.169. Receive and relay adjustments were frozen before validation.

W12 · delivered clock at GNSS PPS

Relative-PPS RMS including bias
60.3 ns
Eligible / logged observations
522 / 600 · 87.0%
Mean / standard deviation
+40.7 / 44.5 ns
Absolute P95 / P99 / maximum
122 / 156 / 180 ns

All 78 source-unavailable records remain. The first window includes a source transition; the stable second window separately measured 38.3 ns RMS across 300/300 observations, with a 121 ns maximum. Across both windows, 466/522 eligible errors are strictly below 100 ns and all 522 are below 200 ns.

After the W12 relay · downstream T096 PPS

Relative-PPS RMS including bias
61.9 ns
Eligible / logged observations
451 / 600 · 75.2%
Mean / standard deviation
−5.5 / 61.7 ns
Absolute P99 / maximum
154 / 211 ns

109 no-map, 36 expired and four reference-ineligible records remain unscorable. One eligible error reaches or exceeds 200 ns.

Sampled transmitted time · measured at downstream RX

Received-output RMS including bias
108.3 ns
Scored / selected transmissions
58 / 120 · 48.3%
Mean / standard deviation
+13.9 / 107.4 ns
Absolute P99 / maximum
226 / 226 ns

61 selected transmissions lack a matching follow-up record; one lacks PPS brackets. Three scored errors reach or exceed 200 ns. This sampled path includes the existing downstream receive correction and does not isolate RF transmit delay.

The UTC-basis repair preserves nanosecond phase: all 522 eligible W12 records have identical relative and UTC fractional phases. The separate GNSS predictor measured 35.2 ns raw / 41.9 ns mapped RMS across 600/600 predictions; it evaluates its own prior model and does not measure delivered BLE accuracy.

The full timing gate remains open because availability gaps, downstream tails and unknown GNSS, GPIO/cable, RF and absolute UTC bounds remain. Physical W12 output timing and GM operation are unqualified. These measurements do not certify other units or profiles.

W12 PPS samples (CSV) · Downstream PPS (CSV) · All selected TX (CSV) · GNSS predictions (CSV) · Full results (JSON) · Provenance and offline verifier

Understand the implementation

Continuous hardware capture, clock bridges and causal fitting are explained in the seven-page low-level timing guide, with counter arithmetic, pseudocode and a validation sequence.

Read the ESP32-S3 implementation summary Read the low-level timing guide (PDF, 7 pages) LaTeX source and implementation references

Precise time through a T096 intermediary

A T096 relay disciplined its own clock from the GM, then generated a new timing train for a second T096 receiver. At that receiver, a separately captured local PPS tested the clock map that was already available at the edge.

  1. T096 GMSource reference
  2. T096 relayRegenerated timing
  3. T096 receiverDelivered clockLocal PPS → shadow query
    Validation only
One intermediary · two BLE links. The receiver’s PPS witness measures the BLE clock and never trains it.
RMS including bias
62.504 ns
Signed mean / demeaned jitter
−18.417 / 59.729 ns
Absolute P95 / P99 / maximum
123.25 / 160.25 / 191 ns
Usable / nominal boundaries
436 / 602 · 72.425%
Observed PPS edges / capture span
504 / 601.576 s

12 October 2026 · native 0.14.43. This is fractional-second phase at the receiver, with separately trained per-link calibration frozen beforehand. It does not establish calibrated absolute UTC, an output-pin error bound, continuous availability or physical two-relay precision.

Observation gaps and calibration

The full accounting is 602 = 436 valid causal comparisons + 65 invalid-reference observations + 3 reference-valid edges with expired maps + 98 conservative unobserved opportunities. The largest observed witness gap was 59.017 s. Invalid-reference excursions remain a separate population, and missing intervals are not interpolated.

A +10,543 ns receive correction per BLE link was trained separately and frozen before this run. The shadow PPS witness supplies no feedback or GNSS fallback to the BLE clock. Sampled counters separately retained six relay and eleven leaf map expiry/recovery episodes.

USB time transfer to a macOS user clock

The fixed concurrent 180-second shared-bus window compares a BLE-fed RC and a direct-GNSS DC USB path. RC/DC causal prediction RMS is 190.050/267.448 ns across 5,054/5,370 predictions. Same-instant shared-source agreement is 325.545/344.546 ns across 145/119 comparisons.

Delivered status samples are 147/177 for RC and 119/177 for DC, with 30/58 withheld. The paths use separate GNSS chains; the bus was shared and extended trials were incomplete at this cutoff. Internal tracking estimates, causal prediction and shared-source agreement remain different metrics. Absolute UTC and continuous delivery are unqualified.

The separate 15 October five-minute result remains available with its 23–339 ns settled estimator RMS, 414 ns agreement across 182/236 comparisons and 328 ns causal prediction RMS. Its startup and withheld samples remain with that dated experiment.

Earlier direct results by device and reference

· These three original configurations retain their own evidence cutoff. In particular, the earlier ESP32-S3 internal prediction is a different measurement from the current delivered W12 PPS result.

Nordic nRF52840 · Warm forward validation

Uncentered RMS
70.197 ns
Scoreable / physical observations
600 / 600
Mean absolute error
57.263 ns
Absolute P99 / maximum
158 / 192 ns

PPS-shadow query of the already-published BLE clock at a separately captured reference edge on the same local hardware counter. Common 10,516 ns edge correction trained separately and frozen before validation; no receiver-GPS training, validation refit, centering, or outlier deletion.

Nordic nRF52840 · Warm reverse validation

Uncentered RMS
76.359 ns
Scoreable / physical observations
598 / 599
Mean absolute error
60.376 ns
Absolute P99 / maximum
219 / 257 ns

PPS-shadow query of the already-published BLE clock at a separately captured reference edge on the same local hardware counter. The same separately trained, frozen 10,516 ns edge correction; one unscoreable causal query retained in the denominator.

Espressif ESP32-S3 · Normal 600-second run

Internal published-prediction RMS
319.5 ns
Scoreable / physical observations
359 / 359
Mean absolute error
244.5 ns
Absolute P99 / maximum
930 / 1010 ns

Internal causal prediction evaluated at physical target events using the previously published clock model. Controller-to-timer mapping.

Fine / coarse / unavailable clock views: 1,123 / 18 / 57 of 1,198 sampled views. Accuracy at scored events does not establish continuous clock availability.

Read the measurement summary (JSON)

How to read a timing result

What does PPS-shadow measure?

A PPS-shadow query evaluates the already-published BLE clock at a separately captured pulse-per-second reference edge on the same local hardware counter. The reference validates the BLE clock and does not train it. T096 RMS includes bias; the validation data were not centered or refitted.

Which clock is being tested?

The current W12 baseline measures delivered BLE time at GNSS PPS edges. Downstream PPS, sampled transmitted time and GNSS prediction remain separate. The earlier W12 result measures internal published predictions. A macOS estimator and a comparison against its own source answer different questions again.

Why keep unavailable observations?

Error statistics describe the observations that could be scored. Coverage records how often that comparison was possible. Missing records and expired maps are retained, so a small RMS is never a claim of uninterrupted precision.

What remains unqualified

  • RMS describes the stated population; it is not a worst-case bound.
  • These measurements do not establish independently calibrated absolute UTC accuracy or independent electrical GPIO output accuracy.
  • W12 retains 78/600 unavailable PPS records; downstream retains 149/600 unscorable records and a 211 ns tail. Sampled TX scores 58/120 with a 226 ns maximum. The full timing gate remains open.
  • Holdover durations describe sampled BLE-relative errors in declared populations; gap interiors, continuous accuracy, temperature compensation and physical GM-transition qualification remain unmeasured.
  • Results apply to the named tested configurations and profiles, not every BLE MCU or multi-hop mesh path.
  • Cross-study RMS, MAE, standard deviation, and GPIO means use different references and conditions; they are not a matched output-accuracy or energy benchmark.

PToB is an experimental protocol proposal, not an adopted Bluetooth SIG or IEEE standard. Indoor positioning, gPTP/TSN integration, medical validation, broadcast genlock, continuous holdover and physical power qualification require their own evidence.

Earlier work & public sources

Earlier editions remain available at their original URLs. Use the dated population that supports your citation; a newer paper does not change an earlier experiment.

Earlier viability paper

16 October 2026 · PDF

The original 15-page system synthesis, before the refreshed 30 October preprint.