Empirical Sub-Microsecond Holdover on BLE-Trained Software Clocks
How long can a frozen BLE-trained software clock retain sampled sub-microsecond P99 error?
Overview
This MCXO-inspired W12 study freezes a software clock model and evaluates its worst sampled error in each run prefix against later authenticated BLE witnesses. The longest fully replayed rolling device frontier reaches 64 seconds across 128 prefixes, with P99 760 ns and maximum 879 ns.
What this paper covers
- Distinguish a rolling device frontier from a fixed mature cohort and a final capture-end window.
- Read P99 across per-run worst sampled errors with failures, unknowns, censoring and witness gaps retained.
- Reproduce the numeric prefixes from compact public evidence without accessing hardware.
Research scope
A separate fixed mature 128-run cohort passes 75 seconds at P99 983 ns and fails at 76 seconds. Later rolling capture-end windows fall to 24 seconds. These populations are not interchangeable. Temperature is shadow-only; sampled BLE-relative error does not establish continuous accuracy, calibrated UTC, physical output or GM-transition acceptance.
Read it in context
The approved seven-page draft measures W12 v0.2.33. Its schematic audits v0.2.34 separately. CoreSE has its own shadow instrument and a separate ≥1.75-second sampled result; W12’s duration cannot be transferred to that target.
Cite this edition
Jonathan Yantis. Empirical Sub-Microsecond Holdover on BLE-Trained Software Clocks. 30 October 2026. Approved research draft. Original PDF.
Summary permalink: https://ptob.dev/research/empirical-ble-holdover/
Jonathan Yantis leads protocol design, hardware development and research. GPT 6.1 contributes implementation, analysis and documentation. This page summarizes the linked edition; retain its original evidence dates and measurement references.