Changelog

Changelog

All notable changes to this project are documented here. The format follows Keep a Changelog and the project uses Semantic Versioning:

• MAJOR: incompatible changes to the Python API, CLI options or file formats written;
• MINOR: new statistics, formats, commands or UI features (backwards compatible);
• PATCH: bug fixes and documentation.

Releases are made from vX.Y.Z tags (see CONTRIBUTING.md). The Release workflow checks the tag against the package version, tests, publishes a GitHub Release with the notes below and uploads to PyPI.

Unreleased

[2.16.0] - 2026-10-02

Research bench v2, part 1 (#29). The new modules are provisional (not yet in ntpstats.api) until 2.17 adds the simulator exchange formats.

Added

• Delay traces (ntpstats.trace, ntpstats trace): per-direction one-way delays extracted from any two-way exchanges.
  • Inputs: NTP and PTP captures, chrony measurements.log, ntpd peerstats/rawstats, and simulator output.
  • The clock offset between the two ends is removed (--detrend floor|linear|none).
  • The asymmetry is an explicit assumption (--asymmetry), since two-way timestamps cannot measure it.
  • ntpstats trace reports floors, PDV percentiles, loss and the correlation of the two directions; --csv exports the trace.
• Trace replay in the bench: ntpstats bench trace:FILE, or a [trace] table in a scenario file.
  • Modes: replay plays the trace in order; bootstrap draws random blocks, keeping the short-term correlation and both directions together.
  • scale multiplies the queueing.
  • Checked by a round trip: delays extracted from a simulated network and replayed give the same estimator scores as the model they came from.
• PTP servos and boundary-clock chains (ntpstats.ptpsim, ntpstats chain): a grandmaster and N boundary clocks, simulated with ground truth.
  • Each node has its own oscillator, the E2E delay mechanism with a moving-median filter, and linuxptp's PI servo (with its default gains) or an adaptive linear-regression servo.
  • Links have asymmetry, timestamp noise, and modelled or replayed PDV.
  • Time-error metrics are given per node and per hop: max|TE|, |cTE|, dTE_L MTIE, dTE_H.
  • Each hop is checked against the commonly quoted G.8273.2 T-BC class limits (A/B/C) or your own limits, and the end of the chain against a budget (default 1.1 µs); exit code 3 on failure.
  • --kp/--ki tune the servo. The warm-up follows the servo's lock time, and the command warns when a run is too short to lock.
• New notebook: a PTP chain against a time-error budget. With linuxptp's default gains, dynamic time error grows much faster than the number of hops (gain peaking): about 50 ns at 10 hops, about 600 ns at 20. A narrower loop keeps 20 hops near 25 ns. The bench notebook now also replays a real log.

Fixed

• Noise identification no longer divides by zero on perfectly alternating or noise-free data (lag-1 autocorrelation of −1, the white-PM limit).
• Commands piped into head and similar tools stop quietly instead of printing a BrokenPipeError traceback.

[2.15.0] - 2026-10-01

Added

• Stable API (#32): from ntpstats import api as nt gives the public surface in one namespace (loading, stability, analysis, metrology, time error and assurance, estimators, simulator, bench, reports, plugin types).

  • Covered by a deprecation policy: one full minor release of NtpstatsDeprecationWarning before a stable name or parameter changes. The warning is a FutureWarning, so it is shown in scripts and notebooks.
  • Helpers for contributors in ntpstats.deprecation: deprecated, renamed_parameter, moved.
  • The surface is frozen in tests/data/api_surface.json; CI fails on incompatible changes.
  • Documented on the new Stable API page; the policy is in CONTRIBUTING.

• Large files (#32, continues #13): load_large and iter_chunks read line-oriented logs in blocks, gzip included, and give the same series as load.

  • In a 300 000-line test, peak memory was about 2.4 times lower and reading twice as fast.
  • load switches to block reading by itself for text logs above 256 MB.
  • Supported formats: ntpd/NTPsec stats, chrony logs, linuxptp, CSV and the 2012 log.

• Notebooks and reproduction gallery (#32, #38): four notebooks, executed in CI on every change:

  • a chrony log to stability with intervals and a noise model;
  • benchmarking your own estimator;
  • compliance evidence (PTP time error, a UTC bound, a mask, an HTML report);
  • a reproduction of the NIST SP 1065 test suites.

  A docs page explains how to contribute a gallery entry.

With these, #32 is complete. Its last step, a 3.0 changelog that lists every change to the stable API, is part of the 3.0 release checklist (CONTRIBUTING). The research bench v2 (#29) is planned for 2.16 and 2.17.

[2.14.0] - 2026-10-01

Added

• In-browser edition (#31): the web UI runs entirely in the browser on the docs site, with Pyodide (WebAssembly). Files never leave the machine. docs/build_app.py builds it from the package wheel; a CI job checks with Playwright and Chromium that the page gives the same numbers as the installed package. Live probes stay in the installed tool.
• Plugins (#37): other packages add parsers, estimators, event detectors, masks and import profiles through entry points (ntpstats.parsers, .estimators, .detectors, .masks, .profiles).
  • ntpstats plugins lists everything, with where it comes from and load errors.
  • A plugin that fails to load is reported and never breaks the tool.
  • Contract tests for plugin authors: pytest --pyargs ntpstats.testing.plugin_contract.
  • Worked example in examples/plugins/ntpstats-toy-csv, installed and checked in CI.
  • The built-in research formats are registered through the same API.
• Dataframes and Parquet (part of #32): TimeSeries.to_pandas() / from_pandas(), StabilityResult.to_dataframe(), DynamicResult.to_xarray(), and Parquet files that keep float64 precision and the metadata (convert --to parquet; Parquet input is auto-detected). Optional extra ntpstats[data]; the package ships a py.typed marker.

Changed

• The web server's /api routing is a plain dispatch() function, shared by the HTTP server and the browser edition. The UI's format list comes from /api/info, so plugin formats appear.

[2.13.0] - 2026-10-01

Added

• Open interop dataset (#30): the first Live interop run of each month writes one JSON record per probe (NTP, interleaved, NTS, NTS pool, NTPv5, Roughtime) to data/interop/YYYY/YYYY-MM-DD.jsonl, committed with [skip ci]. The dataset ships with each release, so it is archived on Zenodo.
  • Schema: data/interop/README.md.
  • New interop input format: one series per test and server, with availability.
  • ntpstats dataset: summary per server.
  • A directory given to any command is read as the concatenation of its files.
• OpenTelemetry export (#36): monitor/watch --otlp [URL] push the metrics as OTLP/HTTP JSON, standard library only; the standard OTEL_* environment variables are honoured. The payload is checked against the official OpenTelemetry protobuf definitions.
• Cross-validation (#28): the exact EDF is compared with the SP 1065 table 5 approximations used by Stable32: within 3 % for white PM/FM, 7 % for flicker/random-walk FM, 16 % for flicker PM.

Fixed

• Roughtime: the client also offers draft-08 (0x80000008), and falls back to Google-Roughtime (the pre-IETF protocol: unframed, 64-byte nonces, microseconds) when a server does not answer the IETF request. Cloudflare's server did not answer in the 2.11 live interop run; the protocol that answered is reported and recorded in the dataset.

[2.12.0] - 2026-10-01

Added

• Power-law noise model (#25): ntpstats noise and ntpstats.noisefit fit h₂…h₋₂ and, optionally, a linear drift to OADEV/MDEV/HDEV curves and the spectrum.
  • The expected estimator values are exact for the discrete Kasdin–Walter model.
  • Model selection uses BIC.
  • Intervals come from a parametric bootstrap. Monte Carlo coverage is ≥ 90 % per coefficient except at the edge of detection.
  • Corner τ values are reported. --scenario writes a simulator clock ([clock.h_alpha]), and simulate --scenario runs it.
  • The UI can overlay the fitted model on the Stability chart.
• Spectra (#25): ntpstats spectrum and ntpstats.spectrum compute S_x(f) and S_y(f) with Welch or sine multitaper, per gap-free stretch, log-binned with degrees of freedom. L(f) in dBc/Hz for a carrier. New Spectrum tab in the UI.
• N-cornered hat (#27): ntpstats hat and ntpstats.hat estimate each source's stability from pairwise differences: Groslambert covariance, three-cornered hat and N-cornered hat by least squares, with intervals. Negative variances are flagged. Sources are aligned from separate logs.
• Holdover prediction (#26): ntpstats holdover and ntpstats.holdover predict TIE(t) with an envelope and the time to violate limits. The 95 % envelope holds 95 % ± 3 % with a known noise model, and 94–100 % with a fitted one.
  • The variance is computed exactly for the fitted noise, including the error of the fitted frequency and drift.
  • The uncertainty of the noise model is mixed into the envelope.
  • --backtest checks calibration on the log itself; --min-holdover exits with code 3 when the limit is not held long enough.
  • Phase can also be integrated from a frequency column (chrony).
  • New Holdover tab in the UI.
• Research data (#34): new formats cggtts (V2E, line checksums), rinex-clock (IGS .clk), circular-t (BIPM, UTC − UTC(k)), ripe-atlas (NTP results) and ntppool (monitor logs). ntpstats cv computes GNSS common-view and all-in-view time transfer, and research.group_summary gives per-server or per-probe distributions. Gzipped inputs are read directly.
• HTOT (#33): Hadamard total deviation (-k htot), bias-corrected per noise type, matching the NIST SP 1065 tables within 0.3 %. compat.allantools.htotdev returns the raw value.
• Docs: Metrology and Research data pages; synthetic examples of every new format in examples/data/.

[2.11.0] - 2026-09-30

Added

• Roughtime (#24): ntpstats roughtime and ntpstats.roughtime (draft-ietf-ntp-roughtime-19). It verifies the delegation and response signatures, MINT/MAXT, and the Merkle proof. It chains nonces over two rounds, runs the causal-ordering check, and writes and verifies malfeasance reports. It gives an authenticated interval for the local clock's error (--check-local), bundles the ecosystem server list and reads the draft's server-list JSON. It is compatible with older-draft servers (nonce leaf, microsecond timestamps).
• NTS pools (#24): NTS-KE NTP Server Deny records (draft-ietf-ntp-nts-keyexchange-pool-01), nts.pool_sessions() and ntpstats query --nts --pool N give independent servers from one pool.
• RFC 9769 interleaved mode (#24): ntpstats query --interleaved and sntp.query_interleaved() use the server's precise transmit timestamp. Captures recognise interleaved exchanges, recompute them with the precise timestamp and add an interleaved column.
• Live interop: NTS pool, Roughtime (chained, with the causal check) and interleaved-mode sections. The NTPv5 client stays on draft-09, the current draft.
• Docs: a Protocols page.

[2.10.0] - 2026-09-30

Added

• UTC traceability audit (#22): ntpstats audit FILE --limit 100us. It computes a per-sample bound (|offset| + path + upstream + reference), stating the rule applied to each term (log columns or user allowances). Windows (1 h) are judged pass/fail/insufficient; gaps count as unmonitored, never as compliant. Output is text, JSON or a self-contained HTML report with the tool version and SHA-256 hashes of the inputs, and the exit code is 3 on failure.
• Change detection (#23): ntpstats events and ntpstats.events. It finds phase steps and spikes (MAD against a rolling slope), frequency changes (binary segmentation, standardised CUSUM), delay-floor (route) changes with an asymmetry hint, and leap smears. Phase and frequency events are marked path changed/unchanged. The UI lists events under the Offset chart.
• GitHub Action (#36): uses: thiagodefreitas/NetworkTime@v2.10.0 runs stability, time error, audit, bounds or events checks, writes a job summary and fails on exit code 3. Arguments go through the environment, never through the shell. CI exercises it on the examples.
• pytest assertions (#36): ntpstats.testing (assert_max_te, assert_time_error_within, assert_stability_within, assert_audit_passes, assert_bounds_valid, assert_no_events) and a timing_log fixture registered as a pytest plugin.
• "Audit, events & CI checks" docs page.

Fixed

• format_seconds(0) printed "0 ps"; it now prints "0 s".

[2.9.0] - 2026-09-30

Added

• Live PTP clients (#21): ntpstats watch ptp4l polls linuxptp through pmc (CURRENT_DATA_SET, TIME_STATUS_NP), and ntpstats watch ptpcheck polls facebook/time ptpcheck stats. Both are negated into reference − local, offered in the UI's Live dialog, and served to Prometheus with --metrics-port.
• Windows w32tm (#21): /stripchart /dataonly text (with midnight rollover and the date taken from the header) and /rdtsc CSV with exact FILETIME timestamps, auto-detected.
• Prometheus import (#21, #36): ntpstats prom URL QUERY saves a range query (e.g. ntpd-rs ntp_source_offset_seconds, chrony_exporter) as JSON, and any command reads it.
• Clock-error bound validation (#21): ntpstats bounds BOUNDS REFERENCE classifies each ClockBound/fbclock/CSV window as inside, violated or indeterminate against a better reference (with its uncertainty), and reports the violation rate, worst excess and tightness. Exit code 3 above --max-violation-rate. Windows are parsed as exact decimals.
• Instrument import profiles (#35): -f profile:NAME|FILE.toml describes column layouts (value/time columns, units, TE or offset sign, ISO/MJD/unix/elapsed time, skipped banner rows, extra columns). Built-ins: te-csv, tic-ns, tic-s, iso-te-ns.
• --negate on every command, for sources that log local − reference.
• Examples: w32tm-stripchart.txt, clockbound.txt with a reference (synthetic), and a "Sources, instruments & bounds" docs page.

[2.8.0] - 2026-09-30

Added

• PTP from packet captures (#19). IEEE 1588 v2/v2.1 over UDP (IPv4/IPv6, ports 319/320) and Ethernet (0x88F7, VLAN) is read from pcap/pcapng. It supports one- and two-step clocks, E2E (Delay_Req/Resp) and P2P (peer delay, one- and two-step), and correction fields. The TAI→UTC offset is taken from Announce or inferred. There is one series per master/slave flow (offset against the capture clock, path/link delay, one-way delays, corrections), and AUTHENTICATION TLVs (NTS4PTP) are flagged. ntpstats.ptp.summary() counts messages.
• Time-error metrics (#20): ntpstats timeerror and ntpstats.timeerror. They give max|TE|, cTE (record and per 1000 s window), TEL/max|TEL| (0.1 Hz first-order filter, restarted after gaps, settling excluded), dTE_L (peak-to-peak, MTIE, TDEV) and dTE_H peak-to-peak. User-supplied limits (max_te, 30ns …) and MTIE/TDEV masks for dTE_L are checked, with exit code 3 on failure. Time-interval counter input is handled with --units ns --input-is-te.
• Time-error cards in the web UI (Overview) and ntpstats report --time-error.
• examples/data/ptp-capture.pcapng (synthetic) and a "PTP captures & time error" docs page.

Changed

• Capture timestamps are kept as integer nanoseconds; a float of POSIX seconds only resolves about 240 ns. NTP offsets from captures are now computed exactly as well.

Fixed

• Personal email addresses removed from public files; the author is referenced by GitHub profile.

[2.7.0] - 2026-09-30

Added

• Cross-validation against NIST SP 1065 (#28). The NBS Monograph 140 data (table 30) and the 1000-point test suite (table 31) are regenerated in the tests. ADEV, OADEV, MDEV, TDEV, HDEV, TOTDEV, MTOT and TTOT agree to the 7 printed digits.
• TTOT (time total deviation), and confidence intervals for MTOT/TTOT from the SP 1065 table 8 EDF.
• Stable32 data files (#33): read_stable32/write_stable32 in ntpstats.interop, -f stable32-phase|stable32-freq, and ntpstats convert (any supported log to a Stable32 file or CSV).
• allantools-compatible API (#33): from ntpstats.compat import allantools as at gives adev, oadev, mdev, tdev, hdev (non-overlapping), ohdev, totdev, mtotdev, ttotdev, theo1, mtie and tierms with allantools signatures.
• Prometheus/OpenMetrics exporter (#36): ntpstats monitor|watch --metrics-port. It serves offset, delay, an error bound (|offset| + delay/2 + root delay/2 + root dispersion), NTS status, rolling TDEV and stddev, and sample and error counters (errors by kind, e.g. kod_rate).
• Grafana dashboard, Prometheus alert rules and a docker-compose stack in contrib/, with a "Monitoring" docs page. Tests check that the dashboard and rules only use exported metrics.
• Docs: "Stable32, TimeLab, allantools" migration page and a tools landscape page.
• Community on-ramp (#38): Code of Conduct (Contributor Covenant 2.1), security policy, issue forms (bug, feature, sample-data contribution), a pull request template, and "ways to contribute" in CONTRIBUTING.
• Releases also publish a container image to ghcr.io/thiagodefreitas/ntpstats.

Changed

• MTOT is bias-corrected by default, dividing by the factor for the noise type identified at each τ (white PM 0.94 … random-walk FM 0.69), as Stable32 and the SP 1065 tables do. This makes MTOT an unbiased estimate of MVAR. --raw-mtot / bias_correction=False give the raw eq. (27) value, which is what earlier versions reported.

[2.6.0] - 2026-09-30

Added

• ntpstats stability --exact and compute(..., max_work=0) / series_stability(..., max_work=0) to force the full MTOT/Theo definitions.

Changed

• MTOT, Theo1, TheoBR and TheoH are much faster (#18). MTOT and Theo1 are vectorised across subsequences. TheoBR computes its bias ratio once per τ grid instead of once per τ. Above a work limit per τ (stability.MAX_WORK), subsequences are strided (stride ≤ m), and the TheoBR ratio averages 64 evenly spaced terms. What was sampled is reported in meta["stride"] and meta["theobr_ratio_terms"], and the CLI prints a note. On 1M samples at octave τ each statistic takes 2–4 s instead of hours; on 4096 samples TheoBR went from 210 s to under 1 s. Below the limit, results match the literal definitions to 1e-12.

2.5.0 - 2026-09-29

Added

• New stability statistics (#15, closes #2): MTOT (modified total deviation), TheoBR (bias-removed Theo1) and TheoH (hybrid ADEV/TheoBR), in the CLI, API and UI.
• Documentation site (MkDocs Material, mkdocs build --strict in CI) with getting started, CLI, formats, statistics, network and estimators, validation, and an API reference generated from docstrings (mkdocstrings). It is deployed to GitHub Pages from master.
• GitHub wiki generated from docs/, CHANGELOG.md and ROADMAP.md (docs/sync_wiki.py, Wiki workflow).
• Type checking: mypy in CI, and the package is now mypy-clean.
• Coverage XML report in CI.

Changed

• Releases are tag-driven (vX.Y.Z), or created from Run workflow with a tag input. The tag must match __version__ and have a CHANGELOG section. Wheels and sdists are checked with twine and published to PyPI through trusted publishing.
• docs extra: mkdocs>=1.5,<2, mkdocs-material, mkdocstrings[python].

Fixed

• Type errors found by mypy: optional defaults, NTS session key handling, and Optional values in report formatting.

2.4.0 - 2026-09-29

Security

• Web UI: export file names are sanitised and RFC 6266-encoded, which prevents response-header injection through crafted upload names; non-ASCII names no longer truncate responses.
• Web UI: a negative Content-Length is rejected instead of blocking a worker thread.
• Web UI: /api/mask treats the body as mask text only and can no longer read files on the server.
• NTS: a Kiss-o'-Death (including NTSN) is honoured only when it echoes the request's Unique Identifier (RFC 8915 §5.7), so a spoofed packet cannot flush the keys.

Fixed

• NTS: TLS failures during NTS-KE raise NTSError, which the CLI and monitor handle, instead of surfacing raw pyOpenSSL exceptions.
• Masks apply to one statistic, named in the header (tau,tdev, tau,mtie, …) or set with --mask-kind (default TDEV). They are no longer compared with estimators of other units.
• CSV export keeps full sub-second time resolution. It previously used %.12g, which kept only 2 decimals of POSIX time.
• CSV import with a header but no time column now requires tau0 instead of using the offset as time.
• ntpstats network removes a linear trend before computing offset statistics, as the UI and report do (--detrend).
• Kalman noise fitting no longer fails on constant offsets (zero ADEV).
• Floor packet percentage no longer drops the last sample when the span is an exact multiple of the window.
• ntpstats simulate no longer mutates the shared presets.
• --tau0 is honoured as the stability grid for multi-column files.
• ntpstats report --ci 0 no longer crashes.
• Live monitor: a late sample from a stopped monitor can no longer land in a newer live dataset.
• ui --host 0.0.0.0 accepts remote Host headers (it is an explicit opt-in). IPv6 bind addresses work.

Performance

• Stability confidence intervals: the EDF uses FFT convolution, so 1M-sample OADEV/MDEV/HDEV/TDEV with CIs takes about 1.2 s (previously 28–121 s).
• Parsers use numpy's C tokenizer, with a tolerant per-line fallback: ~1 s per million loopstats lines, ~2.4 s peerstats and ~3 s chrony (previously ~9–10 s). Closes #13.

Added

• ntpstats compare ESTIMATE REFERENCE: error statistics plus TDEV/MTIE of one source against a reference (e.g. NTP client vs PPS/GNSS).
• CI: ruff lint and coverage reporting; Dockerfile for the UI/CLI/monitor.
• Tests for every fix above (166 tests).

2.3.0 - 2026-09-29

Added

• Research bench: pluggable estimator API (ntpstats.estimators, entry-point group ntpstats.estimators) and ntpstats bench. It runs scenarios × seeds × estimators against ground truth and reports RMS, bias, p95, max, MTIE(1 h) of the error and runtime, with a 30 min warm-up excluded by default. Output is a table, CSV, JSON or a self-contained HTML report (#9).
• Scenario files in TOML/JSON (scenario_from_dict) and two examples in examples/scenarios/ (#9).
• Reference estimators: chrony-style weighted regression with a runs-test window, RADclock-style feed-forward, RFC 5905 clock filter + selection (intersection), cluster and combine for multiple servers with causal frequency propagation, and a median combiner (#10).
• Simulator: flicker FM/PM, temperature-driven frequency (tempco), path events (route change, congestion, outage), multi-server scenarios with falsetickers and stepping servers (simulate_multi), and new presets route-change and falseticker (#11).
• Self-contained HTML reports: ntpstats report FILES -o report.html and a Report button in the UI. Offset, stability with CI bands and a noise table, network wedge, input SHA-256 hashes and parameters are embedded, and the file works offline (#12).
• Live interop workflow: weekly/on-demand queries of public NTPv4 and NTS servers plus the NTPv5 probe from GitHub runners, with a first snapshot in docs/INTEROP.md (all five NTS servers tested authenticated successfully).
• examples/04_custom_estimator.py: plugging your own algorithm into the bench.

Changed

• GitHub Actions updated to actions/checkout@v5 and actions/setup-python@v6 (Node 24).

2.2.0 - 2026-09-29

Fixed

• chrony measurements.log offsets had the wrong sign in 2.0.0–2.1.0. chrony.conf(5) documents θ as "positive indicates that the local clock is slow of the remote source" (the ntpd convention), but the parser negated it. Example data and the cross-format test repeated the same assumption; they are corrected, and a regression test is now pinned to the documentation line.

Added

• PTP: linuxptp ptp4l/phc2sys/ts2phc parser, covering per-sample and summary lines, stdout/syslog/journald, monotonic→UTC mapping, and path delay feeding the network metrics (#5).
• chrony refclocks.log parser (GNSS/PPS reference clocks).
• Live local daemons: ntpstats watch chrony|ntpd (chronyc -c tracking, ntpq -c rv), also available as a UI Live source (#6).
• pcap/pcapng import with stdlib-only dissection (Ethernet/VLAN, Linux cooked, raw IP, IPv4/IPv6). Requests and responses are matched by origin timestamp or v5 client cookie, and offsets are measured from the capture clock (#7).
• NTS client (RFC 8915) behind the optional extra ntpstats[nts]: NTS-KE over TLS 1.3 with ALPN, host-name verification, AES-SIV-CMAC-256 and cookie renewal. query --nts, monitor --nts (#8).
• Experimental NTPv5 client (draft-ietf-ntp-ntpv5-09, wire format cross-checked with ntpd-rs): query --ntpv5, --probe-v5 upgrade negotiation, monitor --ntpv5, UI protocol selector (#8).
• Example logs: linuxptp, chrony refclocks and a pcap capture.

2.1.0 - 2026-09-29

Added

• Exact EDF for ADEV/OADEV/MDEV/TDEV/HDEV confidence intervals (ntpstats.edf): computed from the estimator filter convolved with the discrete power-law noise model, verified by Monte Carlo for every noise type (#1).
• New statistics: TOTDEV (with NIST EDF), Theo1 (τ up to 0.75 × record) and TIErms (#2).
• Dynamic stability (sliding-window ADEV/MDEV/…): ntpstats dynamic, stability.dynamic() and a heat-map in the UI (#3).
• Limit masks from user CSV (tau,limit, log-log interpolation): ntpstats stability --mask FILE (exit code 3 on failure, upper confidence bound used), per-τ margin in the UI and a PASS/FAIL chip (#4).
• UI: estimator chips are populated from the server; mask loading; dynamic view toggle.

Changed

• Confidence intervals now use the exact discrete EDF. The Greenhall–Riley continuous-time approximation over-estimated the EDF by up to ~20 % at small averaging factors for instantaneously sampled phase data. stability.edf_approx is kept as a deprecated alias.

2.0.0 - 2026-09-29

First release of the rewrite. The 2012 Google Summer of Code prototype is kept unchanged in legacy/.

Added

• Parsers for ntpd/NTPsec loopstats, peerstats, rawstats; chrony tracking.log, measurements.log, statistics.log; CSV; 2012 estimators.log. Auto-detection and sign normalisation (server − local).
• Stability statistics: ADEV, OADEV, MDEV, TDEV, HDEV, MTIE with χ² confidence intervals, lag-1 ACF noise identification and gap-aware resampling.
• Network metrics: delay floor, offset/delay wedge, asymmetry indicator, floor packet percentage, min-delay clock filter.
• Kalman filter (delay-weighted, gated, irregular sampling) and RTS smoother.
• Simulator with ground truth and ntpstats simulate --benchmark.
• SNTP client (random transmit timestamp, origin check, KoD, 2036 era) and ntpstats monitor.
• Local web UI (ntpstats ui), CLI, optional matplotlib report figures.
• Test suite, CI, examples, documentation, MIT license.

Fixed (relative to the 2012 prototype)

• Allan deviation τ₀ was hard-coded to 32 s (≈33× error on loopstats data); irregular sampling and gaps were ignored. See docs/STATE_OF_THE_ART.md.