Changelog

Newest first. Entries marked «breaking» changed an interface or a default; while the project is on 0.x those land on minor bumps, which is why pinning matters.

0.7.2 — 24 September 2026#

• Fixed: quality.isolation_distance returned a finite number when the cluster held more than half of all events, where the quantity is undefined. It now returns nan and the quality table prints a dash.
• Fixed: filt.common_median ignored bad_channels from the sidecar when groups was also passed explicitly, so a dead channel could enter the median. Any session referenced that way should be re-run.
• Detection is about 12 % faster: the chunk overlap no longer copies the raw block before filtering it.
• Docs: the methods page gained the trace figure and the section on where alignment fails.

0.7.1 — 12 August 2026#

• Fixed: containers written by 0.7.0 on Windows stored backslashes in the manifest and could not be read anywhere else. Readers now accept both, and the writer normalises.
• Added axonforge selftest, which builds a four-channel synthetic recording in memory and checks the whole path without touching a file.
• io.sanity gained the comb-spectrum score that catches a wrong n_channels on interleaved binary input.

0.7.0 — 30 June 2026#

• Breaking: the keyword thresh_k on detect.threshold is now k. The old name warns and will be removed in 0.8.0.
• Breaking: default restarts in sort.cluster raised from 10 to 20. Clustering is slower by roughly the same factor and noticeably more stable at low signal-to-noise; the figures on the validation page are from the new default.
• NumPy 2.x is supported and tested alongside 1.26.
• Added waveform.align(mode="centre_of_mass") for waveforms whose peak is clipped by the acquisition range.
• Container format v3: provenance is a dictionary per stage rather than one flat string. Version 2 files still read.

0.6.1 — 15 April 2026#

• Fixed: the zero-phase filter produced edge artefacts at chunk boundaries that looked like small spikes roughly every chunk_s seconds. Chunks now overlap by twice the filter length and the overlap is discarded after filtering. This was the worst bug in the project's history: it created events that no neuron produced, at a predictable interval, in every chunked run since 0.3.0.
• Fixed: quality.presence_ratio counted the final partial bin as empty.

0.6.0 — 26 February 2026#

• Breaking: Python 3.9 is no longer supported.
• Breaking: the quality metrics moved out of sort into their own quality module. The old import path warns.
• Added detect.dual_threshold for recordings with slow amplitude changes.
• Added stats.bursts, interval-threshold burst detection.
• Docs: first published methods page.

0.5.0 — 7 November 2025#

• Breaking: UnitSet.save writes an .afz container instead of a pickled archive. Loading old archives was dropped rather than maintained, because the pickle contained class references that we could not promise to keep stable.
• Added the report subcommand: raster, interval histograms and mean waveforms into one document.
• Added stats.cv2 and stats.fano.

0.4.0 — 22 July 2025#

• Clustering replaced: Gaussian mixtures with model order by information criterion, instead of k-means with a k the user had to guess.
• Added quality.l_ratio and quality.isolation_distance.
• The synthetic benchmark is shipped in the source tarball, so published numbers can be reproduced locally.

0.3.0 — 18 March 2025#

• HDF5 reader behind the hdf5 extra.
• Channel groups, and common median referencing that respects them.
• Sub-sample parabolic interpolation in peak alignment. Median timing error on the benchmark fell from 51 µs to 28 µs at a signal-to-noise ratio of 6.

0.2.0 — 5 November 2024#

• Waveform extraction with windows in milliseconds, and alignment on the negative peak.
• Principal-component features and k-means clustering.
• Interspike interval histogram and coefficient of variation.

0.1.0 — 12 June 2024#

• First public release: interleaved binary reader, Butterworth bandpass, threshold detection with dead time, and nothing else.