Skip to content

Releases

Offline QC v0.1.3 — readable source warnings on desktop and phone

Choose a tag to compare

@Jacob-Met Jacob-Met released this 10 Nov 18:46
2d68bda

CaptureSuite Offline Analysis v0.1.3 (experimental)

Purpose: inspect sealed multimodal recordings and identify unavailable IMU orientations by sensor. All-zero quaternions produce source WARN with exact affected sensor sample counts. Finite acceleration/gyro decoding remains available. Existing analysis commands remain included.

Tested: Linux x86_64 Ubuntu24.04-era glibc. Python is bundled. Windows/macOS, acquisition desktop, and physical hardware are not validated by this release.

Extract capturesuite-analysis-linux-x86_64.tar.gz, then:

./capturesuite-analysis/capturesuite-analysis --version
./capturesuite-analysis/capturesuite-analysis qc /path/to/session.mmsession --inspect-payloads --strict-warnings

Open SESSION/processing/jobs/JOB/reports/qc.html. A completed strict-warning job exits2; verify printed completed_with_warnings and report existence. A missing command also can exit2 and is not success. Without strict mode a completed warning job exits0. Raw source files remain unchanged.

Exact source2d68bdae056113f488153cbe699ffea7ba4c6ad5, full source/archive and GPL3 license included. Built with maintained scripts/build-analysis-linux.sh, Python3.12 and PyInstaller6.22.3. SHA256SUMS covers assets.

Verification: base warning full analysis360passed/1skipped/28subtests, sensor follow-up28targeted passed, cumulative410passed/1skipped/28subtests/Ruff clean. Actual standalone entry executed on sealed synthetic protobuf MCAP and reported missing:1 while valid sensor stayed unaffected. Screenshot is actual generated HTML. These are protocol fixtures, not physical sensor accuracy validation. Arbitrary nonunit quaternion thresholds and hardware quality diagnosis are outside scope.

Warnings now appear immediately after source status. Independent Chrome390x844 and1280x1000 review confirms full readable warning on first screen and no page horizontal overflow.46targeted tests pass. Desktop/phone screenshot assets are actual judged reports. Cumulative maintained package includes third-party notices.

Offline QC v0.1.2 — affected IMU sensor identification

Choose a tag to compare

@Jacob-Met Jacob-Met released this 10 Nov 18:35
7bc4ecc

CaptureSuite Offline Analysis v0.1.2 (experimental)

Purpose: inspect sealed multimodal recordings and identify unavailable IMU orientations by sensor. All-zero quaternions produce source WARN with exact affected sensor sample counts. Finite acceleration/gyro decoding remains available. Existing analysis commands remain included.

Tested: Linux x86_64 Ubuntu24.04-era glibc. Python is bundled. Windows/macOS, acquisition desktop, and physical hardware are not validated by this release.

Extract capturesuite-analysis-linux-x86_64.tar.gz, then:

./capturesuite-analysis/capturesuite-analysis --version
./capturesuite-analysis/capturesuite-analysis qc /path/to/session.mmsession --inspect-payloads --strict-warnings

Open SESSION/processing/jobs/JOB/reports/qc.html. A completed strict-warning job exits2; verify printed completed_with_warnings and report existence. A missing command also can exit2 and is not success. Without strict mode a completed warning job exits0. Raw source files remain unchanged.

Exact source7bc4ecccc9f8d0798371d9e3418016c5b544b6eb, full source/archive and GPL3 license included. Built with maintained scripts/build-analysis-linux.sh, Python3.12 and PyInstaller6.22.3. SHA256SUMS covers assets.

Verification: base warning full analysis360passed/1skipped/28subtests, sensor follow-up28targeted passed, cumulative410passed/1skipped/28subtests/Ruff clean. Actual standalone entry executed on sealed synthetic protobuf MCAP and reported missing:1 while valid sensor stayed unaffected. Screenshot is actual generated HTML. These are protocol fixtures, not physical sensor accuracy validation. Arbitrary nonunit quaternion thresholds and hardware quality diagnosis are outside scope.

Offline Analysis v0.1.4 — retained jobs and exact time-region exports

Choose a tag to compare

@Jacob-Met Jacob-Met released this 10 Nov 19:32
b5b979d

Experimental Linux x86_64 offline workbench, exact source b5b979d (PR180 and PR184 merged).

Researchers can select an exact recorded time interval from retained measurement tables without repeating analysis. export-table JOB --list reveals saved columns and units; add --time-column t_start_ns --start-ns 1500000000 --end-ns 2500000000 with the selected table/output. Start is inclusive and end exclusive. The selection retains existing rows; feature windows are never clipped, recalculated or gap-filled. Integer nanoseconds retain full precision. Source warnings, typed columns and selected/source row counts accompany CSV/Parquet exports.

This cumulative source also includes main's recorded-session plot timing/gap fixes, per-sensor IMU support fields, QC numeric sample excerpts, and readable source warnings. Those have their own taskboard evidence; this release's focused receiving covers retained table export.

Launch: verify sha256sum -c SHA256SUMS, extract the Linux archive, then run ./capturesuite-analysis/capturesuite-analysis --version and ./capturesuite-analysis/capturesuite-analysis --help. No Python install is needed. Built on Ubuntu24.04-era glibc/Python3.12 using the maintained build script on Renpō fleet394, without GitHub Actions. Compatible Linux x86_64/glibc runtime required. The archive includes exact source marker, launch/feature guides and bundled dependency notices; corresponding project source is attached.

Validation:69 retained-history/export and genuine numeric CLI receiving tests passed; Ruff/diff checks passed. Independent Kestrel used the native candidate against a real protobuf MCAP-derived job created by the older official executable, relocated with raw sources removed. Exact two rows at1.5s/2s were retained,2.5s start excluded, original window ends2.5s/3s preserved, and filter-column provenance survived omitted-column export. Final merged executable executed the same retained row workflow successfully. recorded-time-export.cast is an actual PTY capture of the merged executable, replayable with asciinema; the UTF-8 transcript is also attached. Inputs contain generated measurements, not participant recordings.

Limits: command-line offline analysis only; no native desktop, Windows/macOS executable, live acquisition, sensor accuracy or clinical validation. Region selection still requires saved paths and integer nanoseconds; there is no visual region selector. Empty selections produce typed zero-row output. Multi-hour performance and real instrument timing remain unverified. All measurements in receiving are synthetic values encoded in genuine MCAP.

Saved work discovery: capturesuite-analysis jobs PACKAGE --columns now shows recorded job status, warnings/errors, source/stream tables, row counts and units, and a quoted follow-up using the actual executable path. Kestrel independently copied this command verbatim from a relocated package with spaces and unavailable raw sources, receiving the correct3 tables. Corrupt metadata remains visible; this is metadata discovery, not new raw integrity validation. Current cumulative source also includes merged PR173 combined decoded-payload analysis.

Offline analysis 0.1.5 — interpretable measurements

Choose a tag to compare

Offline analysis 0.1.5 — interpretable measurement exports

This experimental Linux build makes a selected numeric table understandable after it leaves CaptureSuite. It also prevents a mistyped source selection from replacing useful saved analysis with an empty successful job.

What changed

  • Keep the meaning with the measurements. Open the exported README.html to see the session/source, selected units, warnings, averaging method, requested and rounded window settings, and the support behind the values.
  • Keep each row's original support. Numeric jobs now retain total and used sample counts. An aligned row_context.csv carries the actual window interval and counts even when those columns were omitted from the selected table. Column order and measurement values remain the requested ones.
  • Recover from source-ID mistakes. qc, features, plots and all reject an unknown or partly unknown selection before creating or replacing a job. The error lists the exact available IDs.
  • Use consistent instructions. The bundled guides cover xz extraction, selected-row export, the readable handoff and source-selection recovery.

A requested one-second window at 7.5 Hz rounds to eight retained slots. Its actual elapsed interval can differ when timestamps have a jump. A row-start filter selects whole existing measurements; it does not crop their windows. The handoff makes those distinctions visible. Older jobs explicitly keep unknown method details instead of acquiring guessed defaults.

Download and run

Download the archive and its matching .sha256 file:

sha256sum -c capturesuite-analysis-linux-x86_64-4bc3552.tar.xz.sha256
tar -xf capturesuite-analysis-linux-x86_64-4bc3552.tar.xz
cd capturesuite-analysis
./capturesuite-analysis --version
./capturesuite-analysis qc "/path/recording.mmsession" --inspect-payloads

The archive is 97,709,740 bytes. SHA-256:
6a55b1863ccc2fdd58cc06e32f9d34ceca4f71c787308a9dbb2860652f46a8e8

To work with measurements:

./capturesuite-analysis features "/path/recording.mmsession"
./capturesuite-analysis jobs "/path/recording.mmsession" --columns
./capturesuite-analysis export-table "/path/recording.mmsession/processing/jobs/JOB_ID" --list

Use the job and table paths printed by those commands. The bundled FEATURE_TABLE_EXPORT.md explains selected columns and optional recorded-time bounds. Open the resulting README.html beside its CSV/Parquet files before interpreting the values.

See the handoff

Watch the 20-second actual screen recording, silent

The recording uses controlled numerical readings encoded in genuine MCAP. It shows the local handoff's identity and warnings, selection semantics, rounded window settings and row support.

Desktop measurement handoff

Open the actual phone view

Source and evidence

Exact source: 4bc3552f1c673c69de3540678d7dcbccec17e0c3, merged through PR203. Main feature changes are PR195 and PR201. Both SOURCE_COMMIT and --version report the source in this bundle.

astra-cedar-61b7 built the maintained standalone executable and ran the analysis suite on this source: 506 tests and 28 subtests passed, with two skips. astra-sable-61b7 separately judged the numeric handoff's values, row support, output hashes and actual desktop/phone views, including a 9,375-row export across batches. Cedar independently reproduced the old source-selection failure and judged Sable's repair on two sources: refusal preserved the earlier job, and corrected selection produced usable tables and figures.

Platform and limits

This is a pre-release for Linux x86-64, built and tested on Ubuntu 24.04 with glibc 2.39. It includes its Python runtime and dependencies. Other distributions have not been tested by this team. The bundle includes licenses and notices for its actual bundled dependencies.

This is the offline command-line product. It does not ship the Windows acquisition desktop or live hardware capture. Tests and visual examples use controlled recordings in real MCAP files; they do not establish instrument calibration, synchronization accuracy or clinical validity. Numeric mean/RMS remains the existing equal-sample estimator. Raw source files are retained; running analysis writes derived jobs inside the package.

The published archive was independently downloaded and run with Python absent from PATH. Sable's actual four-row and 9,375-row selected handoffs retained the expected values and support; Cedar's downloaded two-source course preserved the earlier job on a mistaken selection and recovered two tables and three figures after correction. Fresh runs record fresh job IDs; the real screenshots and clip show the same method, selection and support layout. No physical-device or real-user adoption claim is made.

Recorded numeric sample export — experimental Linux

Choose a tag to compare

@Jacob-Met Jacob-Met released this 10 Nov 19:33
f6e525a

Recorded numeric samples are now reusable

The maintained session exporter and native Export dialog now include generic numeric recordings that were previously omitted. Export every recorded channel/frame into CSV with exact original device timestamps, separate MCAP log/publish clocks, channel labels, units, and copied stream/gap provenance.

This experimental Linux x86-64 standalone CLI includes the cumulative offline QC, feature/plot, saved-job, comparison and selected-feature commands at exact source f6e525a. Numeric sample work merged in PR190, merge15753a5144f40632d3660cf5db3618f4d74d59b7. The source also updates the native Qt Export dialog; a packaged desktop is not included in this Linux CLI archive.

Run

Extract capturesuite-analysis-numeric-linux-x86_64-f6e525a.tar.xz, then:

./capturesuite-analysis/capturesuite-analysis --version
./capturesuite-analysis/capturesuite-analysis export-session recording.mmsession exported-samples --modalities numeric --verify

No Python installation is required. Built on Ubuntu24.04 x86-64; requires compatible glibc/system libraries. Read the included NUMERIC_SAMPLE_EXPORT.md before importing nanosecond timestamps into a spreadsheet.

SHA-256: f1d2f8f37764e6f65b23fdab840aef300440459342e988d86791b3d94b04da8c.

Evidence and limits

Producer astra-lark-61c2 ran62 focused real-protobuf-MCAP/native/export regressions and lint checks on fleet339. Independent astra-tern-93ab used an unfamiliar16-value recording and the actual X11 Export dialog, then the exact frozen executable. Accepted exact >2^53 timestamps, Unicode/quotes, separate reset anchors, missing-clock blanks, NaN/backward-step counts, copied gap declarations, original-source hashes and malformed-input refusal. CSV outputs from the frozen binary matched the independently accepted course byte for byte. Included screenshots are Tern's actual native-dialog use; they show source desktop capability, not a bundled desktop.

The measurements used for this course are synthetic values stored in real MCAP. There is no physical instrument, calibration or multi-device synchronization claim. Derived session estimates use native within-batch deltas only; missing clocks stay unknown. Samples inside declared gaps remain present alongside gap provenance; no repair or resampling occurs. Malformed selected input fails without a successful manifest, though partial CSV may remain.

The archive contains exact installed dependency notices and licenses. Original captures remain unchanged.

CaptureSuite native offline workbench — experimental Linux build

Choose a tag to compare

@Jacob-Met Jacob-Met released this 10 Nov 21:25
b520a9c

CaptureSuite native offline workbench — experimental Linux build

Open a recorded session, inspect its signals, run analysis, and export measurements in a native application. This build includes Python, Qt and the analysis dependencies.

Exact source: b520a9c7d4546bb223431fe21828c38700381156
Platform exercised: Ubuntu 24.04, Linux x86_64, glibc 2.39
Native archive SHA-256: 05d75be412fd3b5643ab29eea14001476ea916c70ff1008ab5e721f9edc8fac0

Start

Download capturesuite-desktop-linux-x86_64.tar.gz, extract the whole folder, and run:

tar -xzf capturesuite-desktop-linux-x86_64.tar.gz
cd capturesuite-desktop
./capturesuite-desktop

A graphical Linux desktop and its display/OpenGL drivers and fonts are required. Keep the executable and its _internal folder together. An external Python installation is not required.

You can also open a recording directly:

./capturesuite-desktop --analysis /path/to/recording.mmsession
./capturesuite-desktop --review /path/to/recording.mmsession

Try the included recording

Extract capturesuite-simulated-three-sensors.zip and open its simulated-signals.mmsession folder in Analysis. Choose Features + plots and run it. The recording contains generated load, skin-temperature and pressure signals. Only the load signal has the recorded gap.

The 48-second capturesuite-native-walkthrough.mp4 shows this real workflow and the resulting CSV. It is silent, uses simulated data and removes waiting periods. The exported example selects t_start_ns and pressure_mean with window start in [38.5,39.5) seconds, yielding the two actual rows at 38.5 and 39 seconds.

What is better

  • Saved analysis jobs are reachable from the main desktop. Reopen their figures, tables and warnings without repeating analysis.
  • The timeline shows the three recorded streams with independent units and their actual session times; gap masking is confined to the affected stream. Zoom and pan survive a resize.
  • Select and reorder measurement columns, then export CSV and Parquet over a precise half-open interval. Existing feature windows and original recordings remain unchanged; units, provenance and warnings travel with the export.
  • Raw-session export runs inside the packaged application. Canonical protobuf EMG/IMU recordings and numeric streams are recognized and retain their documented clock basis.
  • The bundle carries the resources and dependency notices needed by the actual frozen application.

Evidence and limits

astra-mizunagi built this source with the maintained Linux packaging script and ran 43 composition tests plus 33 raw/native/numeric tests on estate compute. astra-gawa-4d17 independently used the actual application with PATH=/nonexistent: fresh QC and feature jobs, 3 tables and 4 figures, correctly separated timeline lanes and gaps, exact ordered CSV/Parquet selection, and saved-job reopening in a second process. All 15 original recording files remained unchanged. astra-nami-6c28 independently exercised frozen canonical and numeric export and produced the real-use walkthrough.

This is an experimental offline release. Physical capture, hardware drivers and pose-model weights are not included. Instrument accuracy and physical clock synchronization have not been established by the simulated example. Other operating systems and older glibc versions have not been exercised. Long measurement names can still be truncated in the column chooser; that usability gap is being addressed. Video codec support depends on the included Qt backend and the recording; optional video rewrapping uses an installed ffmpeg, otherwise original MKV files are copied with that behavior recorded.

The source archive corresponds to the exact build above. Later changes on main are not implicitly part of this binary. Application licenses, upstream notices and launch documentation are included in the extracted folder. Verify downloads with SHA256SUMS.

CaptureSuite LSL0.1.2 — retained recordings and offline consumer

Choose a tag to compare

@Jacob-Met Jacob-Met released this 10 Nov 20:18
b1578c6

CaptureSuite LSL component0.1.2 and runnable offline consumer

Source b1578c6, merged through PR189 at058957f6222bd259d187d7a239f6f6c1c370b95d. Includes PR178 live discovery/independent stopping and shared Python SDK0.1.1 recording handoff.

What works

Numeric LSL sources can stop and restart independently. Their Python-worker segments use the native session tree, preserve earlier segment bytes, use the monotonic session anchor and report complete frame counts/time support. Ordinary session analysis can reopen all retained frames rather than silently crop batch tails.

Astra-Lark ran16 focused tests on the Renpo fleet, including actual local liblsl outlets, public Worker RPC, native SessionPackage creation/finalization, ordinary reader/features and unchanged raw hashes. Astra-Heron independently accepted31Hz/3ch and47Hz/2ch, two restarts and stale stops:53/53 and105/105 retained/reopened frames, exact values/times, segment preservation and independently computed featureCSV means. Integrated runtime source is unchanged from that independently tested candidate.

Assets

  • Linux x86_64 archive: standalone existing CaptureSuite analysis executable, installable SDK/protocol wheels, LSL plugin, real software-outlet recording, complete sampleCSV export, selected featureCSV/Parquet and actual QC screenshot.
  • Component archive: SDK/protocol wheels and plugin for a matching existing CaptureSuite Windows installation, plus the recorded example/exports. Requires Python3.12 and the existing native daemon/desktop.
  • SOURCE_COMMIT and SHA256SUMS identify the exact build and archives.

Extract the Linux archive, enter capturesuite-lsl-0.1.2 and run:
./capturesuite-analysis/capturesuite-analysis --version
./capturesuite-analysis/capturesuite-analysis qc recorded-example/Live-LSL.mmsession --inspect-payloads
./capturesuite-analysis/capturesuite-analysis export-session recorded-example/Live-LSL.mmsession new-export --modalities numeric --verify

The controlled example contains40 frames from a23Hz/two-channel outlet and84 from a37Hz/single-channel outlet:80+84 channel-value CSV rows. These are real LSL transport recordings of software-generated values, not physical sensor data. Existing raw recordings are not migrated.

Boundaries

This prerelease establishes the portable local LSL→Python worker→native storage→offline consumer path. Windows named-pipe/daemon/desktop execution, physical hardware timing, remote-clock synchronization and long recording reliability are unverified. The LSL bridge still applies correction before storing device timestamps, and empty inlet timeouts do not create daemon-owned gap records. Derived timing/features are provisional and uncalibrated. The Linux executable is an offline consumer, not a Linux capture daemon.

Offline Analysis: recorded numeric excerpts (Linux x86_64)

Choose a tag to compare

@Jacob-Met Jacob-Met released this 10 Nov 19:12
38edc6e

This experimental Linux x86_64 release adds exact recorded numeric excerpts to the ordinary CaptureSuite QC report. It is built from merged source 38edc6e (PR #179), including earlier dependency notices, early QC warnings, IMU feature support, and PR #177 EMG sample-index continuity warnings.

The report shows the first 256 sample frames and first six channels of a successfully decoded numeric stream. Each lane labels its independent amplitude scale and descriptor units. Dashed lines mark recorded batch boundaries. The horizontal axis is sample position, not time. Expand the source disclosure for MCAP/message identity and unchanged native timestamp integers. QC still checks all later messages.

Download and run

Download capturesuite-analysis-numeric-excerpts-linux-x86_64.tar.gz and SHA256SUMS from this release.

sha256sum -c SHA256SUMS
tar -xzf capturesuite-analysis-numeric-excerpts-linux-x86_64.tar.gz
cd capturesuite-analysis
./capturesuite-analysis --version
./capturesuite-analysis qc examples/numeric-force.mmsession --inspect-payloads

Open the emitted job directory's reports/qc.html in a browser. The included example is clearly labeled simulated force data stored in genuine protobuf MCAP: a narrow excursion, a flat-to-step channel, 512 frames, and a visible 256-frame excerpt limit. To inspect your own session, replace the example path with its .mmsession directory.

The package includes the executable/runtime, source commit pin, QUICK_START.md with exact help/QC/later-window plots commands, usage/limits, dependency terms and the simulated example. The attached desktop and phone screenshots show the actual report generated by the standalone executable.

Supported scope and limits

Tested on Ubuntu 24.04 x86_64 with glibc 2.39; no Python installation is required to run the archive. Other Linux distributions are unqualified. This is not a Windows capture/desktop release.

Numeric excerpts cover only generic.numeric_batch/1, at most 256 leading frames and six channels. They do not show later events, hidden channels, or elapsed gap duration. EMG/IMU retain their existing decoded checks. Use the report's warnings and gap table when judging continuity. A readable trace does not establish full-recording quality, sensor calibration, physiological validity, or physical synchronization. Simulated data is not hardware qualification.

Source and verification

Exact source: 38edc6e
Change: #179

Reproduce the executable with scripts/build-analysis-linux.sh from this pin using the declared Python 3.12 package dependencies, mcap and PyInstaller. The archive additionally includes the tracked docs/NUMERIC_QC_EXCERPTS.md and the attached simulated example.

Astra-Lumen reviewed the change and ran 63 focused real-MCAP/payload/QC/gap/EMG-index tests, Ruff and diff checks on the fleet. The standalone executable completed the example outside the checkout with its Python environment removed; a clock rewind after the retained excerpt failed with strict exit 2 and produced no successful preview. The HTML was inspected at 1440px and 390px with no phone document overflow. An independent actual-use review is tracked separately; no independent verdict is implied here. GitHub Actions were not used.

Downloaded release use

Astra-Lumen downloaded this published archive on hamon-research-168, verified its SHA256 and source pin, extracted it, and ran it with no Python on PATH. Numeric QC completed with exit 0; a later clock rewind and a missing EMG sample-index span both returned strict exit 2. Raw source hashes and native timestamp integers were preserved. The attached report images were regenerated from that downloaded binary at 1440px and 390px; the phone document had no horizontal overflow. DOWNLOAD_USE.json records these observed results.

Offline analysis: per-sensor IMU measurement support

Choose a tag to compare

@Jacob-Met Jacob-Met released this 10 Nov 19:19
7c9c54e

CaptureSuite Offline Analysis — IMU measurement support (experimental Linux x86_64)

This release includes merged PR176 at source 7c9c54e4ee7c38adeff6fb01300205bc8a354da9, the existing selected-table exporter, corrected EMG gap adjacency, actionable source QC warnings and the maintained dependency notice packaging.

A sensor missing from most recorded IMU frames previously produced ordinary feature values alongside stream validity 1.0. IMU feature table version 2 now includes:

  • window_frame_count for the retained frames in each window.
  • Per-sensor acc_sample_count / acc_valid_fraction and gyro_sample_count / gyro_valid_fraction.
  • One complete finite-vector population for acceleration axes and magnitude; independent gyro support.
  • Unknown values when a measurement has zero support. An observed zero remains a zero.
  • An incomplete-support job warning retained in selected CSV/Parquet exports. CLI warning counts now include all saved job warnings.

The denominator is retained frame slots, including gap-masked slots. These fractions do not estimate elapsed-time coverage and do not prove sensor accuracy, synchronization or calibration. Existing valid_fraction retains its stream gap-mask meaning.

Download and launch

Linux x86_64, built and exercised on Ubuntu with Python 3.12 bundled. No Python installation is required to run the extracted executable.

sha256sum -c SHA256SUMS
tar -xzf capturesuite-analysis-linux-x86_64-7c9c54e.tar.gz
./capturesuite-analysis/capturesuite-analysis --version
./capturesuite-analysis/capturesuite-analysis features /path/to/recording.mmsession --strict-warnings

The analysis prints the retained dir= of its job. Inspect and export actual selected columns:

./capturesuite-analysis/capturesuite-analysis export-table /path/to/recording.mmsession/processing/jobs/JOB --list
./capturesuite-analysis/capturesuite-analysis export-table /path/to/recording.mmsession/processing/jobs/JOB \
  --table features/imu/SOURCE/windows.parquet \
  --column t_mid_ns --column SENSOR_acc_mag_mean \
  --column SENSOR_acc_sample_count --column SENSOR_acc_valid_fraction \
  --column window_frame_count --output /path/to/new-export

Replace SOURCE and SENSOR with the literal names shown by --list. Export destinations must be new directories outside the session package. CSV, Parquet, column metadata and the original job warnings travel together. Strict analysis exit 2 means a completed job has warnings; exit 1 means failure.

Validation and limits

  • Producer astra-marten-9c2f: 77 targeted feature/export/plugin regressions, focused Ruff and diff checks passed on estate Python 3.12. Real protobuf MCAP includes partial sensors, nonfinite vectors, declared gaps and complete controls; original source hashes remained unchanged.
  • Independent judge astra-tern-9c2f: unfamiliar simulated readings in real MCAP reproduced the old problem in the published executable. Candidate acceptance covered 75% sensor support, 80% finite gyro support, zero-support unknowns, genuine zero readings, consistent vector populations and typed CSV/Parquet equivalence. A frozen candidate reproduced the source CSV byte-for-byte.
  • Maintained build scripts/build-analysis-linux.sh; no GitHub Actions. Bundled source pin is SOURCE_COMMIT; dependency versions and 42 actual distribution/runtime notice collections are included.
  • This is the offline CLI, not Windows acquisition or a native desktop distribution. Hardware accuracy, real-instrument timing, Windows execution and long-session performance are not claimed.
  • Source-release download/run and demonstration evidence are recorded under task capturesuite-current-workbench-judge-tern9c2f and will be attached after the released bytes are received.

Offline analysis: EMG continuity (Linux experimental)

Choose a tag to compare

@Jacob-Met Jacob-Met released this 10 Nov 20:36
38edc6e

CaptureSuite offline analysis: recorded EMG continuity

Experimental Linux x86-64 build of exact source 38edc6e (merged PR177). This version includes the cumulative offline analysis work at that commit: payload QC, saved feature export, nested signal plots, warning-first responsive reports, numeric excerpts and IMU checks.

What changed

Decoded EMG QC checks each batch's first sample index against the preceding decoded batch's expected end, including transitions between MCAP segments. Missing ranges, overlap and counter resets now produce a warning with exact expected/observed indices. Arbitrary starting indices remain valid. Reset handling re-anchors the next comparison rather than repeating the same error. Raw recording files are unchanged.

Run

Download the Linux archive and SHA256SUMS, then:

sha256sum --check --ignore-missing SHA256SUMS
tar -xzf capturesuite-analysis-linux-x86_64.tar.gz
./capturesuite-analysis/capturesuite-analysis --version
./capturesuite-analysis/capturesuite-analysis qc "/data/session.mmsession" --inspect-payloads --strict-warnings

Open the printed job's reports/qc.html in a browser. Exit 0 means the QC job passed its requested checks; exit 2 means strict warnings. The bundled README documents features, plots, all and export-table.

Observed validation

Producer astra-larch-91e ran the maintained scripts/build-analysis-linux.sh on hamon-build-195, exact source above. 75 relevant source tests and focused Ruff checks passed at the merge. The actual standalone executable ran outside the checkout with PATH=/nonexistent: arbitrary-origin continuous EMG exited0; missing batch expected4/observed8 exited2 with exactly1 forward jump; reset expected104/observed0 exited2 with exactly1 overlap/reset and no spurious follow-on warning. All raw source hashes stayed unchanged. release-use.json records these three results; synthetic input packages are attached. Actual emitted HTML was opened in Chrome at1440x1000 and390x844 and visually inspected; document width equalled viewport and the warning was visible on the first phone screen. Attached screenshots show that final merge report.

Independent judge astra-sable-4f1 separately built/exercised predecessor1884e4ade73507ee7a4721d1c9b26abe2f9ebb72 using heldout multi-segment continuous/missing/overlap/reset recordings. Scoped PASS: exact single discontinuity count, raw unchanged, real desktop/phone report inspected. Task capturesuite-release-operator-judge-sable4f1 retains the independent evidence. The final merged38 build's producer checks are separately identified above; independent judgment is not relabeled as a final38 rebuild.

Platform and limits

Linux x86-64, built with Python3.12 on Ubuntu24.04-era glibc; compatible system libraries required, no installed Python needed. This is the offline command-line analysis package, not a capture desktop or hardware validation. Physical sensors, clinical accuracy, full-duration recordings, Windows/macOS binaries and native desktop operation are not certified by these checks. EMG index discontinuities identify recorded index changes; they do not establish an exact number of physically lost samples. Reports retain a source-oriented run_analysis.py example; use the executable command above for this bundle.

Exact full source archive, dependency versions, GPL license and third-party notices are supplied. This release is explicitly experimental and does not change GitHub's latest release pointer.