imuops 0.3.1__tar.gz

imuops-0.3.1/.gitignore

@@ -0,0 +1,18 @@
+.DS_Store
+.pytest_cache/
+.ruff_cache/
+.venv/
+.coverage
+coverage.xml
+__pycache__/
+*.pyc
+*.pyo
+*.pyd
+*.ipynb_checkpoints
+build/
+dist/
+htmlcov/
+downloads/
+output/
+reports/
+tmp/

imuops-0.3.1/LICENSE

@@ -0,0 +1,22 @@
+MIT License
+
+Copyright (c) 2026 Boz Liu
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+

imuops-0.3.1/PKG-INFO

@@ -0,0 +1,162 @@
+Metadata-Version: 2.4
+Name: imuops
+Version: 0.3.1
+Summary: IMU data QA, reliability scoring, and reproducibility tooling
+Author: Boz Liu
+License: MIT
+License-File: LICENSE
+Keywords: benchmark,data-quality,har,imu,pdr,qa,sensor-fusion,wearable
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering
+Classifier: Topic :: Scientific/Engineering :: Information Analysis
+Requires-Python: >=3.11
+Requires-Dist: ahrs>=0.3.1
+Requires-Dist: h5py>=3.11.0
+Requires-Dist: jinja2>=3.1.4
+Requires-Dist: numpy>=1.26.0
+Requires-Dist: openpyxl>=3.1.0
+Requires-Dist: pandas>=2.2.0
+Requires-Dist: plotly>=5.24.1
+Requires-Dist: pyarrow>=15.0.0
+Requires-Dist: pydantic>=2.8.2
+Requires-Dist: pyyaml>=6.0.2
+Requires-Dist: requests>=2.32.0
+Requires-Dist: scikit-learn>=1.5.2
+Requires-Dist: typer>=0.12.0
+Provides-Extra: dev
+Requires-Dist: pytest-cov>=5.0.0; extra == 'dev'
+Requires-Dist: pytest>=8.2.0; extra == 'dev'
+Description-Content-Type: text/markdown
+
+# imuops
+
+`imuops` is a public alpha for **IMU data QA, reliability scoring, and reproducibility**.
+
+It is built around one core question:
+
+> Is this IMU data trustworthy, why is it failing, and how does that affect baseline algorithms?
+
+The product is intentionally **tabular-first**. The main path is customer-shaped `csv`, `tsv`, or `parquet` data plus a small YAML mapping file. Benchmark adapters stay in the repo for reproducibility demos, but they are secondary to the tabular workflow.
+
+## Status
+
+- alpha / preview release
+- main product path: `tabular`
+- benchmark/demo adapters: `ronin`, `oxiod`, `wisdm`
+- local-only contrib adapter: `legacy_arduino`
+
+Benchmark adapters are validated on fixtures and demo flows. They are not guaranteed across every upstream packaging or layout variant.
+
+## Install
+
+Requires Python 3.11+.
+
+Primary public path:
+
+```bash
+python3.12 -m venv .venv
+source .venv/bin/activate
+pip install .
+```
+
+Use any Python 3.11+ interpreter available on your machine.
+
+Optional `uv` path:
+
+```bash
+uv venv
+source .venv/bin/activate
+uv pip install .
+```
+
+`imuops` is prepared for PyPI publication, but this alpha does not claim a live PyPI package yet.
+
+## Quickstart
+
+After install, the offline first-run path is exactly three commands:
+
+```bash
+imuops ingest tabular examples/sample_tabular_imu.csv --config examples/sample_tabular_config.yaml --out output/sample_tabular_demo
+imuops audit output/sample_tabular_demo --summary-format markdown
+imuops report output/sample_tabular_demo --out output/sample_tabular_demo/report.html
+```
+
+You can also use the bundled demo wrapper:
+
+```bash
+bash examples/run_tabular_demo.sh
+```
+
+## What It Does
+
+- normalizes messy IMU tables into one canonical session format
+- audits timing, clipping, dropout, magnetic disturbance, bias drift, and related issues
+- computes a versioned `trust_score` with explicit penalties, weights, and thresholds
+- replays conservative baseline algorithms for orientation and PDR reproducibility
+- benchmarks task-aware baselines where labels or trajectories exist
+- compares clean vs corrupted or before vs after sessions
+- batches QA over many sessions and writes machine-readable summaries for CI use
+
+## Core Commands
+
+```bash
+imuops ingest tabular /path/to/session.csv --config /path/to/mapping.yaml --out output/session_a
+imuops audit output/session_a --fail-below 0.80 --summary-format markdown
+imuops export output/session_a --profile qa_filtered --format parquet --out output/session_a_clean
+imuops compare output/session_a output/session_b --out output/compare.html
+imuops batch audit output --out output/batch_artifacts
+```
+
+## Included Adapters
+
+### Market-facing default
+
+- `tabular`: customer-shaped `csv`, `tsv`, and `parquet` sources with YAML mapping and unit conversion
+
+### Benchmark/demo adapters
+
+- `ronin`: clean inertial odometry benchmark sessions
+- `oxiod`: clean handheld / phone inertial odometry benchmark sessions
+- `wisdm`: lightweight HAR benchmark sessions
+
+### Contrib/local regression
+
+- `legacy_arduino`: historical Arduino/MPU9255 adapter kept only for local regression and examples
+
+## Trust Score
+
+`imuops` publishes the trust-score contract directly into artifacts and reports:
+
+- per-window formula
+- session aggregation formula
+- penalty totals
+- weight profile
+- thresholds
+
+That is documented in [docs/trustscore.md](docs/trustscore.md), and the current validation tranche is in [docs/trustscore_validation.md](docs/trustscore_validation.md).
+
+## Docs
+
+- [Architecture](docs/architecture.md)
+- [Datasets](docs/datasets.md)
+- [Trust Score](docs/trustscore.md)
+- [Trust-Score Validation](docs/trustscore_validation.md)
+- [Schema Compatibility](docs/schema_compatibility.md)
+- [Release Checklist](docs/release.md)
+- [Contributing](CONTRIBUTING.md)
+
+## Public Alpha Notes
+
+- This repo is a **truthful alpha**, not a commercial deployment claim.
+- Reports support `--redact-source-path` and `--redact-subject-id` for safer sharing.
+- Output quality is strongest for tabular customer data and fixture/demo benchmark layouts.
+
+## License
+
+[MIT](LICENSE)

imuops-0.3.1/README.md

@@ -0,0 +1,126 @@
+# imuops
+
+`imuops` is a public alpha for **IMU data QA, reliability scoring, and reproducibility**.
+
+It is built around one core question:
+
+> Is this IMU data trustworthy, why is it failing, and how does that affect baseline algorithms?
+
+The product is intentionally **tabular-first**. The main path is customer-shaped `csv`, `tsv`, or `parquet` data plus a small YAML mapping file. Benchmark adapters stay in the repo for reproducibility demos, but they are secondary to the tabular workflow.
+
+## Status
+
+- alpha / preview release
+- main product path: `tabular`
+- benchmark/demo adapters: `ronin`, `oxiod`, `wisdm`
+- local-only contrib adapter: `legacy_arduino`
+
+Benchmark adapters are validated on fixtures and demo flows. They are not guaranteed across every upstream packaging or layout variant.
+
+## Install
+
+Requires Python 3.11+.
+
+Primary public path:
+
+```bash
+python3.12 -m venv .venv
+source .venv/bin/activate
+pip install .
+```
+
+Use any Python 3.11+ interpreter available on your machine.
+
+Optional `uv` path:
+
+```bash
+uv venv
+source .venv/bin/activate
+uv pip install .
+```
+
+`imuops` is prepared for PyPI publication, but this alpha does not claim a live PyPI package yet.
+
+## Quickstart
+
+After install, the offline first-run path is exactly three commands:
+
+```bash
+imuops ingest tabular examples/sample_tabular_imu.csv --config examples/sample_tabular_config.yaml --out output/sample_tabular_demo
+imuops audit output/sample_tabular_demo --summary-format markdown
+imuops report output/sample_tabular_demo --out output/sample_tabular_demo/report.html
+```
+
+You can also use the bundled demo wrapper:
+
+```bash
+bash examples/run_tabular_demo.sh
+```
+
+## What It Does
+
+- normalizes messy IMU tables into one canonical session format
+- audits timing, clipping, dropout, magnetic disturbance, bias drift, and related issues
+- computes a versioned `trust_score` with explicit penalties, weights, and thresholds
+- replays conservative baseline algorithms for orientation and PDR reproducibility
+- benchmarks task-aware baselines where labels or trajectories exist
+- compares clean vs corrupted or before vs after sessions
+- batches QA over many sessions and writes machine-readable summaries for CI use
+
+## Core Commands
+
+```bash
+imuops ingest tabular /path/to/session.csv --config /path/to/mapping.yaml --out output/session_a
+imuops audit output/session_a --fail-below 0.80 --summary-format markdown
+imuops export output/session_a --profile qa_filtered --format parquet --out output/session_a_clean
+imuops compare output/session_a output/session_b --out output/compare.html
+imuops batch audit output --out output/batch_artifacts
+```
+
+## Included Adapters
+
+### Market-facing default
+
+- `tabular`: customer-shaped `csv`, `tsv`, and `parquet` sources with YAML mapping and unit conversion
+
+### Benchmark/demo adapters
+
+- `ronin`: clean inertial odometry benchmark sessions
+- `oxiod`: clean handheld / phone inertial odometry benchmark sessions
+- `wisdm`: lightweight HAR benchmark sessions
+
+### Contrib/local regression
+
+- `legacy_arduino`: historical Arduino/MPU9255 adapter kept only for local regression and examples
+
+## Trust Score
+
+`imuops` publishes the trust-score contract directly into artifacts and reports:
+
+- per-window formula
+- session aggregation formula
+- penalty totals
+- weight profile
+- thresholds
+
+That is documented in [docs/trustscore.md](docs/trustscore.md), and the current validation tranche is in [docs/trustscore_validation.md](docs/trustscore_validation.md).
+
+## Docs
+
+- [Architecture](docs/architecture.md)
+- [Datasets](docs/datasets.md)
+- [Trust Score](docs/trustscore.md)
+- [Trust-Score Validation](docs/trustscore_validation.md)
+- [Schema Compatibility](docs/schema_compatibility.md)
+- [Release Checklist](docs/release.md)
+- [Contributing](CONTRIBUTING.md)
+
+## Public Alpha Notes
+
+- This repo is a **truthful alpha**, not a commercial deployment claim.
+- Reports support `--redact-source-path` and `--redact-subject-id` for safer sharing.
+- Output quality is strongest for tabular customer data and fixture/demo benchmark layouts.
+
+## License
+
+[MIT](LICENSE)

imuops-0.3.1/datasets/manifest.toml

@@ -0,0 +1,44 @@
+[tabular]
+adapter = "tabular"
+task = "orientation_or_pdr_or_har"
+reference_type = "optional_labels_or_ground_truth"
+description = "Primary config-driven adapter for customer-shaped CSV, TSV, and Parquet IMU data."
+fetch_mode = "user_supplied"
+body_location = "configurable"
+device_pose = "configurable"
+
+[legacy_arduino]
+adapter = "legacy_arduino"
+task = "pdr"
+reference_type = "gps_optional"
+description = "Experimental contrib adapter for historical Arduino and MPU9255 logs from the local workspace."
+fetch_mode = "local_only"
+body_location = "unknown"
+device_pose = "arbitrary"
+
+[ronin]
+adapter = "ronin"
+task = "pdr"
+reference_type = "trajectory"
+description = "RoNIN-style inertial odometry session directory."
+fetch_script = "scripts/fetch_ronin_sample.py"
+body_location = "body_unknown"
+device_pose = "arbitrary"
+
+[oxiod]
+adapter = "oxiod"
+task = "pdr"
+reference_type = "trajectory"
+description = "OxIOD IMU and visual-inertial ground-truth pair."
+fetch_script = "scripts/fetch_oxiod_sample.py"
+body_location = "handheld"
+device_pose = "arbitrary"
+
+[wisdm]
+adapter = "wisdm"
+task = "har"
+reference_type = "activity_labels"
+description = "WISDM-style accelerometer activity recognition trace."
+fetch_script = "scripts/fetch_wisdm_sample.py"
+body_location = "phone"
+device_pose = "free_carry"

imuops-0.3.1/docs/architecture.md

@@ -0,0 +1,71 @@
+# Architecture
+
+`imuops` keeps each adapter thin and pushes the rest of the system onto one canonical session format. The pipeline is task-aware, but the product is intentionally **tabular-first**: benchmark adapters remain useful for reproducibility demos, while the main public-alpha path is customer-shaped CSV, TSV, and Parquet plus YAML mapping.
+
+## Core flow
+
+1. Adapter ingests source files into:
+   - `session.json`
+   - `imu.parquet`
+   - `gps.parquet`
+   - `ground_truth.parquet`
+2. Audit reads the canonical session and writes:
+   - `issues.json`
+   - `audit_summary.json`
+   - CI-friendly console summaries and exit codes
+3. Replay reads the same session and writes one artifact pair per baseline:
+   - `replay_<baseline>.parquet`
+   - `replay_<baseline>_summary.json`
+4. Benchmark runs task-aware baselines and writes:
+   - `benchmark_summary.json`
+   - `benchmark_<task>_<baseline>.json`
+5. Corruption writes a second canonical session plus:
+   - `corruption.json`
+6. Report combines the session, audit output, benchmark summaries, replay artifacts, and corruption context into a single HTML file.
+
+## Canonical metadata
+
+`session.json` is validated with `pydantic` and includes:
+
+- `schema_version`
+- `trustscore_version`
+- `dataset`
+- `session_id`
+- `task`
+- `reference_type`
+- `nominal_hz`
+- `body_location`
+- `device_pose`
+- `label_namespace`
+- `sensors`
+- `ground_truth_available`
+- `labels_available`
+
+This lets the report and benchmark layers stay generic even when the source datasets have very different conventions.
+
+## Canonical units
+
+- `t_ms`: milliseconds from session start
+- accelerometer: `m/s^2`
+- gyroscope: `rad/s`
+- magnetometer: `uT`
+- pressure: `Pa`
+- temperature: `C`
+
+## Supported tasks and baselines
+
+- `orientation`: `madgwick`, `mahony`
+- `pdr`: step-heading dead reckoning baseline
+- `har`: fixed-window hand-crafted features + random forest baseline
+
+The replay layer is deliberately conservative. It exists to make QA and benchmarking reproducible, not to claim novelty.
+
+## Core adapters vs contrib adapters
+
+- Core:
+  - `tabular`: CSV, TSV, and Parquet plus YAML mapping
+  - `ronin`, `oxiod`, `wisdm`: benchmark/demo adapters validated on fixtures and demo flows
+- Contrib:
+  - `legacy_arduino`: historical Arduino/MPU9255 adapter kept for local regression only
+
+The adapters intentionally stay tolerant enough for small local extracts and hand-curated samples, but only the `tabular` adapter is positioned as the default customer-data interface.

imuops-0.3.1/docs/artifacts/trustscore_validation_summary.json

@@ -0,0 +1,91 @@
+{
+  "all_non_improving": true,
+  "clean": {
+    "status": "pass",
+    "trust_score": 1.0
+  },
+  "presets": [
+    {
+      "baseline_metric_deltas": {
+        "benchmark_pdr_heading_drift_rad": 0.015089873124199377,
+        "benchmark_pdr_path_smoothness": 8.874562068890959e-07,
+        "benchmark_pdr_step_count": 0.0,
+        "benchmark_pdr_trajectory_rmse_m": 0.012373560291695895,
+        "replay_pdr_heading_drift_rad": 0.015089873124199377,
+        "replay_pdr_path_smoothness": 8.874562068890959e-07,
+        "replay_pdr_step_count": 0.0,
+        "replay_pdr_trajectory_rmse_m": 0.012373560291695895
+      },
+      "non_improving": true,
+      "preset": "packet_loss_5",
+      "trust_score": 0.5323465944828356,
+      "trust_score_delta": -0.4676534055171644
+    },
+    {
+      "baseline_metric_deltas": {
+        "benchmark_pdr_heading_drift_rad": 0.0,
+        "benchmark_pdr_path_smoothness": 0.0,
+        "benchmark_pdr_step_count": 0.0,
+        "benchmark_pdr_trajectory_rmse_m": 0.0,
+        "replay_pdr_heading_drift_rad": 0.0,
+        "replay_pdr_path_smoothness": 0.0,
+        "replay_pdr_step_count": 0.0,
+        "replay_pdr_trajectory_rmse_m": 0.0
+      },
+      "non_improving": true,
+      "preset": "timestamp_jitter_3ms",
+      "trust_score": 1.0,
+      "trust_score_delta": 0.0
+    },
+    {
+      "baseline_metric_deltas": {
+        "benchmark_pdr_heading_drift_rad": 2.809794459825378,
+        "benchmark_pdr_path_smoothness": -1.0951255003082612e-06,
+        "benchmark_pdr_step_count": 0.0,
+        "benchmark_pdr_trajectory_rmse_m": 0.0,
+        "replay_pdr_heading_drift_rad": 2.809794459825378,
+        "replay_pdr_path_smoothness": -1.0951255003082612e-06,
+        "replay_pdr_step_count": 0.0,
+        "replay_pdr_trajectory_rmse_m": 0.0
+      },
+      "non_improving": true,
+      "preset": "axis_flip_x",
+      "trust_score": 1.0,
+      "trust_score_delta": 0.0
+    },
+    {
+      "baseline_metric_deltas": {
+        "benchmark_pdr_heading_drift_rad": -0.028779030258548888,
+        "benchmark_pdr_path_smoothness": 6.97476635826116e-07,
+        "benchmark_pdr_step_count": 0.0,
+        "benchmark_pdr_trajectory_rmse_m": 0.0,
+        "replay_pdr_heading_drift_rad": -0.028779030258548888,
+        "replay_pdr_path_smoothness": 6.97476635826116e-07,
+        "replay_pdr_step_count": 0.0,
+        "replay_pdr_trajectory_rmse_m": 0.0
+      },
+      "non_improving": true,
+      "preset": "gyro_bias_small",
+      "trust_score": 1.0,
+      "trust_score_delta": 0.0
+    },
+    {
+      "baseline_metric_deltas": {
+        "benchmark_pdr_heading_drift_rad": -0.1593284368719841,
+        "benchmark_pdr_path_smoothness": -7.053313986061677e-07,
+        "benchmark_pdr_step_count": 0.0,
+        "benchmark_pdr_trajectory_rmse_m": 0.0,
+        "replay_pdr_heading_drift_rad": -0.1593284368719841,
+        "replay_pdr_path_smoothness": -7.053313986061677e-07,
+        "replay_pdr_step_count": 0.0,
+        "replay_pdr_trajectory_rmse_m": 0.0
+      },
+      "non_improving": true,
+      "preset": "mag_bias_30ut",
+      "trust_score": 1.0,
+      "trust_score_delta": 0.0
+    }
+  ],
+  "session_id": "sample_tabular_session",
+  "task": "pdr"
+}

imuops-0.3.1/docs/datasets.md

@@ -0,0 +1,118 @@
+# Datasets
+
+`imuops` is designed to mix customer-shaped tabular IMU data with benchmark sessions. In this public alpha, the default product entrypoint is the config-driven `tabular` adapter.
+
+## Adapter Matrix
+
+| Adapter | Task | Expected source | Reference | Notes |
+| --- | --- | --- | --- | --- |
+| `tabular` | `orientation` / `pdr` / `har` | `csv`, `tsv`, or `parquet` plus YAML config | optional labels / ground truth | Main customer-data path |
+| `ronin` | `pdr` | `info.json` + `data.hdf5` session dir | trajectory | Benchmark/demo adapter validated on fixtures and demo flows |
+| `oxiod` | `pdr` | `imu*.csv` file with matching `vi*.csv` | trajectory | Benchmark/demo adapter validated on fixtures and demo flows |
+| `wisdm` | `har` | WISDM-style `txt` or `csv` | activity labels | Benchmark/demo adapter validated on fixtures and demo flows |
+
+The machine-readable source registry lives in [datasets/manifest.toml](../datasets/manifest.toml).
+
+## Tabular Customer Data
+
+The recommended public workflow is a tabular file plus a YAML mapping file. The repo bundles a small offline sample in [examples/sample_tabular_imu.csv](../examples/sample_tabular_imu.csv) and [examples/sample_tabular_config.yaml](../examples/sample_tabular_config.yaml).
+
+The first public-alpha path is:
+
+```bash
+imuops ingest tabular examples/sample_tabular_imu.csv --config examples/sample_tabular_config.yaml --out output/sample_tabular_demo
+imuops audit output/sample_tabular_demo --summary-format markdown
+imuops report output/sample_tabular_demo --out output/sample_tabular_demo/report.html
+```
+
+The YAML mapping controls:
+
+- timestamp column and time unit
+- accel / gyro / mag column names and units
+- optional temperature, pressure, and label columns
+- optional ground-truth position and heading columns
+- metadata such as task, body location, and device pose
+
+## Contrib / Local Regression
+
+The legacy adapter is intentionally non-core and intentionally local:
+
+- it is available via `imuops.contrib`
+- it is not part of the public quickstart
+- it is useful only for local regression on the historical workspace
+
+You can verify that it is present with:
+
+```bash
+python -c "from imuops.contrib import LegacyArduinoAdapter; print(LegacyArduinoAdapter.name)"
+```
+
+## Public Sample Fetch Helpers
+
+The fetch scripts do not vendor datasets into the repo. They either download a public archive you specify or unpack a local archive you already obtained.
+
+## Privacy and Sharing Notes
+
+- benchmark fetch helpers assume the user is responsible for dataset license compliance
+- `session.json` may contain `source_path` and `subject_id`
+- shared reports should typically be generated with `--redact-source-path --redact-subject-id`
+
+### RoNIN
+
+```bash
+python scripts/fetch_ronin_sample.py --zip-path /path/to/ronin_sample.zip --out downloads/ronin
+imuops ingest ronin downloads/ronin/<session_dir> --out output/ronin_demo
+```
+
+### OxIOD
+
+```bash
+python scripts/fetch_oxiod_sample.py --zip-path /path/to/oxiod_sample.zip --out downloads/oxiod
+imuops ingest oxiod downloads/oxiod/<session_dir>/imu1.csv --out output/oxiod_demo
+```
+
+### WISDM
+
+```bash
+python scripts/fetch_wisdm_sample.py --zip-path /path/to/wisdm_sample.zip --out downloads/wisdm
+imuops ingest wisdm downloads/wisdm/WISDM_ar_v1.1_raw.txt --out output/wisdm_demo
+```
+
+## Demo Commands
+
+### Customer-shaped tabular data
+
+```bash
+imuops ingest tabular /path/to/customer_session.csv --config examples/tabular_config.example.yaml --out output/customer_session
+imuops audit output/customer_session --fail-below 0.80
+imuops export output/customer_session --profile qa_filtered --format parquet --out output/customer_session_clean
+imuops report output/customer_session --out output/customer_session/report.html
+```
+
+### Clean navigation
+
+```bash
+imuops audit output/ronin_demo
+imuops replay output/ronin_demo --baseline madgwick
+imuops benchmark output/ronin_demo --task orientation
+imuops report output/ronin_demo --out output/ronin_demo/report.html
+```
+
+### Clean HAR
+
+```bash
+imuops audit output/wisdm_demo
+imuops benchmark output/wisdm_demo --task har
+imuops report output/wisdm_demo --out output/wisdm_demo/report.html
+```
+
+### Robustness regression
+
+```bash
+imuops corrupt output/ronin_demo --preset packet_loss_5 --out output/ronin_demo_packet_loss
+imuops audit output/ronin_demo_packet_loss
+imuops report output/ronin_demo_packet_loss --out output/ronin_demo_packet_loss/report.html
+imuops compare output/ronin_demo output/ronin_demo_packet_loss --out output/ronin_compare.html
+```
+
+Benchmark adapters are useful for demos and reproducibility, but they are secondary to the tabular customer-data path.