physlint 0.2.0a2__tar.gz → 0.3.2__tar.gz

physlint-0.3.2/.github/ISSUE_TEMPLATE/evidence-submission.yml

@@ -0,0 +1,59 @@
+name: Community evidence submission
+description: Put a public robot dataset or recording on the Physlint Community Evidence Atlas.
+title: "[evidence]: "
+labels: ["evidence", "community"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for contributing to the public map of robot data health. Submit only public, non-sensitive sources that can be revision-pinned. Do not attach private datasets, credentials, or reports containing private paths.
+  - type: input
+    id: source
+    attributes:
+      label: Public dataset or recording
+      description: Link the public repository, dataset, or recording.
+      placeholder: https://huggingface.co/datasets/owner/dataset
+    validations:
+      required: true
+  - type: input
+    id: revision
+    attributes:
+      label: Immutable revision
+      description: Provide a commit, tag, release, checksum, or other stable revision when available.
+      placeholder: Commit SHA, tag, or release identifier
+    validations:
+      required: true
+  - type: dropdown
+    id: profile
+    attributes:
+      label: Data profile
+      options:
+        - LeRobot
+        - Generic MCAP
+        - ROS 2 over MCAP
+        - Another format
+    validations:
+      required: true
+  - type: textarea
+    id: context
+    attributes:
+      label: Robot and collection context
+      description: Tell the community what robot, task, sensors, and collection setup this evidence represents.
+    validations:
+      required: true
+  - type: textarea
+    id: expectation
+    attributes:
+      label: What should we learn from it?
+      description: Describe whether you expect a healthy result, a known defect, or a finding that needs review.
+    validations:
+      required: true
+  - type: checkboxes
+    id: consent
+    attributes:
+      label: Publication agreement
+      options:
+        - label: I confirm this source is public and may be referenced by name in the Community Evidence Atlas.
+          required: true
+        - label: I understand Physlint will publish scoped validation evidence, not a safety certification or universal quality score.
+          required: true

{physlint-0.2.0a2 → physlint-0.3.2}/CHANGELOG.md

@@ -1,5 +1,32 @@
 # Changelog
 
+## 0.3.2 — 2026-09-06
+
+- Launch the Physlint Community Evidence Atlas with 23 published observations, community survey context, real defect stories, and stable evidence links.
+- Add a public evidence-submission workflow for dataset and recording maintainers.
+- Replace the public regressions section with a clearer community narrative while retaining the generated comparison catalog for reproducibility.
+- Refresh the PyPI monthly download snapshot to 289 downloads.
+
+## 0.3.1 — 2026-08-30
+
+- Add a stratified LeRobot compatibility survey, kept separate from the four-snapshot release gate.
+- Accept Hub feature `names` stored as a single string instead of crashing discovery.
+
+## 0.3.0 — 2026-08-30
+
+- Add `physlint compare` for dataset and report regression diffs, including coverage drift without a quality score.
+- Add reviewed baselines via `physlint baseline` and `--baseline`; suppressions require fingerprint, rule ID, author, reason, and optional expiry.
+- Add JUnit, SARIF, and local HTML reporters alongside versioned JSON.
+- Add a plugin loader for `module:Class` specs and `physlint.rules` entry points, plus an example idle-prefix rule.
+- Cache expensive video rule results on disk, keyed by source fingerprint and rule options.
+- Detect exact duplicate numeric episodes with `duplication.exact_episodes`.
+- Publish Observatory regression evidence generated from committed clean-versus-corruption reports.
+
+## 0.2.0a2
+
+- Refined the Physlint Observatory launch experience and branding.
+- Continued validation coverage for LeRobot, MCAP, and ROS 2 workflows.
+
 ## 0.2.0a1 — development
 
 - Add standalone generic MCAP discovery, content fingerprints, CRC-aware scanning, and seven container rules.
@@ -21,7 +48,3 @@
 - Add a reproducible four-dataset and three-corruption release gate.
 
 This alpha supports local LeRobot v3.x directories. MCAP/ROS 2 and Robomimic are roadmap items, not implemented adapters.
-## 0.2.0a2
-
-- Refined the Physlint Observatory launch experience and branding.
-- Continued validation coverage for LeRobot, MCAP, and ROS 2 workflows.

{physlint-0.2.0a2 → physlint-0.3.2}/CONTRIBUTING.md

@@ -9,10 +9,12 @@ mypy
 pytest
 ```
 
-Rules must include positive and negative fixtures, a controlled corruption test where applicable, stable evidence and remediation, documented limitations, and a characterized cost. Adapters must remain read-only, metadata-first, lazy over samples, and explicit about capabilities.
+Rules must include positive and negative fixtures, a controlled corruption test where applicable, stable evidence and remediation, documented limitations, and a characterized cost. Adapters must remain read-only, metadata-first, lazy over samples, and explicit about capabilities. External rules should follow [docs/plugins.md](docs/plugins.md).
 
 New adapter proposals should start with the format-request issue template and include immutable public healthy examples, controlled defects, stream/episode semantics, and the real failure modes the adapter should catch. See `docs/roadmap.md` for the acceptance gate.
 
 The pinned public release gate is intentionally separate from normal CI because it downloads hundreds of MiB. Maintainers can run it locally with `python -m validation.harness` or trigger the manual real-data workflow.
 
+The compatibility survey is also separate from CI and from the release gate. It downloads a stratified public corpus and records outcomes without requiring a pass. Run `python -m validation.survey_harness` and read `docs/validation/compatibility-survey.md` before adding Hub datasets. Do not promote a survey row into the release gate until every finding has been classified by hand.
+
 Do not add network access, telemetry, source-data mutation, large fixtures, or a new default dependency without an issue describing the measured requirement.

{physlint-0.2.0a2 → physlint-0.3.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: physlint
-Version: 0.2.0a2
+Version: 0.3.2
 Summary: Local-first integrity validation for physical-AI and robot-learning data
 Project-URL: Homepage, https://github.com/pranayjoshi/physlint
 Project-URL: Documentation, https://github.com/pranayjoshi/physlint#readme
@@ -54,17 +54,19 @@ Local-first, deterministic validation for physical-AI recordings and robot-learn
 [![Downloads per month](https://img.shields.io/endpoint?url=https%3A%2F%2Fraw.githubusercontent.com%2Fpranayjoshi%2Fphyslint%2Fmaster%2Fdocs%2Fmetrics%2Fdownloads.json&style=flat-square)](https://pypistats.org/packages/physlint)
 ![Python](https://img.shields.io/badge/Python-3.11%2B-3776AB?style=flat-square&logo=python&logoColor=white)
 [![MIT](https://img.shields.io/badge/License-MIT-2ea44f?style=flat-square)](LICENSE)
-![Status](https://img.shields.io/badge/Status-Public%20Alpha-f59e0b?style=flat-square)
+![Status](https://img.shields.io/badge/Status-Released-0f766e?style=flat-square)
 
 [Quickstart](#quickstart) · [Demo](#see-it-catch-a-real-defect) · [Rules](#what-physlint-checks) · [Evidence](#reproducible-public-data-evidence) · [Observatory](#physlint-observatory) · [Roadmap](#format-roadmap) · [Contributing](#contributing)
 
-<img src="docs/assets/launch/exports/validation-summary.png" alt="Physlint public-alpha validation: four of four clean snapshots passed and three of three controlled corruptions were detected across 74 episodes and 31,258 frames." width="100%">
+<img src="docs/assets/launch/exports/evidence-index.png" alt="Physlint evidence index showing robot dataset and recording health checks across LeRobot, MCAP, and ROS 2 data." width="100%">
 
 </div>
 
 Physlint finds concrete integrity defects before robot data reaches training. It explains the impact, identifies the affected episode and stream, recommends remediation, writes a stable JSON report, and returns a CI-safe exit code.
 
-The released `0.1.0a1` public alpha supports **LeRobot Dataset v3.x**. The current `0.2.0a1` development line adds generic **MCAP** container validation and a **ROS 2-over-MCAP** semantic profile, backed by pinned conformance data, a real robot-captured ROS 2 episode, and controlled evidence. Native rosbag2 SQLite remains a conversion workflow for now.
+The released `0.3.2` supports **LeRobot Dataset v3.x**, generic **MCAP** container validation, and a **ROS 2-over-MCAP** semantic profile, plus dataset comparison, reviewed baselines, CI report formats, and a plugin SDK. It also launches the [Community Evidence Atlas](https://physlint.com), where maintainers can explore reproducible findings and submit public robot data for inclusion. Claims stay scoped to pinned conformance data and controlled evidence. Native rosbag2 SQLite remains a conversion workflow.
+
+At the `0.3.2` release snapshot, PyPI Stats reported **289 downloads in the preceding month**.
 
 > [!IMPORTANT]
 > Physlint validates configured data-integrity contracts. A pass does not certify policy quality, task success, or robot safety.
@@ -84,13 +86,13 @@ Physlint requires Python 3.11 or newer.
 ### Install from PyPI
 
 ```bash
-python -m pip install "physlint[video]==0.1.0a1"
+python -m pip install "physlint[video]==0.3.2"
 ```
 
 To test the exact tagged source instead of the PyPI distribution, install the GitHub release directly:
 
 ```bash
-python -m pip install "physlint[video] @ git+https://github.com/pranayjoshi/physlint.git@v0.1.0a1"
+python -m pip install "physlint[video] @ git+https://github.com/pranayjoshi/physlint.git@v0.3.2"
 ```
 
 ### Check a LeRobot dataset or MCAP recording
@@ -106,12 +108,29 @@ physlint check /path/to/recording.mcap
 physlint inspect /path/to/recording.mcap --profile ros2
 ```
 
-Write JSON to an exact destination:
+Write JSON, JUnit, SARIF, or HTML to exact destinations:
 
 ```bash
 physlint check /path/to/lerobot-dataset \
   --output json \
-  --json-output artifacts/physlint-report.json
+  --json-output artifacts/physlint-report.json \
+  --junit-output artifacts/physlint.junit.xml \
+  --sarif-output artifacts/physlint.sarif \
+  --html-output artifacts/physlint.html
+```
+
+Compare two dataset versions or previously written reports:
+
+```bash
+physlint compare /path/to/previous /path/to/current
+physlint compare artifacts/previous.json artifacts/current.json --output json
+```
+
+Accept known findings into a reviewed baseline. New fingerprints of the same rule still fail:
+
+```bash
+physlint baseline /path/to/dataset --author ada --reason "station clock offset" --output .physlint/baseline.yaml
+physlint check /path/to/dataset --baseline .physlint/baseline.yaml
 ```
 
 The source remains untouched. Exit code `0` means the configured contract passed; `1` means validation completed with a blocking finding.
@@ -135,7 +154,7 @@ Physlint owns NaN and infinity semantics in `numeric.finite_values`; the same sa
 
 ## What Physlint checks
 
-Thirty-one deterministic rules are installed. Physlint plans only the rules for the detected adapter and profile:
+Thirty-two deterministic rules are installed. Physlint plans only the rules for the detected adapter and profile:
 
 | Area | Checks |
 |---|---|
@@ -144,6 +163,7 @@ Thirty-one deterministic rules are installed. Physlint plans only the rules for
 | **Temporal** | Strictly monotonic timestamps, FPS cadence, FPS-aware maximum gaps, complete stream overlap, and observation/action delay when independently timestamped |
 | **Numeric** | NaN/Inf, configured physical bounds, and configured discontinuity limits |
 | **Video** | Complete decode, motion-aware frozen-frame runs, and grouped black/near-empty frames |
+| **Duplication** | Exact numeric episode copies |
 | **MCAP** | Record/CRC readability, summary consistency, index coverage, channel/schema coherence, timestamp order, duplicate timestamps, and sequence continuity |
 | **ROS 2** | Portable CDR schemas, full decode, required topics, cadence gaps, configured header/log skew, known message invariants, and TF parent consistency |
 
@@ -158,7 +178,7 @@ physlint explain video.frozen_frames
 
 Rules whose required inputs are unavailable return `not_run` with a reason; they are never misreported as passed. Robot-specific bounds and discontinuity checks stay `not_run` until the user supplies meaningful thresholds.
 
-Read the complete [LeRobot MVP rules](docs/rules/mvp-rules.md) and [MCAP/ROS 2 rules](docs/rules/mcap-ros2-rules.md).
+Read the complete [LeRobot MVP rules](docs/rules/mvp-rules.md), [MCAP/ROS 2 rules](docs/rules/mcap-ros2-rules.md), and [plugin SDK](docs/plugins.md). Task-specific checks stay out of the default contract; load them with `plugins:` or a `physlint.rules` entry point.
 
 For ROS 2 contracts, configure the topics and timing assumptions that Physlint cannot infer honestly:
 
@@ -220,7 +240,16 @@ rules:
 
 reports:
   json: true
+  junit: false
+  sarif: false
+  html: false
   output_dir: .physlint/reports
+cache:
+  enabled: true
+  directory: .physlint/cache
+# baseline: .physlint/baseline.yaml
+# plugins:
+#   - ./rules/idle_prefix.py:IdlePrefixRule
 ```
 
 Use it explicitly when needed:
@@ -237,12 +266,14 @@ The CLI has stable exit codes and writes reports atomically, so a basic GitHub A
 
 ```yaml
 - name: Install Physlint
-  run: python -m pip install "physlint[video]==0.1.0a1"
+  run: python -m pip install "physlint[video]==0.3.2"
 
 - name: Validate robot dataset
   run: |
     physlint check "$DATASET_PATH" \
-      --json-output artifacts/physlint-report.json
+      --json-output artifacts/physlint-report.json \
+      --junit-output artifacts/physlint.junit.xml \
+      --sarif-output artifacts/physlint.sarif
 
 - uses: actions/upload-artifact@v4
   if: always()
@@ -254,7 +285,7 @@ The CLI has stable exit codes and writes reports atomically, so a basic GitHub A
 | Exit code | Meaning |
 |---:|---|
 | `0` | Validation completed and the configured contract passed |
-| `1` | Validation completed and the contract failed |
+| `1` | Validation completed and the contract failed, or `compare` found a regression |
 | `2` | Invalid command or configuration |
 | `3` | Dataset or adapter failure |
 | `4` | Internal Physlint error |
@@ -284,6 +315,8 @@ Everything needed to audit or rerun those claims is versioned:
 
 Runtime measurements are observations from the documented machine and run—not universal performance guarantees.
 
+A larger [compatibility survey](docs/validation/compatibility-survey.md) adds ten revision-pinned public LeRobot v3 snapshots (1,678 episodes, 182,128 frames) plus five fail-closed v2 screens. Five of the ten v3 snapshots passed the default contract; the rest are classified in that report. Those rows are Observatory survey evidence. They do not expand the four-snapshot release claim and they are not a Hugging Face quality score.
+
 The MCAP/ROS 2 release gate additionally verifies an exact Foxglove conformance fixture, a public robot-captured RobotisAI episode, and two deterministic ROS 2 recordings:
 
 | Recording | Profile | Provenance | Checks | Expected outcome |
@@ -297,7 +330,7 @@ See the [MCAP/ROS 2 manifest](validation/mcap_manifest.yaml), [reproduction harn
 
 ## Physlint Observatory
 
-The repository now includes the first [Physlint Observatory](observatory/): a profile-aware public evidence index spanning LeRobot, MCAP, and ROS 2. It deliberately does not collapse unlike contracts into one universal quality score. Every row exposes provenance, applicable checks, findings, and a report link.
+The repository now includes the [Physlint Community Evidence Atlas](observatory/): a profile-aware public map spanning LeRobot, MCAP, and ROS 2, with real defect stories and a contribution path for dataset maintainers. It deliberately does not collapse unlike contracts into one universal quality score. Every row exposes provenance, applicable checks, findings, a stable share link, and a sanitized report.
 
 ## Format roadmap
 
@@ -389,6 +422,7 @@ Contributions are welcome, particularly:
 - False-positive reproductions
 - MCAP/ROS recording schemas and failure modes
 - New deterministic rules with controlled corruptions
+- Plugin rules that follow the [plugin SDK](docs/plugins.md)
 - Documentation, performance characterization, and privacy reviews
 
 Development setup:
@@ -410,7 +444,7 @@ Read [CONTRIBUTING.md](CONTRIBUTING.md), open a format request, or join [GitHub
 
 ## Project status
 
-Physlint is an alpha. Its claims are deliberately limited to the documented LeRobot, generic MCAP, and ROS 2-over-MCAP boundaries and their committed evidence. The project does not train policies, repair data, host datasets, infer task success, produce an opaque quality score, or certify that a robot or policy is safe.
+Physlint is released. Its claims are deliberately limited to the documented LeRobot, generic MCAP, and ROS 2-over-MCAP boundaries and their committed evidence. The project does not train policies, repair data, host datasets, infer task success, produce an opaque quality score, or certify that a robot or policy is safe.
 
 See [CHANGELOG.md](CHANGELOG.md) for release notes.
 

{physlint-0.2.0a2 → physlint-0.3.2}/README.md

@@ -11,17 +11,19 @@ Local-first, deterministic validation for physical-AI recordings and robot-learn
 [![Downloads per month](https://img.shields.io/endpoint?url=https%3A%2F%2Fraw.githubusercontent.com%2Fpranayjoshi%2Fphyslint%2Fmaster%2Fdocs%2Fmetrics%2Fdownloads.json&style=flat-square)](https://pypistats.org/packages/physlint)
 ![Python](https://img.shields.io/badge/Python-3.11%2B-3776AB?style=flat-square&logo=python&logoColor=white)
 [![MIT](https://img.shields.io/badge/License-MIT-2ea44f?style=flat-square)](LICENSE)
-![Status](https://img.shields.io/badge/Status-Public%20Alpha-f59e0b?style=flat-square)
+![Status](https://img.shields.io/badge/Status-Released-0f766e?style=flat-square)
 
 [Quickstart](#quickstart) · [Demo](#see-it-catch-a-real-defect) · [Rules](#what-physlint-checks) · [Evidence](#reproducible-public-data-evidence) · [Observatory](#physlint-observatory) · [Roadmap](#format-roadmap) · [Contributing](#contributing)
 
-<img src="docs/assets/launch/exports/validation-summary.png" alt="Physlint public-alpha validation: four of four clean snapshots passed and three of three controlled corruptions were detected across 74 episodes and 31,258 frames." width="100%">
+<img src="docs/assets/launch/exports/evidence-index.png" alt="Physlint evidence index showing robot dataset and recording health checks across LeRobot, MCAP, and ROS 2 data." width="100%">
 
 </div>
 
 Physlint finds concrete integrity defects before robot data reaches training. It explains the impact, identifies the affected episode and stream, recommends remediation, writes a stable JSON report, and returns a CI-safe exit code.
 
-The released `0.1.0a1` public alpha supports **LeRobot Dataset v3.x**. The current `0.2.0a1` development line adds generic **MCAP** container validation and a **ROS 2-over-MCAP** semantic profile, backed by pinned conformance data, a real robot-captured ROS 2 episode, and controlled evidence. Native rosbag2 SQLite remains a conversion workflow for now.
+The released `0.3.2` supports **LeRobot Dataset v3.x**, generic **MCAP** container validation, and a **ROS 2-over-MCAP** semantic profile, plus dataset comparison, reviewed baselines, CI report formats, and a plugin SDK. It also launches the [Community Evidence Atlas](https://physlint.com), where maintainers can explore reproducible findings and submit public robot data for inclusion. Claims stay scoped to pinned conformance data and controlled evidence. Native rosbag2 SQLite remains a conversion workflow.
+
+At the `0.3.2` release snapshot, PyPI Stats reported **289 downloads in the preceding month**.
 
 > [!IMPORTANT]
 > Physlint validates configured data-integrity contracts. A pass does not certify policy quality, task success, or robot safety.
@@ -41,13 +43,13 @@ Physlint requires Python 3.11 or newer.
 ### Install from PyPI
 
 ```bash
-python -m pip install "physlint[video]==0.1.0a1"
+python -m pip install "physlint[video]==0.3.2"
 ```
 
 To test the exact tagged source instead of the PyPI distribution, install the GitHub release directly:
 
 ```bash
-python -m pip install "physlint[video] @ git+https://github.com/pranayjoshi/physlint.git@v0.1.0a1"
+python -m pip install "physlint[video] @ git+https://github.com/pranayjoshi/physlint.git@v0.3.2"
 ```
 
 ### Check a LeRobot dataset or MCAP recording
@@ -63,12 +65,29 @@ physlint check /path/to/recording.mcap
 physlint inspect /path/to/recording.mcap --profile ros2
 ```
 
-Write JSON to an exact destination:
+Write JSON, JUnit, SARIF, or HTML to exact destinations:
 
 ```bash
 physlint check /path/to/lerobot-dataset \
   --output json \
-  --json-output artifacts/physlint-report.json
+  --json-output artifacts/physlint-report.json \
+  --junit-output artifacts/physlint.junit.xml \
+  --sarif-output artifacts/physlint.sarif \
+  --html-output artifacts/physlint.html
+```
+
+Compare two dataset versions or previously written reports:
+
+```bash
+physlint compare /path/to/previous /path/to/current
+physlint compare artifacts/previous.json artifacts/current.json --output json
+```
+
+Accept known findings into a reviewed baseline. New fingerprints of the same rule still fail:
+
+```bash
+physlint baseline /path/to/dataset --author ada --reason "station clock offset" --output .physlint/baseline.yaml
+physlint check /path/to/dataset --baseline .physlint/baseline.yaml
 ```
 
 The source remains untouched. Exit code `0` means the configured contract passed; `1` means validation completed with a blocking finding.
@@ -92,7 +111,7 @@ Physlint owns NaN and infinity semantics in `numeric.finite_values`; the same sa
 
 ## What Physlint checks
 
-Thirty-one deterministic rules are installed. Physlint plans only the rules for the detected adapter and profile:
+Thirty-two deterministic rules are installed. Physlint plans only the rules for the detected adapter and profile:
 
 | Area | Checks |
 |---|---|
@@ -101,6 +120,7 @@ Thirty-one deterministic rules are installed. Physlint plans only the rules for
 | **Temporal** | Strictly monotonic timestamps, FPS cadence, FPS-aware maximum gaps, complete stream overlap, and observation/action delay when independently timestamped |
 | **Numeric** | NaN/Inf, configured physical bounds, and configured discontinuity limits |
 | **Video** | Complete decode, motion-aware frozen-frame runs, and grouped black/near-empty frames |
+| **Duplication** | Exact numeric episode copies |
 | **MCAP** | Record/CRC readability, summary consistency, index coverage, channel/schema coherence, timestamp order, duplicate timestamps, and sequence continuity |
 | **ROS 2** | Portable CDR schemas, full decode, required topics, cadence gaps, configured header/log skew, known message invariants, and TF parent consistency |
 
@@ -115,7 +135,7 @@ physlint explain video.frozen_frames
 
 Rules whose required inputs are unavailable return `not_run` with a reason; they are never misreported as passed. Robot-specific bounds and discontinuity checks stay `not_run` until the user supplies meaningful thresholds.
 
-Read the complete [LeRobot MVP rules](docs/rules/mvp-rules.md) and [MCAP/ROS 2 rules](docs/rules/mcap-ros2-rules.md).
+Read the complete [LeRobot MVP rules](docs/rules/mvp-rules.md), [MCAP/ROS 2 rules](docs/rules/mcap-ros2-rules.md), and [plugin SDK](docs/plugins.md). Task-specific checks stay out of the default contract; load them with `plugins:` or a `physlint.rules` entry point.
 
 For ROS 2 contracts, configure the topics and timing assumptions that Physlint cannot infer honestly:
 
@@ -177,7 +197,16 @@ rules:
 
 reports:
   json: true
+  junit: false
+  sarif: false
+  html: false
   output_dir: .physlint/reports
+cache:
+  enabled: true
+  directory: .physlint/cache
+# baseline: .physlint/baseline.yaml
+# plugins:
+#   - ./rules/idle_prefix.py:IdlePrefixRule
 ```
 
 Use it explicitly when needed:
@@ -194,12 +223,14 @@ The CLI has stable exit codes and writes reports atomically, so a basic GitHub A
 
 ```yaml
 - name: Install Physlint
-  run: python -m pip install "physlint[video]==0.1.0a1"
+  run: python -m pip install "physlint[video]==0.3.2"
 
 - name: Validate robot dataset
   run: |
     physlint check "$DATASET_PATH" \
-      --json-output artifacts/physlint-report.json
+      --json-output artifacts/physlint-report.json \
+      --junit-output artifacts/physlint.junit.xml \
+      --sarif-output artifacts/physlint.sarif
 
 - uses: actions/upload-artifact@v4
   if: always()
@@ -211,7 +242,7 @@ The CLI has stable exit codes and writes reports atomically, so a basic GitHub A
 | Exit code | Meaning |
 |---:|---|
 | `0` | Validation completed and the configured contract passed |
-| `1` | Validation completed and the contract failed |
+| `1` | Validation completed and the contract failed, or `compare` found a regression |
 | `2` | Invalid command or configuration |
 | `3` | Dataset or adapter failure |
 | `4` | Internal Physlint error |
@@ -241,6 +272,8 @@ Everything needed to audit or rerun those claims is versioned:
 
 Runtime measurements are observations from the documented machine and run—not universal performance guarantees.
 
+A larger [compatibility survey](docs/validation/compatibility-survey.md) adds ten revision-pinned public LeRobot v3 snapshots (1,678 episodes, 182,128 frames) plus five fail-closed v2 screens. Five of the ten v3 snapshots passed the default contract; the rest are classified in that report. Those rows are Observatory survey evidence. They do not expand the four-snapshot release claim and they are not a Hugging Face quality score.
+
 The MCAP/ROS 2 release gate additionally verifies an exact Foxglove conformance fixture, a public robot-captured RobotisAI episode, and two deterministic ROS 2 recordings:
 
 | Recording | Profile | Provenance | Checks | Expected outcome |
@@ -254,7 +287,7 @@ See the [MCAP/ROS 2 manifest](validation/mcap_manifest.yaml), [reproduction harn
 
 ## Physlint Observatory
 
-The repository now includes the first [Physlint Observatory](observatory/): a profile-aware public evidence index spanning LeRobot, MCAP, and ROS 2. It deliberately does not collapse unlike contracts into one universal quality score. Every row exposes provenance, applicable checks, findings, and a report link.
+The repository now includes the [Physlint Community Evidence Atlas](observatory/): a profile-aware public map spanning LeRobot, MCAP, and ROS 2, with real defect stories and a contribution path for dataset maintainers. It deliberately does not collapse unlike contracts into one universal quality score. Every row exposes provenance, applicable checks, findings, a stable share link, and a sanitized report.
 
 ## Format roadmap
 
@@ -346,6 +379,7 @@ Contributions are welcome, particularly:
 - False-positive reproductions
 - MCAP/ROS recording schemas and failure modes
 - New deterministic rules with controlled corruptions
+- Plugin rules that follow the [plugin SDK](docs/plugins.md)
 - Documentation, performance characterization, and privacy reviews
 
 Development setup:
@@ -367,7 +401,7 @@ Read [CONTRIBUTING.md](CONTRIBUTING.md), open a format request, or join [GitHub
 
 ## Project status
 
-Physlint is an alpha. Its claims are deliberately limited to the documented LeRobot, generic MCAP, and ROS 2-over-MCAP boundaries and their committed evidence. The project does not train policies, repair data, host datasets, infer task success, produce an opaque quality score, or certify that a robot or policy is safe.
+Physlint is released. Its claims are deliberately limited to the documented LeRobot, generic MCAP, and ROS 2-over-MCAP boundaries and their committed evidence. The project does not train policies, repair data, host datasets, infer task success, produce an opaque quality score, or certify that a robot or policy is safe.
 
 See [CHANGELOG.md](CHANGELOG.md) for release notes.
 

{physlint-0.2.0a2 → physlint-0.3.2}/docs/metrics/downloads.json

@@ -1,9 +1,9 @@
 {
   "schemaVersion": 1,
   "label": "downloads/month",
-  "message": "85",
+  "message": "289",
   "color": "brightgreen",
   "cacheSeconds": 86400,
-  "updated_at": "2026-08-27T19:40:47.363603+00:00",
+  "updated_at": "2026-09-07T02:45:03.175664+00:00",
   "source": "https://pypistats.org/api/packages/physlint/recent?period=month"
 }

physlint-0.3.2/docs/plugins.md

@@ -0,0 +1,44 @@
+# Plugin SDK
+
+Physlint loads extra rules without putting task-specific heuristics into the default contract. A plugin is a class with `metadata: RuleMetadata` and `run(dataset, options, severity) -> list[Finding]`.
+
+## Load from a quality contract
+
+```yaml
+config_version: 1
+plugins:
+  - ./rules/idle_prefix.py:IdlePrefixRule
+```
+
+The spec is `module:Class` or `path.py:Class`. Unknown option keys are rejected. Plugin rule IDs must not collide with built-ins.
+
+```bash
+physlint rules --config physlint.yaml
+physlint explain example.idle_prefix --config physlint.yaml
+physlint check /path/to/dataset --config physlint.yaml
+```
+
+## Install via entry points
+
+Packaged plugins can register:
+
+```toml
+[project.entry-points."physlint.rules"]
+idle_prefix = "my_org_rules:IdlePrefixRule"
+```
+
+Installed entry points are loaded on every run, then config `plugins:` are appended.
+
+## Contract
+
+- Read-only. Never mutate source data.
+- Deterministic. The same dataset, options, and Physlint version must produce the same fingerprints.
+- Bounded. Honor `max_findings` and avoid embedding images or full samples.
+- Honest. If required capabilities are missing, the engine records `not_run`; do not pass.
+- Documented. Include remediation, limitations, and option defaults.
+
+Set `metadata.adapters` to restrict a plugin to `lerobot` or `mcap`. Leave it empty to participate on every adapter that satisfies the required capabilities.
+
+## Example
+
+`examples/plugins/idle_prefix.py` flags long zero-action prefixes. It is a warning, not a built-in, because idle length is a collection policy rather than a universal integrity failure.

{physlint-0.2.0a2 → physlint-0.3.2}/docs/rules/mvp-rules.md

@@ -36,3 +36,4 @@ Fails for non-positive episode lengths, reversed ranges, a mismatch between decl
 - `numeric.discontinuity`: adjacent values respect configured maximum absolute deltas.
 - `video.decode`: decoded frame count matches episode length within tolerance.
 - `video.black_frames`: frame mean and standard deviation are not both near zero; adjacent bad frames are grouped.
+- `duplication.exact_episodes`: numeric observation/action trajectories are unique across episodes.

physlint-0.3.2/docs/validation/compatibility-survey.md

@@ -0,0 +1,84 @@
+# LeRobot compatibility survey
+
+Status: first wave, 2026-08-30  
+Tier: compatibility observations — **not the public release gate**
+
+The release gate remains the four pinned snapshots in [`validation/manifest.yaml`](../../validation/manifest.yaml). This survey asks a different question: does the LeRobot v3 adapter stay honest across embodiments, producers, and cadences that those four rows do not cover?
+
+It does not attempt to validate Hugging Face robotics data. Hub-scale crawls measure repository hygiene, not Physlint trust.
+
+## Method
+
+1. Screen Hub candidates from metadata only (`meta/info.json`, size, version, robot, FPS, license/gate).
+2. Keep a snapshot only when it fills an empty diversity cell and is small enough to review findings by hand.
+3. Pin the exact revision. Download into `validation/.work/survey/`, which is gitignored.
+4. Run the default `physlint check`. Record sanitized reports. Do not require a pass.
+5. For published v2.x layouts, download only `meta/info.json` and prove discovery fails closed.
+6. Classify findings against source values before promoting any row into the release gate.
+
+## First-wave cells
+
+| Cell | Snapshot |
+|---|---|
+| Official 10 FPS / 2D | `lerobot/pusht` |
+| Official xArm / 15 FPS | `lerobot/xarm_lift_medium` |
+| Official humanoid / 50 FPS | `lerobot/unitreeh1_two_robot_greeting` |
+| Official SO-100 / 30 FPS | `lerobot/svla_so100_pickplace` |
+| Community bimanual / 50 FPS | `macrodata/aloha_static_battery_ep000_004` |
+| Industrial arm / 6 FPS | `wenyixu101/ur10e-robotiq-2f85` |
+| Community SO-101 / 25 FPS | `wenyixu101/so101-pick-and-place` |
+| Reachy / 60 FPS | `tavis-benchmark/tavis-head-sample-reachy2` |
+| Mobile base / 30 FPS | `vietnguyen28/lekiwi_go2` |
+| NVIDIA Franka suite | `nvidia/LIBERO_LeRobot_v3` (`libero_spatial`) |
+| Fail-closed v2.0 / v2.1 | Koch, Reachy kitchen, RoboTwin, large SO-101 |
+
+Deferred on purpose: official mobile Aloha with `observation.effort` (8.6 GB), 1080p LeKiwi (3.9 GB), 10 GB SO-101 sim, 60 GB BridgeData V2, and multi-terabyte Hub dumps. Those belong in a later size/memory wave, not this survey.
+
+## Claims that stay false
+
+- “Physlint validated Hugging Face robotics datasets.”
+- “N / M Hub datasets pass.”
+- “N terabytes scanned.”
+
+After this wave the honest statement is: the release gate is unchanged; a stratified public survey of N revision-pinned v3 snapshots plus M version-edge screens is committed under `validation/reports/lerobot-survey-2026-08-30/`.
+
+## Reproduce
+
+```bash
+python -m pip install -e '.[video,dev]'
+python -m pip install huggingface_hub
+python -m validation.survey_harness
+```
+
+Use `--only pusht --only koch-v21` to rerun one cell. Pass `--dataset-root ID=PATH` for an offline snapshot. The survey writes sanitized JSON next to `summary.json`; it does not update README release-gate counts.
+
+## Results
+
+First wave, Physlint `0.3.0`, run `lerobot-survey-2026-08-30`.
+
+Checked snapshots: **5 passed, 5 failed, 0 errored** after the string-`names` adapter fix. Version-edge screens: **5/5 rejected**. Coverage: 1,678 episodes and 182,128 frames. The four-snapshot release gate is unchanged.
+
+| Snapshot | Result | Applicable | Findings | Classification |
+|---|---|---:|---:|---|
+| `lerobot/pusht` | Fail | 15 | 50 (capped) | Likely false positive. 2D low-texture sim; 6–8 frame freezes with `motion_fraction` 1.0 against a 0.5 pixel MAD. |
+| `lerobot/xarm_lift_medium` | Pass | 15 | 0 | Clean official xArm / 15 FPS. |
+| `lerobot/unitreeh1_two_robot_greeting` | Pass | 15 | 0 | Clean official humanoid / 50 FPS. |
+| `lerobot/svla_so100_pickplace` | Pass | 15 | 0 | Clean official SO-100 / 30 FPS. |
+| `macrodata/aloha_static_battery_ep000_004` | Pass | 15 | 0 | Clean community bimanual / 50 FPS. |
+| `wenyixu101/ur10e-robotiq-2f85` | Fail | 15 | 4 | Review. Four 6-frame holds at 6 FPS (~1 s) during motion; possible sim hold. |
+| `wenyixu101/so101-pick-and-place` | Fail | 15 | 50 (capped) | Likely false positive. All cited ranges are the wrist camera, often 40+ frames, `motion_fraction` ~0.4. |
+| `tavis-benchmark/tavis-head-sample-reachy2` | Fail | 5 | 3 | Confirmed defect. Episode metadata names `file-001/002/004.parquet`; the Hub tree only publishes `file-000.parquet`. |
+| `vietnguyen28/lekiwi_go2` | Pass | 15 | 0 | Clean community mobile base. |
+| `nvidia/LIBERO_LeRobot_v3` `libero_spatial` | Fail | 15 | 1 | Review. One 6-frame freeze at an episode start, `motion_fraction` 0.4. |
+
+### Version-edge screens
+
+Discovery rejected `v2.0` and `v2.1` metadata for Koch, Reachy kitchen, RoboTwin, and the large SO-101 pi0.5 corpus. That is the required fail-closed behavior.
+
+### Product issues
+
+1. **Fixed — string feature names.** Reachy stores `action.names` as `"env actions"`. Discovery previously raised a Pydantic `ValidationError`. The adapter now coerces a string name to a one-element list and wraps other invalid feature declarations as `AdapterError`.
+2. **Open — frozen-frame defaults on 2D / wrist / episode-start holds.** Motion correlation is not enough when pixel MAD stays below 0.5. Do not change the default contract from this wave alone; gather more wrist-camera and 2D-sim cases first.
+3. **Open — 50-finding cap.** PushT and the SO-101 teleop set both hit the cap, so the true freeze count is unknown.
+
+No failing survey row is promoted into the release gate. Sanitized reports are in [`validation/reports/lerobot-survey-2026-08-30/`](../../validation/reports/lerobot-survey-2026-08-30/).