robotruth 0.1.6__tar.gz

robotruth-0.1.6/.gitignore

@@ -0,0 +1,18 @@
+.venv/
+__pycache__/
+*.pyc
+*.egg-info/
+dist/
+build/
+.pytest_cache/
+.ruff_cache/
+reports/
+data/
+*.safetensors
+*.ckpt
+*.pt
+*.pth
+.env
+
+# policyci pulled results (large: videos, episode logs)
+results/

robotruth-0.1.6/CHANGELOG.md

@@ -0,0 +1,58 @@
+# Changelog
+
+## 0.1.6 (2026-09-21)
+
+- First release published to PyPI, so `pip install robotruth` works. Publishing goes through PyPI trusted publishing from GitHub Actions, which means no API token exists in this repository or its secrets.
+- **Baseline study.** Our monitor being at chance on real task failures turned out to be a fact about the signal rather than about our scorers. Nine other detectors covering the families used in recent work, the chunk consistency of VLA-FAIL, the distance, density and action-variance signals of FAIL-Detect, random network distillation, a GRU sequence model, and isolation forest, one-class SVM and PCA reconstruction, were each fitted on successful episodes only and scored on the same held-out episodes. All ten catch injected execution faults. All ten are at chance on the failures the robots actually had, with AUROC from 0.31 to 0.57 across BotFails and DROID.
+- Nine of the twenty detector and corpus pairs land below 0.5, which is not the same as no signal: the score runs backwards. On DROID a failed episode runs 188 steps against 239 and covers 5.2 units of total motion against 7.6, so a policy that fails does less and any score built on movement reads failure as quiet and quiet as normal. A monitor tuned on that data would spend its alarm budget on the successes.
+- `ops/baseline_comparison.py` runs the comparison and `ops/why_below_chance.py` measures the mechanism; full output is in `examples/validation/2026-09-20-paper`.
+
+## 0.1.5 (2026-09-20)
+
+- **The guard's scope is corrected, and the correction is measured.** Replayed against real robot logs it caught 5/145 of BotFails' own failures and 0/58 of DROID's. A threshold-free AUROC per scorer head came out at 0.50 to 0.51 and 0.33 to 0.44, so there is no signal to recover: a robot that fails a task keeps moving normally, and the action stream does not show it. The guard detects execution faults, where the robot stops behaving like itself, and the README no longer implies otherwise.
+- **That capability does transfer to real robots.** Stalls and erratic control were detected in 153/153, 72/72 and 136/137 real episodes across three physical-robot corpora within 0.1 to 0.2 s, with false alarms inside the 0.05 bound. Constant offsets need about two standard deviations, and on one corpus were missed at every magnitude tried.
+- **False alarms measured on a pool that can carry the claim**: 0/338 held-out successes, interval [0.000, 0.011]. The 0.1.4 bound rested on 18 episodes and its interval reached 0.176, wide enough to hide a 17 percent true rate.
+- **Gradual faults are detected too.** Every previous number used a step change. Faults ramping to full strength over 0.5 to 4 s are still detected 100% of the time, and the delay tracks how long the fault takes to leave the nominal envelope rather than being a fixed lag.
+- **`robotruth.fingerprint.DriftDetector`**: a calibrated drift alarm for when you have a population of fingerprints instead of a known tolerance. Median and scaled MAD for the centre and spread, so one already-drifted unit cannot mask the drift it represents, and a split-conformal threshold on the worst standardised per-joint deviation. On eight real datasets and 270 held-out clean episodes it gives 3/270 false alarms against 23/270 for a three-sigma rule, with detection of a 0.25 sd offset unchanged at 268/270. It reports itself saturated rather than claiming a rate too few fingerprints can certify.
+- The policy loader reads min/max scaling and the newer lerobot layout that keeps statistics in separate preprocessor files, taking the scheme from the policy's own config instead of guessing.
+- Known limit, unchanged and now stated plainly: every live guard number comes from one policy on one task. Three attempts at a second combination failed, one of them because `lerobot/diffusion_pusht` raises an uncorrectable ECC error on an A10 and poisons the GPU for other processes.
+
+## 0.1.4 (2026-09-19)
+
+- **The stall blind spot is closed.** A `StagnationScorer` watches windowed motion of the action stream, and `calibrate_guard` now gives it its own conformal threshold in a `MultiHeadScorer` instead of averaging it into the composite, alarming on the union of the two heads with alpha split between them. Measured on the same live policy and fault protocol as 0.1.3, on a 150-episode calibration pool: stall, offset and noise faults all detected 10/10, pooled 30/30 = 1.000 [0.886, 1.000], median latency 0.04 s, 0/18 false alarms on held-out successes.
+- The dedicated head is the reason it works. With the stagnation score averaged into the composite, freeze detection was only 1/10: a frozen action stream is self-consistent and in-distribution, so three of the four component scores report nominal and dilute the one that sees the stall. Both runs are published side by side in `examples/validation/2026-09-19/guard_detection_v2/`.
+- The stall head reports itself as saturated below roughly 40 calibration episodes rather than claiming an uncertifiable guarantee.
+- **Cross-dataset benchmark**: `ops/benchmark_fetch.py`, `ops/benchmark_report.py` and `ops/botfails_bench.py` fetch public LeRobot datasets (metadata and parquet only), run every module over them and render one HTML report. First run covers 54 datasets from RoboMIND and AgiBotWorld task ports, OpenX conversions, DROID, BotFails and DAgger logs, with zero ingest failures.
+- **HTML reports redesigned**: markdown in report sections now renders as HTML instead of being dumped as raw text, with a document header, table of contents, numbered sections, sentence-cased headings and column headers, and print styles.
+- `ops/guard_detection.py` saves every rollout as npz, so threshold changes can be replayed offline instead of costing GPU hours.
+
+## 0.1.3 (2026-09-19)
+
+- Guard detection measured with policy-breaking faults injected into executed actions on a live policy (all 30 fault episodes failed the task). On a 164-episode calibration pool: offset (miscalibration) faults 10/10 detected, noise (erratic policy) faults 10/10 with max thresholds, median detection latency 0.04 s, 0/17 false alarms on held-out successes. Stall (freeze) faults 0/10: a frozen action stream is self-consistent and in-distribution, so the current scorers cannot see it; a stagnation scorer is the planned fix and the blind spot is documented in Known limits.
+- Calibration-size guidance is now measured: 45-episode pools leaked up to 3/17 false alarms in the same experiment while the 164-episode pool leaked none. Calibrate on 150 or more nominal episodes.
+
+## 0.1.2 (2026-09-19)
+
+- Fused judge validated on real labeled data: on 323 BotFails episodes the fused and calibrated judge reaches 0.921 [0.617, 0.972] balanced accuracy at 25.8 percent coverage with 2.73 false alarms per hour, against 0.614 and 11.13 for the uncalibrated vision channel alone. The action channel alone abstains on everything, which is why fusion is the shipped default.
+- Guard conformal methods measured across three independent calibration draws on a live policy. The bonferroni method alarmed on about 20 percent of held-out successes at 44 to 49 calibration episodes because every time bin was saturated, so its alpha did not hold; `max` remains the default and the caveat is documented in `guard/conformal.py`. Detection remains unmeasured because the perturbation broke the policy in only 1 of 12 episodes.
+- Memory-safe frame sampling in the evaluation scripts: sampling frames from a decoded video kept the whole video alive through numpy views, which grew to 140 GB across a dataset. Frames are now streamed and copied.
+- Em dashes removed from all documentation.
+
+## 0.1.1 (2026-09-18)
+
+- Open-weight VLM backend for the judge (`robotruth.judge.open_vlm.OpenVLMBackend`, Qwen-VL class via transformers): the vision channel runs on your own GPU with no API key. Benchmarked zero-shot on 140 labeled real UR5 episodes: balanced accuracy 0.555 [0.442, 0.663] at 3.5 s per episode on one A10 (frontier ceiling on FailBench: 0.77). The Claude backend remains available behind the `vlm` extra.
+- Live policy-loop validation: a real ACT policy in gym-aloha with the guard attached to every step and the judge scored against simulator truth. Judge balanced accuracy 0.929 [0.651, 0.987] over 24 live episodes with zero false alarms; guard false-alarm bound held in both runs; run-to-run threshold variance at small calibration sizes documented.
+- openpi extractor validated against the real pi0_base orbax checkpoint (12 GB hashed; norm_stats for nine robot assets detected and flagged as part of the executable policy).
+- The live test reproduced the contract failure class in the wild: lerobot 0.6.1 silently drops this checkpoint's normalization buffers, taking an 83 percent policy to 0 percent with only a log warning.
+
+## 0.1.0 (2026-09-18)
+
+First public release.
+
+- Contract checker: ExecSpec manifests for LeRobot (old and new preprocessor formats), GR00T, openpi and custom checkpoints; probe-directory source so weights never move; fail-closed comparison. Validated against five public checkpoints (ACT, SmolVLA, pi0, pi05, GR00T N1.5).
+- Statistics: Wilson and Clopper-Pearson intervals, paired and bootstrap differences, anytime-valid empirical Bernstein confidence sequences, sequential paired test, sample-size planner, Kaplan-Meier and log-rank on censored time to success, Bradley-Terry with bootstrap errors, blinded interleaved schedules, claims audit (Newcombe intervals and Fisher tests over reported counts).
+- Episode records: schema with 12 failure classes and 60 subclasses, interventions with source, provenance hashes; JSONL log; MCAP bridge; fleet metrics with intervals (success, autonomous fraction, interventions per hour, MTBI, failure Pareto); LeRobot dataset ingest (v2 and v3, success and intervention columns, DROID column aliases).
+- Fingerprints: robot unit fingerprint from an excitation log (joint lag and backlash by regression, RMSE, offset, gain) and workspace cell fingerprint from a camera frame (ArUco markers, exposure, sharpness, colour balance) with drift tolerances.
+- Judge: action-stream features, VLM backend protocol with a Claude backend, logistic fusion, split-conformal calibration with abstain, evaluation with balanced accuracy and false alarms per hour, FailBench-style loader.
+- Guard: Mahalanobis (Ledoit-Wolf, optional CUDA), chunk-consistency and action-stat scorers, time-uniform conformal thresholds (max and Bonferroni), contrast-set calibration, hysteresis and hard limits, offline replay, metrics.
+- Ops: GPU-box scripts for checkpoint probing, dataset fetching, sync-and-test, and end-to-end validation on public LeRobot datasets.

robotruth-0.1.6/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.