fieldtrial 0.1.0.dev0__tar.gz → 0.2.0__tar.gz

fieldtrial-0.2.0/.gitignore

@@ -0,0 +1,39 @@
+# Secrets: tokens and credentials stay local
+.env
+.env.*
+!.env.example
+.pypirc
+
+# Python bytecode
+__pycache__/
+*.py[cod]
+
+# Virtual environments (uv)
+.venv/
+venv/
+
+# Build artifacts (uv build / hatchling)
+build/
+dist/
+*.egg-info/
+*.whl
+
+# Test, lint and type-check caches
+.pytest_cache/
+.hypothesis/
+.mypy_cache/
+.ruff_cache/
+.import_linter_cache/
+.coverage
+.coverage.*
+htmlcov/
+coverage.xml
+
+# Docs build (mkdocs)
+/site/
+
+# Editors and OS clutter
+.idea/
+.vscode/
+.DS_Store
+Thumbs.db

fieldtrial-0.2.0/CHANGELOG.md

@@ -0,0 +1,181 @@
+# Changelog
+
+All notable changes to this project are documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this
+project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html). Before 1.0,
+minor releases may contain breaking changes.
+
+## [Unreleased]
+
+## [0.2.0] - 2026-10-02
+
+Second release: crossover rounds, checkpoint ladders and group-sequential stopping; real
+blinding with switching runners; evaluation-camera capture, rig drift checks and links to
+LeRobot datasets. Studies, design hashes and results files from 0.1 keep working.
+
+### Added
+
+- Crossover rounds (`design.type: crossover_rounds`) for tasks whose scene carries over
+  between trials. Each arm runs whole rounds in cycles of two, in randomized, balanced
+  order. The analysis is a period-adjusted difference with an exact randomization test,
+  and its interval inverts the same test. A new `crossover-rounds` template. The console
+  shows the round and when to reset the scene.
+- Checkpoint ladders (`analysis.ladder`): Mantel's test of a linear association between
+  training step and success, and plateau detection by fixed-sequence non-inferiority
+  against the final checkpoint. Without a comparison, the association test is the
+  primary analysis. Reports gain a success-by-step chart and a plateau table.
+- Group-sequential stopping (`analysis.stopping: {rule: group_sequential, looks: K}`):
+  - Lan–DeMets error-spending boundaries (O'Brien–Fleming or Pocock type).
+  - Interim looks via `fieldtrial interim`, the console and `POST /api/v1/studies/{study}/interim`.
+    A look reveals only "continue" or "stop" and cancels the remaining trials on a stop.
+  - A stage-wise p-value and a repeated confidence interval in the final analysis.
+  - The simulator runs planned looks as they come due (`--no-interim` to skip them).
+- Real blinding with switching runners. The console and the REST API start and stop them
+  with each trial. `fieldtrial check-runners` checks the setup and names arms by blind
+  code only.
+  - `command` runner: runs `runners.command.template` for each trial with the arm's
+    `policy` and `serving` values. The template is split like a shell would but never
+    runs in one, and placeholders are checked at validation.
+    - The command gets its own process group and is stopped with SIGINT, then SIGTERM,
+      then SIGKILL. Logs go to `logs/`, and the environment carries the blind code only.
+    - `launch: per_arm` keeps one process per arm and sends it start and stop lines.
+    - With `success_exit_code`, a successful exit preselects the success stage.
+  - `openpi_router` runner: a websocket proxy in front of one openpi policy server per
+    arm.
+    - It refuses arms whose metadata differ, passes `Api-Key` headers through and forwards
+      frames unchanged to the current trial's arm.
+    - It records request latency per trial. Install with `fieldtrial[openpi]`.
+  - A runner that cannot start a trial marks it invalid with the reason, so the slot is
+    rescheduled.
+  - Reports gain a descriptive runner table per arm: requests, errors, latency and
+    abnormal exits.
+- Rig drift checks. Compare a photo of the rig with a reference photo: shift by phase
+  correlation, brightness change, and structural similarity (SSIM) after alignment.
+  - Run them with `fieldtrial rig-check DIR PHOTO` (or `--camera`), or attach a photo
+    when a console session starts.
+  - A flagged check shows as a console warning and as a report deviation, and the report
+    lists every check.
+- Evaluation-camera capture (`fieldtrial[capture]`: OpenCV and PyAV). With
+  `capture.camera` in `study.yaml`, every trial is recorded from Start to Stop and the
+  clip is attached to the trial. The rig is checked from the camera at session start and
+  every `capture.drift.every_trials` trials. A camera failure never blocks a trial.
+- Links between trials and LeRobot v3.0 dataset episodes (`fieldtrial[lerobot]`:
+  pyarrow).
+  - `fieldtrial link-episodes DIR DATASET` matches trials to episodes in run order, or
+    from a `trial,episode_index` mapping file, and shows the plan by blind code before
+    writing.
+  - Episodes with DAgger `intervention` flags get intervention counts, and reports gain a
+    descriptive dataset-episodes table.
+- A `capture:` section in `study.yaml`, always left out of the design hash.
+- `fieldtrial.stats`:
+  - `spending_boundaries`, `constant_boundaries`, `crossing_probabilities`,
+    `sequential_test`, `repeated_interval`
+  - `trend_test`, `stratified_trend_test`, `plateau`, `plateau_paired`
+  - `crossover_test`
+- `Results` gains optional `ladder`, `crossover` and `sequential` blocks, the lists
+  `runner`, `rig_checks` and `episodes` (empty by default), and the primary methods
+  `ladder`, `crossover` and `group_sequential` (schema version unchanged: additions only).
+
+### Changed
+
+- `limits.reset: carry_over` is now accepted together with `design.type: crossover_rounds`.
+- Design hashes of existing studies are unchanged: new settings are left out of the hash
+  while they are unused.
+- `pillow` is declared as a dependency; it was already required by matplotlib.
+
+## [0.2.0rc1] - 2026-10-02
+
+Release candidate of 0.2.0, published to PyPI for testing. Its changes are listed under
+0.2.0; the release is identical apart from this changelog.
+
+## [0.2.0a2] - 2026-10-02
+
+Second v0.2 pre-release: real blinding with the command runner and the openpi router. Its
+changes are listed under 0.2.0.
+
+## [0.2.0a1] - 2026-10-02
+
+First v0.2 pre-release: crossover rounds, checkpoint ladders and group-sequential stopping.
+Its changes are listed under 0.2.0.
+
+## [0.1.0] - 2026-10-02
+
+First release. Studies: design, lock, fill and analyze end to end, run them from a
+phone-friendly operator console or your own runtime, and share a self-contained report.
+
+### Added
+
+- `study.yaml` schema with line-numbered validation errors, and three templates (`basic`,
+  `checkpoint-ladder`, `serving-sweep`).
+- Randomized complete block schedules with Williams-balanced arm order and blind codes.
+  The schedule is identical on every platform and numpy version for a given seed.
+- Locking and amendments: the design, its hash and the schedule are stored in a SQLite
+  database in the study folder; every change is one transaction plus one event-log row.
+- Invalid trials are kept and rescheduled at the end of their block.
+- A simulated runner and auto-operator, and CSV import (all or nothing, `--map`) and
+  CSV/JSONL export.
+- The analysis engine: the primary analysis follows the locked design (McNemar with a Tango
+  interval, CMH, or Cochran's Q with Holm; a threshold test for single-arm studies), with
+  an independent Boschloo/Newcombe sensitivity analysis, stage funnels, time to success,
+  per-condition and per-session tables, drift and invalid-trial checks, deviations and
+  provenance. The `Results` model is versioned, and its JSON schema is published.
+- Report wording from one tested template per situation, and a Markdown report.
+- New statistics: stage distribution and funnel, Brunner–Munzel on stages, success-time
+  curve and bootstrap median time, CMH test, and homogeneity tests for drift.
+- Commands `init`, `validate`, `plan`, `lock`, `amend`, `simulate`, `import`, `export`,
+  `status`, `unblind`, `analyze` and `report`.
+- Documentation: "Running a study" and "Analysis and reports".
+- The operator console (`fieldtrial serve`): start a session with the rig checklist, run
+  trials by blind code with a timer, label the furthest stage, termination and failure tags,
+  undo within 10 seconds, mark invalid trials, correct labels with a logged reason, a live
+  mirror screen, keyboard and foot-pedal keys, light and dark themes. `--lan` serves phones
+  on the local network behind an access token with a QR code.
+- `fieldtrial demo`: a half-run simulated study in the console.
+- REST API v1 with an OpenAPI document, Server-Sent Events, idempotency keys and optimistic
+  concurrency, and `fieldtrial.client`, a dependency-free Python client.
+- Reports list label corrections made after unblinding as deviations.
+- Documentation: "Operator console", "REST API and client" and a phone test checklist.
+- A self-contained HTML report with seven charts (`fieldtrial report`, now the default
+  format, and in the console): no scripts and nothing loaded from other hosts.
+- An example re-analysis of Dream Machines' published pi0.5 fine-tuning results.
+- Documentation: quickstart, concepts, guides (planning, comparing checkpoints, ladders,
+  serving sweeps, LeRobot, openpi, custom runtimes), and screenshots.
+- `CITATION.cff`.
+
+## [0.1.0rc1] - 2026-10-02
+
+Release candidate of 0.1.0, published to PyPI for testing. Its changes are listed under
+0.1.0; the release is identical apart from this changelog.
+
+## [0.1.0a1] - 2026-10-01
+
+First alpha: the statistics core and the calculator commands.
+
+### Added
+
+- `fieldtrial.stats`, the statistics core (numpy and scipy only):
+  - one-arm intervals: Wilson (default), Clopper–Pearson, Jeffreys, Agresti–Coull, and an
+    exact binomial test against a threshold
+  - two independent arms: Newcombe difference CI, Boschloo exact test (primary), Fisher's
+    exact test and the conditional odds ratio
+  - paired designs: exact McNemar, Tango score CI, Cochran's Q with pairwise McNemar
+  - multiplicity adjustments: Holm, Bonferroni, Benjamini–Hochberg
+  - planning: sample size, power and minimum detectable effect (pooled-z, Fleiss
+    continuity-corrected, arcsine; unequal allocation), exact power for Boschloo and
+    McNemar, and a seeded simulation cross-check
+  - Bayesian summaries (descriptive only): P(p_B > p_A) and credible intervals
+- Calculator commands `fieldtrial ci`, `compare`, `paired`, `power`, `mde` and `adjust`,
+  each with `--json` output.
+- Documentation: a statistics reference page per method, a command-line page and the API
+  reference.
+- Project skeleton: packaging, `fieldtrial --version`, CI and the documentation site.
+
+[Unreleased]: https://github.com/rokbenko/fieldtrial/compare/v0.2.0...HEAD
+[0.2.0]: https://github.com/rokbenko/fieldtrial/compare/v0.1.0...v0.2.0
+[0.2.0rc1]: https://github.com/rokbenko/fieldtrial/compare/v0.2.0a2...v0.2.0rc1
+[0.2.0a2]: https://github.com/rokbenko/fieldtrial/compare/v0.2.0a1...v0.2.0a2
+[0.2.0a1]: https://github.com/rokbenko/fieldtrial/compare/v0.1.0...v0.2.0a1
+[0.1.0]: https://github.com/rokbenko/fieldtrial/compare/v0.1.0a1...v0.1.0
+[0.1.0rc1]: https://github.com/rokbenko/fieldtrial/compare/v0.1.0a1...v0.1.0rc1
+[0.1.0a1]: https://github.com/rokbenko/fieldtrial/releases/tag/v0.1.0a1

fieldtrial-0.2.0/PKG-INFO

@@ -0,0 +1,204 @@
+Metadata-Version: 2.5
+Name: fieldtrial
+Version: 0.2.0
+Summary: Find out whether your robot policy actually got better: statistically rigorous real-world evaluation of robot policies.
+Project-URL: Homepage, https://github.com/rokbenko/fieldtrial
+Project-URL: Repository, https://github.com/rokbenko/fieldtrial
+Project-URL: Issues, https://github.com/rokbenko/fieldtrial/issues
+Project-URL: Changelog, https://github.com/rokbenko/fieldtrial/blob/main/CHANGELOG.md
+Author: Rok Benko
+License-Expression: Apache-2.0
+License-File: LICENSE
+Keywords: confidence intervals,hypothesis testing,lerobot,policy evaluation,robot learning,robotics,statistics,vla
+Classifier: Development Status :: 2 - Pre-Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3 :: Only
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: Scientific/Engineering :: Mathematics
+Classifier: Typing :: Typed
+Requires-Python: >=3.11
+Requires-Dist: alembic>=1.13
+Requires-Dist: fastapi>=0.115
+Requires-Dist: jinja2>=3.1
+Requires-Dist: matplotlib>=3.8
+Requires-Dist: numpy>=1.26
+Requires-Dist: pillow>=10
+Requires-Dist: pydantic>=2.7
+Requires-Dist: python-multipart>=0.0.9
+Requires-Dist: pyyaml>=6.0
+Requires-Dist: qrcode>=7.4
+Requires-Dist: rich>=13.8
+Requires-Dist: scipy>=1.13
+Requires-Dist: sqlalchemy>=2.0.30
+Requires-Dist: sse-starlette>=2.1
+Requires-Dist: typer>=0.15
+Requires-Dist: uvicorn[standard]>=0.30
+Provides-Extra: capture
+Requires-Dist: av>=12; extra == 'capture'
+Requires-Dist: opencv-python-headless>=4.8; extra == 'capture'
+Provides-Extra: dev
+Requires-Dist: httpx>=0.28; extra == 'dev'
+Requires-Dist: hypothesis>=6.130; extra == 'dev'
+Requires-Dist: import-linter>=2.1; extra == 'dev'
+Requires-Dist: mypy>=2.0; extra == 'dev'
+Requires-Dist: pre-commit>=4.0; extra == 'dev'
+Requires-Dist: pytest-cov>=6.0; extra == 'dev'
+Requires-Dist: pytest>=8.3; extra == 'dev'
+Requires-Dist: ruff>=0.16; extra == 'dev'
+Requires-Dist: scipy-stubs>=1.15; extra == 'dev'
+Requires-Dist: statsmodels>=0.14.4; extra == 'dev'
+Requires-Dist: types-pyyaml>=6.0; extra == 'dev'
+Provides-Extra: docs
+Requires-Dist: mkdocs-material>=9.6; extra == 'docs'
+Requires-Dist: mkdocs<2,>=1.6; extra == 'docs'
+Requires-Dist: mkdocstrings[python]>=0.29; extra == 'docs'
+Provides-Extra: lerobot
+Requires-Dist: pyarrow>=15; extra == 'lerobot'
+Provides-Extra: openpi
+Requires-Dist: websockets>=13; extra == 'openpi'
+Description-Content-Type: text/markdown
+
+# fieldtrial
+
+**Find out whether your robot policy actually got better.**
+
+[![CI](https://github.com/rokbenko/fieldtrial/actions/workflows/ci.yml/badge.svg)](https://github.com/rokbenko/fieldtrial/actions/workflows/ci.yml)
+[![PyPI](https://img.shields.io/pypi/v/fieldtrial.svg)](https://pypi.org/project/fieldtrial/)
+[![License: Apache-2.0](https://img.shields.io/badge/license-Apache--2.0-blue.svg)](https://github.com/rokbenko/fieldtrial/blob/main/LICENSE)
+
+fieldtrial is an open-source Python framework for statistically rigorous, real-world
+evaluation of robot policies. LeRobot trains the policy; fieldtrial tells you whether it
+actually got better.
+
+## Why
+
+Real-world evaluation is slow, noisy and ad hoc. With 40 rollouts, the 95% confidence
+interval for a success rate is 12 to 15 percentage points wide on each side, so many apparent
+improvements from a fine-tuning run are just noise. fieldtrial helps at every stage:
+
+- **Before:** how many rollouts do I need? What is the smallest difference this
+  evaluation can detect?
+- **During:** a randomized, blinded schedule, and a phone-friendly operator console that
+  records the outcome, the furthest stage reached and the failure mode of each rollout.
+- **After:** the right statistics for the design (exact tests, paired analyses,
+  multiplicity control), honest wording, and a self-contained report.
+
+## 30-second demo
+
+```console
+$ uvx fieldtrial compare 74/80 91/120
+Arm 1: 74/80 = 92.5% (95% CI 84.6%–96.5%)
+Arm 2: 91/120 = 75.8% (95% CI 67.4%–82.6%)
+Difference +16.7 pp, Newcombe 95% CI [+6.2 pp, +26.0 pp]
+Boschloo p = 0.0020 (two-sided); rejects H0 at α = 0.05
+Fisher p = 0.0022
+
+$ uvx fieldtrial power --p1 0.76 --p2 0.90      # rollouts needed to detect 76% → 90%
+112 per arm (pooled-z)
+
+$ uvx fieldtrial demo                           # a simulated study in the console
+```
+
+`fieldtrial demo` opens a half-run, blinded study with a simulated robot. Run the rest
+from the keyboard (Space, Space, Enter), unblind, and read the report.
+
+<p>
+  <img src="https://raw.githubusercontent.com/rokbenko/fieldtrial/main/docs/assets/console-running.png" alt="The operator console on a phone: a trial in progress with its blind code, timer and a large Stop button" width="260">
+  <img src="https://raw.githubusercontent.com/rokbenko/fieldtrial/main/docs/assets/console-label.png" alt="Labelling a trial: furthest stage reached, why it ended, failure tags" width="260">
+</p>
+<p>
+  <img src="https://raw.githubusercontent.com/rokbenko/fieldtrial/main/docs/assets/report.png" alt="The HTML report: summary, success rate per arm with confidence intervals, primary analysis and a forest plot" width="560">
+</p>
+
+<!-- A short GIF of a trial run in the console goes here. -->
+
+## Quickstart
+
+```console
+$ uv tool install fieldtrial              # or: pip install fieldtrial
+$ fieldtrial init my-study                # writes my-study/study.yaml
+$ fieldtrial plan my-study --baseline 0.75
+$ fieldtrial lock my-study                # freezes the design and randomizes the schedule
+$ fieldtrial serve my-study --lan         # scan the QR code with a phone at the robot
+$ fieldtrial unblind my-study
+$ fieldtrial report my-study              # my-study/reports/report.html
+```
+
+A study is one folder: `study.yaml` (the design) and a SQLite database. Everything is
+local and works offline; fieldtrial sends no telemetry.
+
+- [Documentation](https://rokbenko.github.io/fieldtrial/): quickstart, concepts, guides
+  and a statistics reference.
+- [Re-analysis of Dream Machines' published pi0.5 results](https://github.com/rokbenko/fieldtrial/tree/main/examples/dream-machines-pi05):
+  which of 30 published comparisons the data actually resolve.
+
+## What you get
+
+- **Design:**
+  - randomized complete blocks with balanced arm order
+  - crossover rounds for tasks whose scene carries over
+  - checkpoint ladders
+  - planned interim looks with early stopping
+  - blind codes, a design hash, locking and logged amendments
+- **Real blinding:** a command runner that launches your rollout command for each arm,
+  and an openpi router that sends each trial's requests to that arm's policy server.
+- **Console:** sessions with a rig checklist, a timer, stage and failure-tag labels,
+  10-second undo, invalid trials with automatic rescheduling, a live mirror screen,
+  keyboard and foot-pedal keys, LAN access with a QR code.
+- **Evidence:** an evaluation camera that records every trial, rig checks against a
+  reference photo, and links from trials to LeRobot dataset episodes (with DAgger
+  interventions).
+- **Analysis:** the primary test follows from the locked design:
+  - exact McNemar with a Tango interval, Cochran–Mantel–Haenszel, or Cochran's Q with Holm
+  - group-sequential boundaries
+  - the association of success with training step, and plateau detection
+  - a period-adjusted crossover test
+
+  Every analysis also gets an independent-samples sensitivity analysis, stage funnels,
+  time to success, drift checks and a list of every deviation from the plan.
+- **Reports:** self-contained HTML with charts, Markdown for pull requests, and a
+  versioned JSON results model.
+- **Integration:** a REST API with a dependency-free Python client for custom runtimes,
+  CSV import and export, and `fieldtrial.stats` as a library.
+
+Every statistical function is tested against an independent reference implementation.
+Reports only describe a difference when the pre-registered test rejects; otherwise they
+say what the study could have detected.
+
+## How it fits with LeRobot and openpi
+
+fieldtrial complements LeRobot and openpi and never forks them. It is not a training
+framework, a simulation benchmark, a robot driver, a labeling platform or a cloud
+service. In manual mode, fieldtrial schedules and records the trials and you run the
+robot however you like. For real blinding, the `command` runner launches your rollout
+command (for example `lerobot-rollout`) for each arm, and the `openpi_router` runner
+routes an openpi client's traffic to each trial's policy server; see
+[Real blinding with runners](https://rokbenko.github.io/fieldtrial/guides/runners/).
+
+## How to cite
+
+If fieldtrial helps your research, please cite it (see
+[CITATION.cff](https://github.com/rokbenko/fieldtrial/blob/main/CITATION.cff)):
+
+```bibtex
+@software{benko_fieldtrial,
+  author  = {Benko, Rok},
+  title   = {fieldtrial: statistically rigorous real-world evaluation for robot policies},
+  url     = {https://github.com/rokbenko/fieldtrial},
+  license = {Apache-2.0}
+}
+```
+
+For evaluation practice in general, see Kress-Gazit et al. (2024), *Robot Learning as an
+Empirical Science: Best Practices for Policy Evaluation*, arXiv:2409.09491.
+
+## Development
+
+See [CONTRIBUTING.md](https://github.com/rokbenko/fieldtrial/blob/main/CONTRIBUTING.md).

fieldtrial-0.2.0/README.md

@@ -0,0 +1,137 @@
+# fieldtrial
+
+**Find out whether your robot policy actually got better.**
+
+[![CI](https://github.com/rokbenko/fieldtrial/actions/workflows/ci.yml/badge.svg)](https://github.com/rokbenko/fieldtrial/actions/workflows/ci.yml)
+[![PyPI](https://img.shields.io/pypi/v/fieldtrial.svg)](https://pypi.org/project/fieldtrial/)
+[![License: Apache-2.0](https://img.shields.io/badge/license-Apache--2.0-blue.svg)](https://github.com/rokbenko/fieldtrial/blob/main/LICENSE)
+
+fieldtrial is an open-source Python framework for statistically rigorous, real-world
+evaluation of robot policies. LeRobot trains the policy; fieldtrial tells you whether it
+actually got better.
+
+## Why
+
+Real-world evaluation is slow, noisy and ad hoc. With 40 rollouts, the 95% confidence
+interval for a success rate is 12 to 15 percentage points wide on each side, so many apparent
+improvements from a fine-tuning run are just noise. fieldtrial helps at every stage:
+
+- **Before:** how many rollouts do I need? What is the smallest difference this
+  evaluation can detect?
+- **During:** a randomized, blinded schedule, and a phone-friendly operator console that
+  records the outcome, the furthest stage reached and the failure mode of each rollout.
+- **After:** the right statistics for the design (exact tests, paired analyses,
+  multiplicity control), honest wording, and a self-contained report.
+
+## 30-second demo
+
+```console
+$ uvx fieldtrial compare 74/80 91/120
+Arm 1: 74/80 = 92.5% (95% CI 84.6%–96.5%)
+Arm 2: 91/120 = 75.8% (95% CI 67.4%–82.6%)
+Difference +16.7 pp, Newcombe 95% CI [+6.2 pp, +26.0 pp]
+Boschloo p = 0.0020 (two-sided); rejects H0 at α = 0.05
+Fisher p = 0.0022
+
+$ uvx fieldtrial power --p1 0.76 --p2 0.90      # rollouts needed to detect 76% → 90%
+112 per arm (pooled-z)
+
+$ uvx fieldtrial demo                           # a simulated study in the console
+```
+
+`fieldtrial demo` opens a half-run, blinded study with a simulated robot. Run the rest
+from the keyboard (Space, Space, Enter), unblind, and read the report.
+
+<p>
+  <img src="https://raw.githubusercontent.com/rokbenko/fieldtrial/main/docs/assets/console-running.png" alt="The operator console on a phone: a trial in progress with its blind code, timer and a large Stop button" width="260">
+  <img src="https://raw.githubusercontent.com/rokbenko/fieldtrial/main/docs/assets/console-label.png" alt="Labelling a trial: furthest stage reached, why it ended, failure tags" width="260">
+</p>
+<p>
+  <img src="https://raw.githubusercontent.com/rokbenko/fieldtrial/main/docs/assets/report.png" alt="The HTML report: summary, success rate per arm with confidence intervals, primary analysis and a forest plot" width="560">
+</p>
+
+<!-- A short GIF of a trial run in the console goes here. -->
+
+## Quickstart
+
+```console
+$ uv tool install fieldtrial              # or: pip install fieldtrial
+$ fieldtrial init my-study                # writes my-study/study.yaml
+$ fieldtrial plan my-study --baseline 0.75
+$ fieldtrial lock my-study                # freezes the design and randomizes the schedule
+$ fieldtrial serve my-study --lan         # scan the QR code with a phone at the robot
+$ fieldtrial unblind my-study
+$ fieldtrial report my-study              # my-study/reports/report.html
+```
+
+A study is one folder: `study.yaml` (the design) and a SQLite database. Everything is
+local and works offline; fieldtrial sends no telemetry.
+
+- [Documentation](https://rokbenko.github.io/fieldtrial/): quickstart, concepts, guides
+  and a statistics reference.
+- [Re-analysis of Dream Machines' published pi0.5 results](https://github.com/rokbenko/fieldtrial/tree/main/examples/dream-machines-pi05):
+  which of 30 published comparisons the data actually resolve.
+
+## What you get
+
+- **Design:**
+  - randomized complete blocks with balanced arm order
+  - crossover rounds for tasks whose scene carries over
+  - checkpoint ladders
+  - planned interim looks with early stopping
+  - blind codes, a design hash, locking and logged amendments
+- **Real blinding:** a command runner that launches your rollout command for each arm,
+  and an openpi router that sends each trial's requests to that arm's policy server.
+- **Console:** sessions with a rig checklist, a timer, stage and failure-tag labels,
+  10-second undo, invalid trials with automatic rescheduling, a live mirror screen,
+  keyboard and foot-pedal keys, LAN access with a QR code.
+- **Evidence:** an evaluation camera that records every trial, rig checks against a
+  reference photo, and links from trials to LeRobot dataset episodes (with DAgger
+  interventions).
+- **Analysis:** the primary test follows from the locked design:
+  - exact McNemar with a Tango interval, Cochran–Mantel–Haenszel, or Cochran's Q with Holm
+  - group-sequential boundaries
+  - the association of success with training step, and plateau detection
+  - a period-adjusted crossover test
+
+  Every analysis also gets an independent-samples sensitivity analysis, stage funnels,
+  time to success, drift checks and a list of every deviation from the plan.
+- **Reports:** self-contained HTML with charts, Markdown for pull requests, and a
+  versioned JSON results model.
+- **Integration:** a REST API with a dependency-free Python client for custom runtimes,
+  CSV import and export, and `fieldtrial.stats` as a library.
+
+Every statistical function is tested against an independent reference implementation.
+Reports only describe a difference when the pre-registered test rejects; otherwise they
+say what the study could have detected.
+
+## How it fits with LeRobot and openpi
+
+fieldtrial complements LeRobot and openpi and never forks them. It is not a training
+framework, a simulation benchmark, a robot driver, a labeling platform or a cloud
+service. In manual mode, fieldtrial schedules and records the trials and you run the
+robot however you like. For real blinding, the `command` runner launches your rollout
+command (for example `lerobot-rollout`) for each arm, and the `openpi_router` runner
+routes an openpi client's traffic to each trial's policy server; see
+[Real blinding with runners](https://rokbenko.github.io/fieldtrial/guides/runners/).
+
+## How to cite
+
+If fieldtrial helps your research, please cite it (see
+[CITATION.cff](https://github.com/rokbenko/fieldtrial/blob/main/CITATION.cff)):
+
+```bibtex
+@software{benko_fieldtrial,
+  author  = {Benko, Rok},
+  title   = {fieldtrial: statistically rigorous real-world evaluation for robot policies},
+  url     = {https://github.com/rokbenko/fieldtrial},
+  license = {Apache-2.0}
+}
+```
+
+For evaluation practice in general, see Kress-Gazit et al. (2024), *Robot Learning as an
+Empirical Science: Best Practices for Policy Evaluation*, arXiv:2409.09491.
+
+## Development
+
+See [CONTRIBUTING.md](https://github.com/rokbenko/fieldtrial/blob/main/CONTRIBUTING.md).